Method and device for parapapillary atrophy segmentation and structure quantification

Abstract

The application discloses a disc parapapillary atrophy segmentation and structure quantitative analysis method and equipment, and belongs to the field of computer vision. A parapapillary atrophy image dataset of myopic teenagers is constructed; a parapapillary atrophy region segmentation model is constructed, and training is completed, so that the segmentation of the parapapillary atrophy region in a cross-domain fundus image is realized; the parapapillary atrophy region segmented is subjected to multi-item structural quantification, and correlation analysis is performed on the diopter, age and eye axial length recorded in the parapapillary atrophy image dataset of myopic teenagers, so that quantitative analysis used for the progress evaluation of myopia of teenagers and the prediction of the diopter state is obtained. The steps are simple, convenient to use, and effective technical means are provided for the analysis of the parapapillary atrophy region image.

Description

Technical Field

[0001] This invention relates to a method and equipment for segmentation and quantitative analysis of optic disc paralobule atrophy, belonging to the field of computer technology. Background Technology

[0002] Adolescence is a critical period for the occurrence and progression of myopia. During this time, the refractive state of the adolescents is highly plastic and developmental. Therefore, early identification of structural changes is crucial for individualized myopia prevention and control.

[0003] Existing studies have shown that the enlargement of peripapillary atrophy (PPA) is a key structural marker of myopia progression and pathological myopia, and its morphological changes are closely related to myopia progression, axial elongation, and various fundus lesions.

[0004] In this context, the segmentation and quantitative analysis of PPA are particularly crucial.

[0005] Mainstream models rely on a large amount of fine annotation to maintain stable performance, but the lesion outlines in fundus images are thin and the boundaries are weak, making professional annotation extremely costly and making it difficult to obtain sufficient annotation data in real clinical environments.

[0006] Furthermore, segmentation models often rely on convolution stacking to obtain local structural features, which has a large parameter scale and computational overhead. However, when annotation is limited and there are significant differences across device domains, the expressive power of convolution is easily limited, especially for regions with fine-grained structural changes such as spectral shrunkenness, where the performance is often not robust enough.

[0007] In PPA segmentation tasks, high-quality samples that match myopia biological parameters are scarce. Public datasets, such as pathological myopia datasets, contain only a small number of PPA lesions and are mostly concentrated in cases of high myopia atrophy, which is of insufficient research value for low to moderate myopia.

[0008] To alleviate the performance bottleneck caused by scarce annotations, various semi-supervised learning (SSL) methods have emerged in recent years. These methods apply different enhancements to unlabeled samples to establish prediction consistency, thereby improving the generalization ability of the model without increasing the annotation cost.

[0009] However, in PPA segmentation, not only are annotations scarce, but the significant domain differences caused by multi-device acquisition also make it difficult for the model to generalize across institutions. Traditional SSL methods are unable to effectively alleviate this problem of cross-domain performance degradation.

[0010] In addition, the PPA is located close to the optic disc, and its morphology is significantly affected by individual differences and disease progression, with considerable variations in shape, extent, and brightness.

[0011] Existing SSL or cross-domain methods often fail to fully utilize global fundus structural information and do not effectively constrain intra-class consistency of lesions across patients and domains. Summary of the Invention

[0012] Summary of the invention: To address the shortcomings of the above-mentioned technologies, a method and device for segmentation and quantitative analysis of optic disc paralobule atrophy are developed. The method is simple to implement and has good results. It will provide important support for revealing early structural changes in myopia and improving the quantitative basis for risk prediction.

[0013] To achieve the above technical objectives, this invention discloses a method for segmentation and quantitative structural analysis of optic disc paralobule atrophy, the specific steps of which are as follows: Step 1: Construct an image dataset of paradiscal atrophy in myopic adolescents; Step 2: Construct a segmentation model for the paradiscal atrophy region and complete its training to achieve segmentation of the paradiscal atrophy region in cross-domain fundus images; Step 3: Perform multiple structural quantifications on the segmented paradiscal atrophy region and conduct correlation analysis with the refractive error, age, and axial length recorded in the paradiscal atrophy image dataset of myopic adolescents to obtain quantitative analysis for assessing myopia progression and predicting refractive status in adolescents.

[0014] Furthermore, the construction process of the paradiscal atrophy PPA dataset for myopic adolescents is as follows: fundus images of myopic adolescents are collected and organized, and corresponding refractive biological parameters are recorded to construct a dedicated dataset for paradiscal atrophy of myopic adolescents containing fundus images, annotations of paradiscal atrophy lesions, and biological parameters such as refractive power.

[0015] Furthermore, the PPA dataset was divided into training and testing sets in an 8:2 ratio, and the samples in the training set were horizontally flipped and standardized for data augmentation. The training set contains N s One from the source domain Annotated fundus images – annotation pairs As a sample, and the target domain of Annotated fundus images - annotation pairs and Unlabeled fundus images As a sample, ; Indicates that the source domain has labeled images, Representing pixel-level annotations of the source domain image. Indicates that the target domain has been labeled with images. Representing pixel-level annotations of the target domain image. This indicates an unlabeled image in the target domain; The resolution of the fundus image is , These represent height, width, and channels, respectively; the pixel-level annotation corresponding to image x. For the semantic annotation map of the paradiscal atrophy region with the same size as the input image, the pixel values ​​satisfy... k represents the number of categories.

[0016] Furthermore, the visual disk-side shrinkage region segmentation model includes a frequency domain enhancement processing module SAFA, an improved mean teacher network model, a feature fusion module FPG, and a cross-domain prototype alignment module CDPA; the improved mean teacher network model includes two pairs of UKAN structures, encoder I and decoder I and encoder II and decoder II, as well as encoder III of a transformer model structure; the FPG module is located at the bottleneck layer position of encoder I, decoder I, and encoder II and decoder II. The SAFA module processes the input source-domain labeled fundus images. Combined with labeled fundus images of the target domain This generates an enhanced image that differs from the original image but maintains structural consistency in the peridiscal atrophy region. Enhance the image and the target domain labeled fundus images Form a sample group; Sample group and unlabeled fundus images Simultaneous input of encoders I and III, unlabeled fundus images The input encoder II, encoder I, encoder II and encoder III convert the input fundus image into a multi-level semantic feature representation through convolution and downsampling operations; The FPG module includes a query matrix Q, a key matrix K, and a value matrix V. The key matrix K is transposed to obtain Kᵀ, which is used for matrix multiplication with the query matrix Q to calculate attention weights. The encoder III outputs global anatomical features. Segmentation features of encoder I respectively Or the segmentation features of encoder II The FPG module performs scale unification and structure enhancement on the input features, including segmentation features. Input query matrix Q, global anatomical features Input the transpose key matrix K respectively T Sum matrix V, transpose bond matrix K T Multiplying the query matrix Q yields the feature relevance matrix, which is then normalized using the Softmax activation function to obtain the attention weight matrix. This attention weight matrix is ​​then multiplied by the value matrix V to generate intermediate segmentation features that fuse global anatomical information. Intermediate segmentation features The input is fed into decoder I and decoder II for upsampling and reconstruction processing. Decoder I recovers the spatial resolution of the features step by step and outputs the source domain segmented image. Target domain segmentation image and unlabeled segmented images Decoder II recovers the spatial resolution of features step by step and outputs an unlabeled segmented image. Generate and input the original fundus image The predicted segmentation results of the shrunk regions next to the visual disk of consistent size; the predicted segmentation results of the shrunk regions next to the visual disk of decoder I and decoder II are input into the CDPA module, and through prototype learning and loss function constraints, intra-class feature aggregation and inter-class feature separation are achieved; The CDPA module handles data from the source domain. and target domain The segmentation features are used to construct semantic structure prototypes, and prototype matching and alignment are performed on the predicted segmentation results of the paradiscal atrophy region under different imaging conditions to reduce the self-source domain. and target domain The impact of feature distribution differences between images on the accuracy of paradiscal atrophy segmentation; The para-optic disc atrophy region segmentation model is trained by jointly constraining the segmentation results of labeled and unlabeled fundus images through supervised loss and unsupervised consistency loss, and finally outputs the segmentation mask results for quantitative analysis of the para-optic disc atrophy region structure.

[0017] Furthermore, the SAFA module works as follows: Obtain the source domain from the training dataset samples , source domain The sample is a labeled fundus image acquired under the first imaging condition and containing the paradiscal atrophy region; each pixel in the labeled fundus image is labeled as whether the pixel is the paradiscal atrophy region; Source domain via SAFA module samples Frequency domain transformation is performed to extract the low-frequency amplitude components from the spectrum. These low-frequency amplitude components characterize the brightness distribution pattern and color attenuation trend of the optic disc sidelobe atrophy region during imaging. Furthermore, from the target domain... samples Extract the corresponding low-frequency amplitude components in the target domain. samples For fundus images acquired under the second imaging condition, the target domain Imaging equipment, imaging parameters, or acquisition environment and source domain of medium-sized samples different; Using the SAFA module, first apply Fourier transform style transfer to the source domain. medium sample Low-frequency amplitude components and target domain medium sample The low-frequency amplitude components are swapped to obtain a preliminary enhanced image. To simulate the overall brightness attenuation and color shift changes in the optic disc paralobule atrophy region under different imaging devices and acquisition conditions, while maintaining the spatial structure and anatomical morphology of the source domain optic disc paralobule atrophy region, the study aimed to achieve the desired results. medium sample Low-frequency degradation features under target imaging conditions are obtained, and enhanced samples are constructed to simulate the differences in lesion manifestations across imaging conditions. After obtaining the initial enhanced image Then, through a preset global interpolation function Weighted interpolation is performed on the images before and after component swapping. The randomness of M is adjusted by the policy controller to construct a composite frequency domain component that simultaneously fuses the imaging degradation features of the source and target domains. Enhanced samples with consistent optic disc sidelobe atrophy structural features are obtained using the following formula. : .

[0018] Furthermore, encoder I in the improved mean-based teacher network model is used to process the input enhanced image. and the target domain labeled fundus images Segmentation features were extracted from paradiscal atrophy images from the sample group. , Representing channels, height, and width respectively, encoder II processes the input unlabeled image of the target domain. Extracting segmentation features from paradiscal atrophy images Encoder III is an independent pre-trained semantically guided encoder. Encoder III is a VisionFM model. It is used to encode features from fundus images and extract global anatomical features related to the atrophic area of ​​the optic disc paralobule. B represents the number of input images, with a value of 3. The FPG module represents the number of samples, channels, height, and width; it guides the segmentation of paradiscal atrophy images by incorporating global anatomical features from encoder III output. Segmentation features of images with paradiscal atrophy The enhanced features of encoder I are obtained. The FPG module provides global anatomical features. Dismantle from Extracted features B=1, characteristic Used for feature fusion at the bottleneck layer of encoder II to obtain enhanced features of encoder II. .

[0019] Furthermore, the enhanced features of decoder I... Or the enhanced features of encoder II The data are input to decoder I or decoder II for upsampling and reconstruction, and the final prediction result is obtained. Decoder I outputs source domain segmented image Target domain segmentation image Unlabeled segmented images The initial segmentation prediction map of the conventional optic disc sidelobe atrophy region is obtained; the UKAN structure decoder I has 4 layers of features, of which the penultimate layer of features As input features to the CDPA module, lesion prototype constraint segmentation branches are constructed in the atrophic region of the optic disc paralobule.

[0020] Furthermore, the CDPA module sets m prototype vectors. Describe the diverse feature distribution of the optic disc paralobule atrophy region under different morphological characteristics, boundary morphology, and imaging conditions; the CDPA module calculates the input features. With m prototype vectors Similarity value between Selection and Features The prototype vector with the highest similarity As a category reference prototype for pixels, the category prediction results of pixels are corrected based on the category reference prototype to generate a second segmentation prediction result based on the prototype constraint. ; Based on current pixel features Similarity value between the corresponding prototype vector A weighted update is performed, causing the category prototypes to gradually shift towards the cluster centers represented by the category prototypes: , In the formula: It is a category prototype. These are weighting coefficients, set to 0.99. It is the number of pixels, It is the similarity value between pixel features and prototype vectors; Decoder I generates predicted segments of the peridiscal atrophy region and calculates the segmentation baseline loss based on the predicted segmentation results. Prototype prediction loss This is used to supervise the accuracy of the output of the optic disc atrophy region segmentation model; the intermediate features of decoder I are input into the CDPA module, and the loss function is obtained through prototype learning and cross-domain prototype alignment. Structural consistency constraints are applied to the feature representations; the segmentation loss function is used to supervise the prediction results of the output layer, and the CDPA module is used to supervise the learning of the feature layer representations, as shown in the following equation: , In the formula, Similarity value between pixel features and prototype vector composition, This represents the similarity value of the entire image; express Matrix transpose; It is a category prototype; This is the weighting coefficient, set to 0.1; The prototype of a non-target category is used to represent typical feature patterns that are easily misclassified as other categories during the current category segmentation process. The set of "negative prototypes" represents the set of prototype vectors for all non-target categories. The "negative prototype set" represents the set of prototype vectors for all categories other than the category to which the current pixel belongs, and is used as a comparative reference feature for the wrong category. The negative prototype set represents the target category. The target category negative prototype set represents the set of prototype vectors of other categories that do not match the current pixel. It is used to suppress the situation where pixel features are misclassified into non-optic disc sidelobe atrophy regions. The temperature coefficient is used to scale the similarity calculation results to control the concentration of different class probability distributions in the segmentation prediction results, thereby avoiding excessive smoothing or excessive dispersion of the segmentation results and improving the classification stability of the boundary region of the optic disc sidelobe atrophy area. Baseline segmentation loss Prototype prediction loss and prototype constraint loss Combined into the total loss function It is used to constrain the overall error between the prediction results of the optic disc sidelobe atrophy segmentation model and the actual annotation, so as to guide the optic disc sidelobe atrophy segmentation model to obtain stable and accurate optic disc sidelobe atrophy segmentation capability under different imaging conditions. , In the formula, ce represents the cross-entropy loss. Represents source domain segmentation image, Represents the target domain segmentation image, This represents an unlabeled segmented image. The weights are represented by the second-to-last layer features of Decoder I, which provide additional supervision from the perspective of feature clustering under the guidance of prototype learning. The loss function... Represented as: , In the formula Represent source domain segmentation images Target domain segmentation image Unlabeled segmented images The prediction results for the corresponding prototype branch; Baseline segmentation loss This is used to measure the difference between the initial segmentation prediction result and the corresponding ground truth annotation, ensuring that the optic disc sidelobe atrophy segmentation model has basic segmentation capabilities for optic disc sidelobe atrophy regions; prototype prediction loss. Used to measure the second segmentation prediction result generated based on the multi-prototype learning module CDPA. With real labeling The differences between them are used to enhance the ability of the optic disc sidelobe atrophy segmentation model to finely segment the optic disc sidelobe atrophy region under complex boundary morphologies and different imaging conditions; prototype constraint loss Used to measure the difference between prototype vectors, guiding the source and target domains to form a consistent prototype representation space, thereby reducing the impact of different imaging devices and variations in imaging parameters on the segmentation results; total function Represented as: , In the formula, , The weights represent the weights for different loss functions. Takes the value 1. The value is 0.5; Based on the total function The aforementioned feature mapping network was trained, and the loss function was minimized using the SGD optimizer to optimize the network parameters and obtain the optimal parameters for the optic disc sidelobe atrophy segmentation model.

[0021] Furthermore, the trained paradiscal atrophy region segmentation model is used to identify the input paradiscal atrophy image. The paradiscal atrophy image directly input to encoder I and encoder III can be output by the CDPA module as the image after paradiscal atrophy segmentation.

[0022] Furthermore, the specific steps of step 3 are as follows: Step 3a: Evaluate the performance of the model in the optic disc paralobule atrophy segmentation model, using Dice coefficient (DSC), intersection-over-union ratio (IoU), Precision, Recall, and 95% Hausdorff distance (HD95) as evaluation indicators. Step 3b: After obtaining the segmentation results of the target region, referring to previous work, perform structured quantitative analysis on the paradiscal atrophy region, specifically including: calculating the pixel area of ​​the segmented paradiscal atrophy region to obtain the lesion area parameter LA; calculating the minimum and average thickness values ​​of the paradiscal atrophy region in the radial direction to obtain the thickness parameter LT; obtaining the relative width parameter LW of the paradiscal atrophy region based on the proportional relationship between the thickness parameter and the optic disc diameter; and obtaining the angular extension parameter AD based on the angular coverage range of the paradiscal atrophy region.

[0023] A computer device includes a processor and a memory, the processor being electrically connected to the memory, the memory storing instructions and data, and the processor executing a method for segmenting and quantitatively analyzing the structure of optic disc paralobule atrophy. Beneficial Effects: This invention constructs the PPA dataset for paradiscal atrophy images in myopic adolescents, used for image segmentation and structural feature analysis of paradiscal atrophy, providing a data foundation for the automated recognition and quantification of paradiscal atrophy images. This invention designs a paradiscal atrophy region segmentation model, which localizes and understands the context of paradiscal atrophy image regions, achieving feature alignment and pseudo-label consistency constraints, thereby improving the model's cross-domain generalization performance. Based on the segmentation results, this invention extracts corresponding quantization parameters, providing a quantitative reference for research on paradiscal atrophy in myopic adolescents. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the structure of the visual disc-side shrinkage region segmentation model of the present invention. Detailed Implementation

[0025] The embodiments of the present invention will be further described below with reference to the accompanying drawings: This invention discloses a method for segmentation and quantitative structural analysis of optic disc paralobule atrophy, the steps of which are as follows: Step 1: Construct a dataset of paradiscal atrophy images for myopic adolescents; collect and organize fundus images of myopic adolescents, and record the corresponding refractive biological parameters to construct a dedicated dataset for paradiscal atrophy in myopic adolescents, including fundus images, annotations of paradiscal atrophy lesions, and biological parameters such as refractive power.

[0026] Step 2: Construct a segmentation model for the paradiscal atrophy region and complete its training to achieve segmentation of the paradiscal atrophy region in cross-domain fundus images; The PPA dataset was divided into training and test sets in an 8:2 ratio, and the samples in the training set were horizontally flipped and standardized for data augmentation. The training set contains N... s One from the source domain Annotated fundus images – annotation pairs As a sample, and the target domain of Annotated fundus images - annotation pairs and Unlabeled fundus images As a sample, ; Indicates that the source domain has labeled images, Representing pixel-level annotations of the source domain image. Indicates that the target domain has been labeled with images. Representing pixel-level annotations of the target domain image. This indicates an unlabeled image in the target domain; The resolution of the fundus image is , These represent height, width, and channels, respectively; the pixel-level annotation corresponding to image x. For the semantic annotation map of the paradiscal atrophy region with the same size as the input image, the pixel values ​​satisfy... k represents the number of categories.

[0027] like Figure 1 As shown, the visual disk-side shrinkage region segmentation model includes a frequency domain enhancement processing module SAFA, an improved mean teacher network model, a feature fusion module FPG, and a cross-domain prototype alignment module CDPA. The improved mean teacher network model includes two pairs of UKAN structures: encoder I and decoder I, and encoder II and decoder II, as well as encoder III of a transformer model structure. The FPG module is located at the bottleneck layer position of encoder I, decoder I, and encoder II and decoder II. The SAFA module processes the input source-domain labeled fundus images. Combined with labeled fundus images of the target domain This generates an enhanced image that differs from the original image but maintains structural consistency in the peridiscal atrophy region. Enhance the image and the target domain labeled fundus images Form a sample group; Sample group and unlabeled fundus images Simultaneous input of encoders I and III, unlabeled fundus images The input encoder II, encoder I, encoder II and encoder III convert the input fundus image into a multi-level semantic feature representation through convolution and downsampling operations; The FPG module includes a query matrix Q, a key matrix K, and a value matrix V. The key matrix K is transposed to obtain Kᵀ, which is used for matrix multiplication with the query matrix Q to calculate attention weights. The encoder III outputs global anatomical features. Segmentation features of encoder I respectively Or the segmentation features of encoder II The FPG module performs scale unification and structure enhancement on the input features, including segmentation features. Input query matrix Q, global anatomical features Input the transpose key matrix K respectively T Sum matrix V, transpose bond matrix K T Multiplying the query matrix Q yields the feature relevance matrix, which is then normalized using the Softmax activation function to obtain the attention weight matrix. This attention weight matrix is ​​then multiplied by the value matrix V to generate intermediate segmentation features that fuse global anatomical information. Intermediate segmentation features The input is fed into decoder I and decoder II for upsampling and reconstruction processing. Decoder I recovers the spatial resolution of the features step by step and outputs the source domain segmented image. Target domain segmentation image and unlabeled segmented images Decoder II recovers the spatial resolution of features step by step and outputs an unlabeled segmented image. Generate and input the original fundus image The predicted segmentation results of the shrunk regions next to the visual disk of consistent size; the predicted segmentation results of the shrunk regions next to the visual disk of decoder I and decoder II are input into the CDPA module, and through prototype learning and loss function constraints, intra-class feature aggregation and inter-class feature separation are achieved; The CDPA module handles data from the source domain. and target domain The segmentation features are used to construct semantic structure prototypes, and prototype matching and alignment are performed on the predicted segmentation results of the paradiscal atrophy region under different imaging conditions to reduce the self-source domain. and target domain The impact of feature distribution differences between images on the accuracy of paradiscal atrophy segmentation; The para-optic disc atrophy region segmentation model is trained by jointly constraining the segmentation results of labeled and unlabeled fundus images through supervised loss and unsupervised consistency loss, and finally outputs the segmentation mask results for quantitative analysis of the para-optic disc atrophy region structure.

[0028] The SAFA module works as follows: Obtain the source domain from the training dataset samples , source domain The sample is a labeled fundus image acquired under the first imaging condition and containing the paradiscal atrophy region; each pixel in the labeled fundus image is labeled as whether the pixel is the paradiscal atrophy region; Source domain via SAFA module samples Frequency domain transformation is performed to extract the low-frequency amplitude components from the spectrum. These low-frequency amplitude components characterize the brightness distribution pattern and color attenuation trend of the optic disc sidelobe atrophy region during imaging. Furthermore, from the target domain... samples Extract the corresponding low-frequency amplitude components in the target domain. samples For fundus images acquired under the second imaging condition, the target domain Imaging equipment, imaging parameters, or acquisition environment and source domain of medium-sized samples different; Using the SAFA module, first apply Fourier transform style transfer to the source domain. medium sample Low-frequency amplitude components and target domain medium sample The low-frequency amplitude components are swapped to obtain a preliminary enhanced image. To simulate the overall brightness attenuation and color shift changes in the optic disc paralobule atrophy region under different imaging devices and acquisition conditions, while maintaining the spatial structure and anatomical morphology of the source domain optic disc paralobule atrophy region, the study aimed to achieve the desired results. medium sample Low-frequency degradation features under target imaging conditions are obtained, and enhanced samples are constructed to simulate the differences in lesion manifestations across imaging conditions. After obtaining the initial enhanced image Then, through a preset global interpolation function Weighted interpolation is performed on the images before and after component swapping. The randomness of M is adjusted by the policy controller to construct a composite frequency domain component that simultaneously fuses the imaging degradation features of the source and target domains. Enhanced samples with consistent optic disc sidelobe atrophy structural features are obtained using the following formula. : .

[0029] Encoder I in the improved mean-based teacher network model is used to process the input augmented image. and the target domain labeled fundus images Segmentation features were extracted from paradiscal atrophy images from the sample group. , Representing channels, height, and width respectively, encoder II processes the input unlabeled image of the target domain. Extracting segmentation features from paradiscal atrophy images Encoder III is an independent pre-trained semantically guided encoder. Encoder III is a VisionFM model. It is used to encode features from fundus images and extract global anatomical features related to the atrophic area of ​​the optic disc paralobule. B represents the number of input images, with a value of 3. The FPG module represents the number of samples, channels, height, and width; it guides the segmentation of paradiscal atrophy images by incorporating global anatomical features from encoder III output. Segmentation features of images with paradiscal atrophy The enhanced features of encoder I are obtained. The FPG module provides global anatomical features. Dismantle from Extracted features B=1, characteristic Used for feature fusion at the bottleneck layer of encoder II to obtain enhanced features of encoder II. .

[0030] Enhanced features of decoder I Or the enhanced features of encoder II The data are input to decoder I or decoder II for upsampling and reconstruction, and the final prediction result is obtained. Decoder I outputs source domain segmented image Target domain segmentation image Unlabeled segmented images The initial segmentation prediction map of the conventional optic disc sidelobe atrophy region is obtained; the UKAN structure decoder I has 4 layers of features, of which the penultimate layer of features As input features to the CDPA module, lesion prototype constraint segmentation branches are constructed in the atrophic region of the optic disc paralobule.

[0031] The CDPA module sets m prototype vectors. Describe the diverse feature distribution of the optic disc paralobule atrophy region under different morphological characteristics, boundary morphology, and imaging conditions; the CDPA module calculates the input features. With m prototype vectors Similarity value between Selection and Features The prototype vector with the highest similarity As a category reference prototype for pixels, the category prediction results of pixels are corrected based on the category reference prototype to generate a second segmentation prediction result based on the prototype constraint. ; Based on current pixel features Similarity value between the corresponding prototype vector A weighted update is performed, causing the category prototypes to gradually shift towards the cluster centers represented by the category prototypes: , In the formula: It is a category prototype. These are weighting coefficients, set to 0.99. It is the number of pixels, It is the similarity value between pixel features and prototype vectors; Decoder I generates predicted segments of the peridiscal atrophy region and calculates the segmentation baseline loss based on the predicted segmentation results. Prototype prediction loss This is used to supervise the accuracy of the output of the optic disc atrophy region segmentation model; the intermediate features of decoder I are input into the CDPA module, and the loss function is obtained through prototype learning and cross-domain prototype alignment. Structural consistency constraints are applied to the feature representations; the segmentation loss function is used to supervise the prediction results of the output layer, and the CDPA module is used to supervise the learning of the feature layer representations, as shown in the following equation: , In the formula, Similarity value between pixel features and prototype vector composition, This represents the similarity value of the entire image; express Matrix transpose; It is a category prototype; This is the weighting coefficient, set to 0.1; The prototype of a non-target category is used to represent typical feature patterns that are easily misclassified as other categories during the current category segmentation process. The set of "negative prototypes" represents the set of prototype vectors for all non-target categories. The "negative prototype set" represents the set of prototype vectors for all categories other than the category to which the current pixel belongs, and is used as a comparative reference feature for the wrong category. The negative prototype set represents the target category. The target category negative prototype set represents the set of prototype vectors of other categories that do not match the current pixel. It is used to suppress the situation where pixel features are misclassified into non-optic disc sidelobe atrophy regions. The temperature coefficient is used to scale the similarity calculation results to control the concentration of different class probability distributions in the segmentation prediction results, thereby avoiding excessive smoothing or excessive dispersion of the segmentation results and improving the classification stability of the boundary region of the optic disc sidelobe atrophy area. Baseline segmentation loss Prototype prediction loss and prototype constraint loss Combined into the total loss function It is used to constrain the overall error between the prediction results of the optic disc sidelobe atrophy segmentation model and the actual annotation, so as to guide the optic disc sidelobe atrophy segmentation model to obtain stable and accurate optic disc sidelobe atrophy segmentation capability under different imaging conditions. , In the formula, ce represents the cross-entropy loss. Represents source domain segmentation image, Represents the target domain segmentation image, This represents an unlabeled segmented image. The weights are represented by the second-to-last layer features of Decoder I, which provide additional supervision from the perspective of feature clustering under the guidance of prototype learning. The loss function... Represented as: , In the formula Represent source domain segmentation images Target domain segmentation image Unlabeled segmented images The prediction results for the corresponding prototype branch; Baseline segmentation loss This is used to measure the difference between the initial segmentation prediction result and the corresponding ground truth annotation, ensuring that the optic disc sidelobe atrophy segmentation model has basic segmentation capabilities for optic disc sidelobe atrophy regions; prototype prediction loss. Used to measure the second segmentation prediction result generated based on the multi-prototype learning module CDPA. With real labeling The differences between them are used to enhance the ability of the optic disc sidelobe atrophy segmentation model to finely segment the optic disc sidelobe atrophy region under complex boundary morphologies and different imaging conditions; prototype constraint loss Used to measure the difference between prototype vectors, guiding the source and target domains to form a consistent prototype representation space, thereby reducing the impact of different imaging devices and variations in imaging parameters on the segmentation results; total function Represented as: , In the formula, , The weights represent the weights for different loss functions. Takes the value 1. The value is 0.5; Based on the total function The aforementioned feature mapping network was trained, and the loss function was minimized using the SGD optimizer to optimize the network parameters and obtain the optimal parameters for the optic disc sidelobe atrophy segmentation model.

[0032] Step 3: Perform multiple structural quantifications on the segmented paradiscal atrophy region and conduct correlation analysis with the refractive error, age, and axial length recorded in the paradiscal atrophy image dataset of myopic adolescents to obtain quantitative analysis for assessing myopia progression and predicting refractive status in adolescents.

[0033] The performance of the model in the PPA segmentation task was evaluated using Dice coefficient (DSC), intersection-over-union ratio (IoU), precision, recall, and 95% Hausdorff distance (HD95) as evaluation metrics.

[0034] Based on previous work, the PPA region was quantitatively described.

[0035] The proposed metrics include lesion area (LA), optic disc area (DA), angular extension (AD), thickness (LT), relative width (LW), and area ratio (AR).

[0036] To verify the applicability and stability of the above method, public datasets and self-built datasets were selected as experimental data.

[0037] The public dataset PALM is used to provide source domain samples across devices and populations, while the self-built MTPPA dataset is used to provide target domain samples.

[0038] Based on this data combination, a comparative training and evaluation were conducted using a semi-supervised segmentation method, a domain adaptation method, and a scheme based on a base model. Upper and lower bound experiments were constructed to observe the changes in model performance under domain difference conditions, thereby verifying the effectiveness of the method of the present invention on fundus images of cross-domain myopia populations.

[0039] Step 4: After obtaining the segmentation results of the target region, extract its quantitative features, which can be used to reflect the relationship between structural changes in the target region and the refractive state.

[0040] The above description is merely one embodiment of the present invention and is not intended to limit the present invention. Any minor modifications, equivalent substitutions, and improvements made to the above embodiment based on the technical essence of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for segmentation and quantitative structural analysis of optic disc paralobule atrophy, characterized in that, The specific steps are as follows: Step 1: Construct an image dataset of paradiscal atrophy in myopic adolescents; Step 2: Construct a segmentation model for the paradiscal atrophy region and complete its training to achieve segmentation of the paradiscal atrophy region in cross-domain fundus images; Step 3: Perform multiple structural quantifications on the segmented paradiscal atrophy region and conduct correlation analysis with the refractive error, age, and axial length recorded in the paradiscal atrophy image dataset of myopic adolescents to obtain quantitative analysis for assessing myopia progression and predicting refractive status in adolescents.

2. The method for segmentation and quantitative structural analysis of optic disc paralobule atrophy according to claim 1, characterized in that, The process of constructing the paradiscal atrophy PPA dataset for myopic adolescents is as follows: fundus images of myopic adolescents are collected and organized, and corresponding refractive biological parameters are recorded to construct a dedicated dataset for paradiscal atrophy of myopic adolescents containing fundus images, annotations of paradiscal atrophy lesions, and biological parameters such as refractive power.

3. The method for segmentation and quantitative structural analysis of optic disc paralobule atrophy according to claim 2, characterized in that, The PPA dataset was divided into training and test sets in an 8:2 ratio, and the samples in the training set were horizontally flipped and standardized for data augmentation. The training set contains N s One from the source domain Annotated fundus images – annotation pairs As a sample, and the target domain of Annotated fundus images - annotation pairs and Unlabeled fundus images As a sample, ; Indicates that the source domain has labeled images, Representing pixel-level annotations of the source domain image. Indicates that the target domain has been labeled with images. Representing pixel-level annotations of the target domain image. This indicates an unlabeled image in the target domain; The resolution of the fundus image is , These represent height, width, and channels, respectively; the pixel-level annotation corresponding to image x. For the semantic annotation map of the paradiscal atrophy region with the same size as the input image, the pixel values ​​satisfy... k represents the number of categories.

4. The method for segmentation and quantitative structural analysis of optic disc paralobule atrophy according to claim 3, characterized in that, The visual disk-side shrinkage region segmentation model includes a frequency domain enhancement processing module (SAFA), an improved mean teacher network model, a feature fusion module (FPG), and a cross-domain prototype alignment module (CDPA). The improved mean teacher network model includes two pairs of UKAN structures: encoder I and decoder I, and encoder II and decoder II, as well as encoder III of a transformer model structure. The FPG module is located at the bottleneck layer position of encoder I, decoder I, and encoder II and decoder II. The SAFA module processes the input source-domain labeled fundus images. Combined with labeled fundus images of the target domain This generates an enhanced image that differs from the original image but maintains structural consistency in the peridiscal atrophy region. Enhance the image and the target domain labeled fundus images Form a sample group; Sample group and unlabeled fundus images Simultaneous input of encoders I and III, unlabeled fundus images The input encoder II, encoder I, encoder II and encoder III convert the input fundus image into a multi-level semantic feature representation through convolution and downsampling operations; The FPG module includes a query matrix Q, a key matrix K, and a value matrix V. The key matrix K is transposed to obtain Kᵀ, which is used for matrix multiplication with the query matrix Q to calculate attention weights. The encoder III outputs global anatomical features. Segmentation features of encoder I respectively Or the segmentation features of encoder II The FPG module performs scale unification and structure enhancement on the input features, including segmentation features. Input query matrix Q, global anatomical features Input the transpose key matrix K respectively T Sum matrix V, transpose bond matrix K T Multiplying the query matrix Q yields the feature relevance matrix, which is then normalized using the Softmax activation function to obtain the attention weight matrix. This attention weight matrix is ​​then multiplied by the value matrix V to generate intermediate segmentation features that fuse global anatomical information. Intermediate segmentation features The input is fed into decoder I and decoder II for upsampling and reconstruction processing. Decoder I recovers the spatial resolution of the features step by step and outputs the source domain segmented image. Target domain segmentation image and unlabeled segmented images Decoder II recovers the spatial resolution of features step by step and outputs an unlabeled segmented image. Generate and input the original fundus image Predicted segmentation results for optic disc atrophy regions of uniform size; The predicted segmentation results of the visual disc shrinkage region from decoder I and decoder II are input into the CDPA module. Through prototype learning and loss function constraints, intra-class feature aggregation and inter-class feature separation are achieved. The CDPA module handles data from the source domain. and target domain The segmentation features are used to construct semantic structure prototypes, and prototype matching and alignment are performed on the predicted segmentation results of the paradiscal atrophy region under different imaging conditions to reduce the self-source domain. and target domain The impact of feature distribution differences between images on the accuracy of paradiscal atrophy segmentation; The para-optic disc atrophy region segmentation model is trained by jointly constraining the segmentation results of labeled and unlabeled fundus images through supervised loss and unsupervised consistency loss, and finally outputs the segmentation mask results for quantitative analysis of the para-optic disc atrophy region structure.

5. The method for segmentation and quantitative structural analysis of optic disc paralobule atrophy according to claim 4, characterized in that, The SAFA module works as follows: Obtain the source domain from the training dataset samples , source domain The sample is a labeled fundus image acquired under the first imaging condition and containing the paradiscal atrophy region; each pixel in the labeled fundus image is labeled as whether the pixel is the paradiscal atrophy region; Source domain via SAFA module samples Frequency domain transformation is performed to extract the low-frequency amplitude components from the spectrum. These low-frequency amplitude components characterize the brightness distribution pattern and color attenuation trend of the optic disc sidelobe atrophy region during imaging. Furthermore, from the target domain... samples Extract the corresponding low-frequency amplitude components in the target domain. samples For fundus images acquired under the second imaging condition, the target domain Imaging equipment, imaging parameters, or acquisition environment and source domain of medium-sized samples different; Using the SAFA module, first apply Fourier transform style transfer to the source domain. medium sample The low-frequency amplitude components and the target domain medium sample The low-frequency amplitude components are swapped to obtain a preliminary enhanced image. To simulate the overall brightness attenuation and color shift changes in the optic disc paralobule atrophy region under different imaging devices and acquisition conditions, while maintaining the spatial structure and anatomical morphology of the source domain optic disc paralobule atrophy region, the study aimed to achieve the desired results. medium sample Low-frequency degradation features under target imaging conditions are obtained, and enhanced samples are constructed to simulate the differences in lesion manifestations across imaging conditions. After obtaining the initial enhanced image Then, through a preset global interpolation function Weighted interpolation is performed on the images before and after component swapping. The randomness of M is adjusted by the policy controller to construct a composite frequency domain component that simultaneously fuses the imaging degradation features of the source and target domains. Enhanced samples with consistent optic disc sidelobe atrophy structural features are obtained using the following formula. : 。 6. The method for segmentation and quantitative structural analysis of optic disc paralobule atrophy according to claim 5, characterized in that, Encoder I in the improved mean teacher network model is used to process the input augmented image. and the target domain labeled fundus images Segmentation features were extracted from paradiscal atrophy images from the sample group. , Representing channels, height, and width respectively, encoder II processes the input unlabeled image of the target domain. Extracting segmentation features from paradiscal atrophy images Encoder III is an independent pre-trained semantically guided encoder. Encoder III is a VisionFM model. It is used to encode features from fundus images and extract global anatomical features related to the atrophic area of ​​the optic disc paralobule. B represents the number of input images, with a value of 3. The FPG module represents the number of samples, channels, height, and width; it guides the segmentation of paradiscal atrophy images by incorporating global anatomical features from encoder III output. Segmentation features of images with paradiscal atrophy The enhanced features of encoder I are obtained. The FPG module provides global anatomical features. Dismantle from Extracted features B=1, characteristic Used for feature fusion at the bottleneck layer of encoder II to obtain enhanced features for encoder II. .

7. The method for segmentation and quantitative structural analysis of optic disc paralobule atrophy according to claim 6, characterized in that, Enhanced features of decoder I Or the enhanced features of encoder II The data are input to decoder I or decoder II for upsampling and reconstruction, and the final prediction result is obtained. Decoder I outputs source domain segmented image Target domain segmentation image Unlabeled segmented images The initial segmentation prediction map of the conventional optic disc sidelobe atrophy region is obtained; the UKAN structure decoder I has 4 layers of features, of which the penultimate layer of features As input features to the CDPA module, lesion prototype constraint segmentation branches are constructed in the atrophic region of the optic disc paralobule.

8. The method for segmentation and quantitative structural analysis of optic disc paralobule atrophy according to claim 7, characterized in that, The CDPA module sets m prototype vectors. Describe the diverse feature distribution of the optic disc paralobule atrophy region under different morphological characteristics, boundary morphology, and imaging conditions; the CDPA module calculates the input features. With m prototype vectors Similarity values ​​between Selection and Features The prototype vector with the highest similarity As a category reference prototype for pixels, the category prediction results of pixels are corrected based on the category reference prototype to generate a second segmentation prediction result based on the prototype constraint. ; Based on current pixel features Similarity value between the corresponding prototype vector A weighted update is performed, causing the category prototypes to gradually shift towards the cluster centers represented by the category prototypes: ， In the formula: It is a category prototype. These are weighting coefficients, set to 0.

99. It is the number of pixels, It is the similarity value between pixel features and prototype vectors; Decoder I generates predicted segments of the peridiscal atrophy region and calculates the segmentation baseline loss based on the predicted segmentation results. Prototype prediction loss This is used to supervise the accuracy of the output of the optic disc atrophy region segmentation model; the intermediate features of decoder I are input into the CDPA module, and the loss function is obtained through prototype learning and cross-domain prototype alignment. Structural consistency constraints are applied to the feature representations; the segmentation loss function is used to supervise the prediction results of the output layer, and the CDPA module is used to supervise the learning of the feature layer representations, as shown in the following equation: ， In the formula, Similarity value between pixel features and prototype vector composition, This represents the similarity value of the entire image; express Matrix transpose; It is a category prototype; This is the weighting coefficient, set to 0.1; The prototype of a non-target category is used to represent typical feature patterns that are easily misclassified as other categories during the current category segmentation process. This represents the set of "negative prototypes" for all non-target categories. The "negative prototype set" represents the set of prototype vectors for all categories other than the category to which the current pixel belongs, and is used as a comparative reference feature for incorrect categories. The negative prototype set represents the target category. The target category negative prototype set represents the set of prototype vectors of other categories that do not match the current pixel. It is used to suppress the situation where pixel features are misclassified into non-optic disc sidelobe atrophy regions. The temperature coefficient is used to scale the similarity calculation results to control the concentration of different class probability distributions in the segmentation prediction results, thereby avoiding excessive smoothing or excessive dispersion of the segmentation results and improving the classification stability of the boundary region of the optic disc sidelobe atrophy area. Baseline segmentation loss Prototype prediction loss and prototype constraint loss Combined into the total loss function It is used to constrain the overall error between the prediction results of the optic disc sidelobe atrophy segmentation model and the actual annotation, so as to guide the optic disc sidelobe atrophy segmentation model to obtain stable and accurate optic disc sidelobe atrophy segmentation capability under different imaging conditions. ， In the formula, ce represents the cross-entropy loss. Represents source domain segmentation image, Represents the target domain segmentation image, This represents an unlabeled segmented image. The weights are represented by the second-to-last layer features of Decoder I, which provide additional supervision from the perspective of feature clustering under the guidance of prototype learning. The loss function... Represented as: ， In the formula Represent source domain segmentation images Target domain segmentation image Unlabeled segmented images The prediction results for the corresponding prototype branch; Baseline segmentation loss This is used to measure the difference between the initial segmentation prediction result and the corresponding ground truth annotation, ensuring that the optic disc sidelobe atrophy segmentation model has basic segmentation capabilities for optic disc sidelobe atrophy regions; prototype prediction loss. Used to measure the second segmentation prediction result generated based on the multi-prototype learning module CDPA. With real labeling The differences between them are used to enhance the ability of the optic disc sidelobe atrophy segmentation model to finely segment the optic disc sidelobe atrophy region under complex boundary morphologies and different imaging conditions; prototype constraint loss It is used to measure the difference between prototype vectors and to guide the source domain and target domain to form a consistent prototype representation space, thereby reducing the impact of different imaging devices and changes in imaging parameters on the segmentation results. Total function Represented as: ， In the formula, , The weights represent the weights for different loss functions. Takes the value 1. The value is 0.5; Based on the total function The aforementioned feature mapping network was trained, and the loss function was minimized using the SGD optimizer to optimize the network parameters and obtain the optimal parameters for the optic disc sidelobe atrophy segmentation model.

9. The method for segmentation and quantitative structural analysis of optic disc paralobule atrophy according to claim 1, characterized in that, The trained paradiscal atrophy region segmentation model is used to identify the input paradiscal atrophy image. The paradiscal atrophy image directly input to encoder I and encoder III can be output by the CDPA module as the image after paradiscal atrophy segmentation.

10. A computer device, characterized in that, It includes a processor and a memory, the processor being electrically connected to the memory, the memory being used to store instructions and data, and the processor being used to execute the optic disc paralobule atrophy segmentation and structural quantitative analysis method according to any one of claims 1-9.

Images