A pathological myopia atrophy lesion segmentation method and system based on implicit dual-guided attention

Abstract

The present application relates to the technical field of medical image processing, in particular to a pathological myopia atrophy lesion segmentation method and system based on implicit double-guided attention; first, a pathological myopia fundus image is acquired and preprocessed, and a segmentation network composed of a ConvNeXt-Tiny encoder and a U-Net++ decoder is input to extract features; then, a shared convolution branch is used to generate a difficult attention map and an edge attention map through self-learning, which are used as guide signals to perform channel attention calibration of difficult guidance and spatial attention calibration of edge guidance on the features; finally, the calibrated features are weighted and fused and reconstructed by a learnable scalar, and an accurate lesion segmentation result is output. The present application accurately allocates attention through the double-guided mechanism, effectively solves the problems of fuzzy lesion boundary and background interference, and compensates for information loss through feature reconstruction, significantly improving the accuracy and continuity of segmentation.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method and system for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention. Background Technology

[0002] Pathological myopia is one of the leading eye diseases causing irreversible visual impairment and has become a serious public health problem worldwide. As the disease progresses, patients gradually develop typical pathological changes in the fundus, such as diffuse atrophy, patchy atrophy, and macular involvement. The distribution, morphology, and structure of the atrophic lesions are highly correlated with the degree of visual function decline. Achieving stable and precise pixel-level segmentation of atrophic lesions is not only the core basis for objective staging and dynamic monitoring of pathological myopia, but also an important prerequisite for clinical efficacy evaluation and prognosis, playing a crucial role in improving the diagnosis and treatment of pathological myopia.

[0003] Accurate segmentation of atrophic lesions in fundus images of pathological myopia presents significant technical challenges. Lesion boundaries often exhibit gradual transitions, lacking clear anatomical demarcations, and manual annotation is highly subjective and prone to variability. Furthermore, lesions vary greatly in scale, morphology, and spatial distribution, ranging from focal small patches to large-scale map-like expansions, making segmentation extremely difficult. In addition, background interference from fundus image noise, uneven illumination, and vascular and pigment textures further weakens lesion features, leading to insufficient stability of traditional threshold segmentation, region growing, and active contour models in complex clinical scenarios, making them unsuitable for practical applications.

[0004] In recent years, deep learning technology based on encoder-decoder architecture has driven the rapid development of medical image segmentation. Methods represented by U-Net and its derivatives have achieved good results in fundus lesion segmentation tasks due to their advantages in multi-scale feature extraction and fusion. However, for weakly boundary targets such as pathological myopic atrophy lesions, existing methods still have significant shortcomings: First, boundary enhancement strategies heavily rely on explicit boundary branches and boundary loss functions, requiring stringent annotation quality. Annotation noise can easily lead to the model learning unstable contours, resulting in poor segmentation robustness. Second, conventional attention modules only allocate weights based on the statistical information of the features themselves, lacking directional guidance that matches the difficulties in lesion identification. Attention shifts easily occur in low-contrast and blurred boundary regions, leading to missed lesion detection, false background detection, and insufficient boundary fit. Therefore, improving the model's discrimination and segmentation accuracy for weakly boundary atrophy lesions while reducing reliance on explicit and precise boundary supervision has become a key technical problem urgently needing to be solved in the field of intelligent diagnosis of pathological myopia. Summary of the Invention

[0005] The purpose of this invention is to propose a method and system for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention, which can maintain the integrity of the lesion area, effectively reduce background false detections, and improve the stability of boundary localization.

[0006] According to a first aspect of the present disclosure, a method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention is provided, comprising the following steps: Acquire fundus images of pathological myopia and perform preprocessing; The preprocessed pathological myopia fundus image is input into the segmentation network, multi-scale features are extracted and the terminal feature map X is output; Model the terminal feature map X using shared convolutional branches to generate a difficult attention map. and edge attention map ; Using the difficult attention map and the edge attention map as guiding signals, the features are subjected to difficult-guided channel attention calibration and edge-guided spatial attention calibration to obtain the calibrated features; A learnable scalar is introduced to perform weighted fusion and reconstruction of the calibrated features, compensating for information loss and outputting lesion segmentation results.

[0007] In one embodiment, shared convolutional branch modeling is performed on the terminal feature map X to generate a difficult attention map. and edge attention map Specifically: Model the terminal feature map X using shared convolutional branches to extract intermediate features. ; In intermediate features Based on this, edge attention maps are generated through two independent prediction branches. and difficult attention maps : in, This represents the Sigmoid activation function. This represents a 3×3 convolution operation in the edge branch. This represents a 1×1 convolution operation in the edge branch; This represents a 3×3 convolution operation in a difficult branch. This represents a 1×1 convolution operation in a difficult branch.

[0008] In one embodiment, a difficult attention map is used. and edge attention map Element-wise interactions are performed and applied to the terminal feature map X as residuals. After 1×1 convolution and batch normalization recalibration, the features are obtained. : in, Features after residual interaction; Features Perform channel attention calibration for difficult guidance and spatial attention calibration for edge guidance separately: In channel attention calibration, the difficult attention map is... Mapped to channel guiding vector The final channel attention weights are obtained by combining them with the basic channel descriptors. ; In spatial attention calibration, the edge attention map is... Convolutional mapping is used to create a spatial guiding graph. The final spatial attention weights are obtained by combining them with the basic spatial descriptor. .

[0009] In one embodiment, the channel attention weight Spatial attention weights The corresponding feature update relationship is represented as follows: in, This represents the Sigmoid activation function. The base channel descriptor is obtained by global average pooling and global max pooling. This is the basic space descriptor obtained by concatenating average pooling and max pooling; This is the feature map after channel attention enhancement. This enhances spatial features.

[0010] In one embodiment, a learnable scalar is introduced to perform weighted fusion and reconstruction of the calibrated features, specifically as follows: Introducing a learnable scalar α to enhance spatial features and characteristics Weighted fusion is performed to obtain Then use residual convolution blocks to... Deep features are obtained by performing nonlinear reconstruction. Channel weights are generated through lightweight channel attention. Perform global calibration and then back-inject the calibrated reconstruction results as residuals. The final output is obtained from the process. : Here, α is constrained to the range of 0 to 1 by the Sigmoid function, and is used for automatic learning of spatial enhancement features. With original features The ratio between them; β is a learnable parameter, and σ(β) is used for adaptive control compensation intensity.

[0011] In one embodiment, the encoder of the segmentation network is a ConvNeXt-Tiny network, which extracts multi-scale features and then maps the features of each layer to a unified channel through convolution before inputting them into the decoder; the decoder is a U-Net++ structure, which achieves step-by-step fusion of multi-scale features through nested skip connections and sets up a deep supervision branch to assist training.

[0012] In one embodiment, during training, an equally weighted combination of BCE loss and Dice loss is used as the basic segmentation loss. Furthermore, a deep supervision branch is introduced in the intermediate layer of the decoder, and the total loss is a weighted sum of the main branch loss and the auxiliary branch loss: in, This indicates the main output prediction result. This indicates accurate labeling. Indicates the first The prediction results of each auxiliary branch, where λ represents the deep supervision weight coefficient, used to control the proportion of auxiliary branch loss in the total loss.

[0013] According to a second aspect of the present disclosure, a pathological myopic atrophy lesion segmentation system based on implicit dual-guided attention is provided, comprising: The data acquisition and preprocessing module acquires fundus images of pathological myopia and performs preprocessing. The multi-scale feature extraction module inputs the preprocessed pathological myopia fundus image into the segmentation network, extracts multi-scale features, and outputs the terminal feature map X. The implicit prior modeling module performs shared convolutional branch modeling on the terminal feature map X to generate a difficult attention map. and edge attention map ; The dual-guided attention calibration module uses the difficult attention map and the edge attention map as guidance signals to perform difficult-guided channel attention calibration and edge-guided spatial attention calibration on the features to obtain calibrated features; The results output module introduces a learnable scalar to perform weighted fusion and reconstruction of the calibrated features, compensates for information loss, and outputs the lesion segmentation results.

[0014] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the aforementioned method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention.

[0015] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the aforementioned method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention.

[0016] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. This invention employs shared feature branches to learn edge priors and difficult region priors, eliminating the need for additional labeled boundary information or complex post-processing steps. This mechanism effectively addresses common clinical scenarios such as blurred lesion boundaries, low contrast with the background, and noisy annotations, significantly reducing the model's dependence on high-quality labeled data and improving its applicability and generalization ability in real medical environments.

[0017] 2. This invention proposes a dual-guided attention calibration strategy. It utilizes the difficult-to-distinguish region guide channel attention to focus on lesion areas that are difficult to differentiate, while simultaneously using edge priors to guide spatial attention and enhance boundary feature responses. This targeted attention allocation method enables the network to more accurately capture lesion features, effectively suppress background interference, and thus significantly improve the ability to distinguish lesions from the background, reducing missed detections and false detections.

[0018] 3. To address the potential information loss caused by feature selection via attention mechanisms, this invention introduces a learnable scalar for adaptive feature enhancement and reconstruction. This intelligently compensates for key features weakened by attention weighting, effectively improving the continuity of responses within lesions, making segmentation boundaries more closely match the actual lesion contour, avoiding gaps or breaks in the segmentation results, and enhancing the visual quality and clinical usability of the segmentation results.

[0019] 4. This invention demonstrates excellent performance in the automatic segmentation of pathological myopic atrophy lesions, significantly improving key evaluation indicators such as crossover ratio, Dice coefficient, accuracy, and specificity. The high-precision automatic segmentation results provide doctors with reliable quantitative analysis data, assisting in clinical diagnosis and disease assessment, and possess promising clinical application prospects and widespread value. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0021] Figure 1 This is a schematic diagram of the segmentation network structure in this invention; Figure 2 This is a schematic diagram of the implicit dual-guided attention principle in this invention; Figure 3 This is a comparison chart of the qualitative segmentation results of the method of the present invention and different segmentation methods; Figure 4 This is a schematic diagram of the attention heatmap overlay result corresponding to the ablation experiment of this invention. Detailed Implementation

[0022] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0026] Example 1: This embodiment provides a method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention, including the following steps: S1. Acquire fundus images of pathological myopia and perform preprocessing; Specifically, this embodiment uses the publicly available dataset PALM, which contains 1200 color fundus images and provides information such as pathological myopia classification labels, optic disc annotations, foveal location, and lesion-level pixel annotations. Based on the annotations related to atrophic lesions, 870 image samples containing atrophic lesion annotations were selected and organized; these samples were then proportionally divided into a training set of 610 images, a validation set of 130 images, and a test set of 130 images. The input images were normalized and tensorized, and uniformly scaled to a resolution of 384×384; data augmentation was performed using random horizontal flipping, random vertical flipping, and random rotation to improve the model's generalization ability.

[0027] S2. Input the preprocessed pathological myopia fundus image into the segmentation network, extract multi-scale features and output the terminal feature map X; like Figure 1 As shown, the segmentation network IDGA-Net adopts an encoder-decoder architecture, specifically including a ConvNeXt-Tiny encoder, a U-Net++ decoder, and a terminal cascaded feature refinement module. The encoder extracts four levels of features from the input fundus image, with 96, 192, 384, and 768 channels respectively. First, a 3×3 convolution is used to uniformly map each level of features to 64 channels, and then the data is fed into the decoder for multi-scale feature fusion. The decoder uses a nested skip connection structure of U-Net++, with each fusion block having a uniform output channel of 64. The upsampling process uses bilinear interpolation, and a deep supervision mechanism is introduced in the intermediate layers. After decoding, the terminal feature map X is obtained.

[0028] S3. Model the terminal feature map X using shared convolutional branches to generate a difficult attention map. and edge attention map ; like Figure 2 As shown, for the terminal feature map X, intermediate features are first extracted through a shared branch. The shared branch consists of two 3×3 convolutional layers, reducing the number of channels to 16, setting the channel compression ratio to r=4, and introducing a random deactivation layer (Dropout) to mitigate overfitting. Based on intermediate features... Edge attention maps are generated using two independent prediction branches with different parameters. and difficult attention maps ,in Used to delineate the potential edge location of lesions. This is used to characterize regions with low contrast, complex textures, or those prone to misjudgment. Subsequently, the two priors are interacted element-wise and applied as residuals to the input feature X. Finally, after 1×1 convolution and batch normalization recalibration, the output feature is obtained. .

[0029] S4. Using the difficult attention map and the edge attention map as guiding signals, perform difficult-guided channel attention calibration and edge-guided spatial attention calibration on the features to obtain the calibrated features; like Figure 2 As shown, for features Global average pooling and global max pooling are performed separately, and then passed through a shared multilayer perceptron (MLP) to obtain the basic channel attention descriptor; at the same time, difficult priors are applied. The channel guidance vector is obtained through global average pooling and isomorphic mapping. This is combined with the basic channel descriptor to generate channel attention weights. Based on this, the features Weighted enhancement is performed to obtain channel enhancement features. .

[0030] Features along the channel dimension Average pooling and max pooling are performed separately and concatenated, then a 7×7 convolution is performed to obtain the basic spatial attention descriptor; at the same time, the edge prior is... Mapped to a spatial guiding graph via 7×7 convolution. This is combined with the basic spatial descriptor to generate spatial attention weights. Based on this, channel enhancement features are defined. Weighted augmentation is performed to obtain spatial augmentation features. .

[0031] S5. Introduce a learnable scalar to perform weighted fusion and reconstruction of the calibrated features, compensate for information loss, and output the lesion segmentation results.

[0032] Furthermore, learnable scalars are introduced. right and Weighted fusion is performed to obtain fusion features. ;Utilizing residual convolution blocks Nonlinear reconstruction is performed to obtain deep features. Then, channel weights are generated using lightweight channel attention. Perform global calibration and then back-inject the calibrated reconstruction results as residuals. To obtain the final output To improve the continuity and integrity of the lesion area; A weighted combination of BCE loss and Dice loss is used as the basic segmentation loss, and three deep supervision branches are introduced in the intermediate layer of the decoder. The total loss is a weighted average of the main branch loss and the auxiliary branch loss, with the deep supervision weight coefficient set to 0.4. The model encoder uses a ConvNeXt-Tiny network pre-trained on ImageNet; the optimizer is AdamW, with an initial learning rate of [missing value]. The weight decay is set to The first 10 epochs of training employed a linear warm-up strategy, followed by a cosine annealing strategy to gradually decay the learning rate. The training batch size is set to 24, and the total number of training rounds is set to 300.

[0033] The method of this invention was compared with U-Net, U-Net++, U-Net3+, AttentionU-Net, nnU-Net, DeepLabV3+, BASNet, TransUNet, Swin-Unet, ConvUNeXt, VM-UNet, MASUNet, MPBA-Net, and Swin-UMamba under the same data partitioning method and evaluation protocol. Evaluation metrics included IoU, Dice, Accuracy, Sensitivity, Specificity, Precision, HD95, and ASD. The experimental results are shown in Table 1. The results indicate that the method of this invention outperforms the method in terms of lesion region integrity, background false detection suppression, and overall segmentation stability.

[0034] Compared with existing segmentation methods, combining Figure 3 It can be seen that the method provided in this embodiment can obtain more stable segmentation results under weak boundary and complex background conditions, and performs well in key indicators such as IoU, Dice, Precision and Specificity, which can effectively improve the integrity of lesion area, boundary fit and background false detection suppression ability.

[0035] Table 1. Quantitative comparison results of the method of the present invention with different segmentation methods. As shown in Table 1, the method of the present invention achieves optimal values ​​in all five indicators: IoU, Dice, Accuracy, Specificity, and Precision, which are 0.8143, 0.8977, 0.9770, 0.9921, and 0.9357, respectively. This indicates that the method of the present invention has superior comprehensive performance in terms of lesion area integrity, false detection suppression, and overall segmentation stability.

[0036] To further verify the effectiveness of each functional module, ablation experiments were conducted under the same experimental settings. Using ConvNeXt+U-Net++ as the baseline model, explicit boundary and uncertainty branches, implicit edge / difficult attention, dual-guided CBAM, and adaptive feature enhancement modules were introduced sequentially. The ablation experiment results are shown in Table 2.

[0037] Table 2 Ablation Experiment Results of the Method of the Present Invention As shown in Table 2, the model performance continuously improves with the gradual introduction of implicit prior modeling, dual-guided attention calibration, and adaptive feature enhancement modules. The final model outperforms the baseline model in terms of IoU, Dice, Specificity, Precision, and Accuracy, validating the good complementarity and synergistic effect among the modules.

[0038] like Figure 4 The attention heatmap analysis shown further demonstrates that with the gradual addition of the three modules—implicit prior modeling, dual-guided attention calibration, and adaptive enhancement compensation—the model's ability to focus on lesion-related regions is continuously enhanced, and the background interference response is significantly reduced. This indicates that the method of this invention can achieve more stable and reliable lesion segmentation under complex backgrounds and weak boundary conditions.

[0039] Example 2: This embodiment provides a pathological myopic atrophy lesion segmentation system based on implicit dual-guided attention, including: The data acquisition and preprocessing module acquires fundus images of pathological myopia and performs preprocessing. The multi-scale feature extraction module inputs the preprocessed pathological myopia fundus image into the segmentation network, extracts multi-scale features, and outputs the terminal feature map X. The implicit prior modeling module performs shared convolutional branch modeling on the terminal feature map X to generate a difficult attention map. and edge attention map ; The dual-guided attention calibration module uses the difficult attention map and the edge attention map as guidance signals to perform difficult-guided channel attention calibration and edge-guided spatial attention calibration on the features to obtain calibrated features; The results output module introduces a learnable scalar to perform weighted fusion and reconstruction of the calibrated features, compensates for information loss, and outputs the lesion segmentation results.

[0040] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.

[0041] Example 3: An electronic device is provided for running the aforementioned "a method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention". The electronic device includes: a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S5 of the method described in Embodiment 1, specifically including but not limited to: S1. Acquire fundus images of pathological myopia and perform preprocessing; S2. Input the preprocessed pathological myopia fundus image into the segmentation network, extract multi-scale features and output the terminal feature map X; S3. Model the terminal feature map X using shared convolutional branches to generate a difficult attention map. and edge attention map ; S4. Using the difficult attention map and the edge attention map as guiding signals, perform difficult-guided channel attention calibration and edge-guided spatial attention calibration on the features to obtain the calibrated features; S5. Introduce a learnable scalar to perform weighted fusion and reconstruction of the calibrated features, compensate for information loss, and output the lesion segmentation results.

[0042] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage (acquiring observation data and outputting inversion results) through a communication interface. The above hardware configuration does not constitute a limitation of the present invention.

[0043] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S5 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.

[0044] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.

[0045] The above description is merely a preferred embodiment of this application and is not intended to limit 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 protection scope of this application.

[0046] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention, characterized in that, Includes the following steps: Acquire fundus images of pathological myopia and perform preprocessing; The preprocessed pathological myopia fundus image is input into the segmentation network, multi-scale features are extracted and the terminal feature map X is output; Model the terminal feature map X using shared convolutional branches to generate a difficult attention map. and edge attention map ; Using the difficult attention map and the edge attention map as guiding signals, the features are subjected to difficult-guided channel attention calibration and edge-guided spatial attention calibration to obtain the calibrated features; A learnable scalar is introduced to perform weighted fusion and reconstruction of the calibrated features, compensating for information loss and outputting lesion segmentation results.

2. The method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention as described in claim 1, characterized in that, Model the terminal feature map X using shared convolutional branches to generate a difficult attention map. and edge attention map Specifically: Model the terminal feature map X using shared convolutional branches to extract intermediate features. ; In intermediate features Based on this, edge attention maps are generated through two independent prediction branches. and difficult attention maps : in, This represents the Sigmoid activation function. This represents a 3×3 convolution operation in the edge branch. This represents a 1×1 convolution operation in the edge branch; This represents a 3×3 convolution operation in a difficult branch. This represents a 1×1 convolution operation in a difficult branch.

3. The method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention as described in claim 1, characterized in that, Difficult attention map and edge attention map Element-wise interactions are performed and applied to the terminal feature map X as residuals. After 1×1 convolution and batch normalization recalibration, the features are obtained. : in, Features after residual interaction; Features Perform channel attention calibration for difficult guidance and spatial attention calibration for edge guidance separately: In channel attention calibration, the difficult attention map is... Mapped to channel guiding vector The final channel attention weights are obtained by combining them with the basic channel descriptors. ; In spatial attention calibration, the edge attention map is... Convolutional mapping is used to create a spatial guiding graph. The final spatial attention weights are obtained by combining them with the basic spatial descriptor. .

4. The method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention according to claim 3, characterized in that, The channel attention weight Spatial attention weights The corresponding feature update relationship is represented as follows: in, This represents the Sigmoid activation function. The base channel descriptor is obtained by global average pooling and global max pooling. This is the basic space descriptor obtained by concatenating average pooling and max pooling; This is the feature map after channel attention enhancement. This enhances spatial features.

5. The method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention according to claim 1, characterized in that, A learnable scalar is introduced to perform weighted fusion and reconstruction of the calibrated features, specifically as follows: Introducing a learnable scalar α to enhance spatial features and characteristics Weighted fusion is performed to obtain Then use residual convolution blocks to... Deep features are obtained by performing nonlinear reconstruction. Channel weights are generated through lightweight channel attention. Perform global calibration and then back-inject the calibrated reconstruction results as residuals. The final output is obtained from the process. : Here, α is constrained to the range of 0 to 1 by the Sigmoid function, and is used for automatic learning of spatial enhancement features. With original features The ratio between them; β is a learnable parameter, and σ(β) is used for adaptive control compensation intensity.

6. The method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention according to claim 1, characterized in that, The encoder of the segmentation network is a ConvNeXt-Tiny network, which extracts multi-scale features and then maps the features of each layer into a unified channel through convolution before inputting them into the decoder. The decoder is a U-Net++ structure, which achieves step-by-step fusion of multi-scale features through nested skip connections and sets up a deep supervision branch to assist training.

7. The method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention according to claim 6, characterized in that, During training, an equal-weighted combination of BCE loss and Dice loss is used as the basic segmentation loss. Furthermore, a deep supervision branch is introduced in the intermediate layer of the decoder, and the total loss is a weighted sum of the main branch loss and the auxiliary branch loss: in, This indicates the main output prediction result. This indicates accurate labeling. Indicates the first The prediction results of each auxiliary branch, where λ represents the deep supervision weight coefficient, used to control the proportion of auxiliary branch loss in the total loss.

8. A pathological myopic atrophy lesion segmentation system based on implicit dual-guided attention, characterized in that, include: The data acquisition and preprocessing module acquires fundus images of pathological myopia and performs preprocessing. The multi-scale feature extraction module inputs the preprocessed pathological myopia fundus image into the segmentation network, extracts multi-scale features, and outputs the terminal feature map X. The implicit prior modeling module performs shared convolutional branch modeling on the terminal feature map X to generate a difficult attention map. and edge attention map ; The dual-guided attention calibration module uses the difficult attention map and the edge attention map as guidance signals to perform difficult-guided channel attention calibration and edge-guided spatial attention calibration on the features to obtain calibrated features; The results output module introduces a learnable scalar to perform weighted fusion and reconstruction of the calibrated features, compensates for information loss, and outputs the lesion segmentation results.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the method for segmenting pathological myopic atrophy lesions based on implicit dual-guided attention as described in any one of claims 1-7.

Images