An eye fundus image recognition lesion segmentation analysis system based on data enhancement

Abstract

This invention relates to the fields of medical image processing and artificial intelligence-assisted diagnosis, specifically to a data-augmented fundus image recognition lesion segmentation and analysis system. The system includes: an image acquisition module for acquiring the target fundus image to be analyzed and extracting its texture distribution features; a drift measurement module for calling a preset synthetic training domain feature distribution and generating a feature domain offset index from the homology distance; a feature decoupling module for dynamically adjusting the activation weights of the enhancement-sensitive feature channel and the topological feature channel in the segmentation network based on the feature domain offset index, generating an adaptive feature tensor; and a lesion segmentation module for performing pixel-level classification of the target fundus image based on the adaptive feature tensor, outputting a lesion segmentation mask that conforms to anatomical constraints. This invention effectively solves the problems of model aging and environmental adaptability during long-term operation.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and artificial intelligence-assisted diagnosis technology, specifically to a data-enhanced fundus image recognition lesion segmentation and analysis system. Background Technology

[0002] In the current field of fundus image-assisted diagnosis technology, in order to cope with the low-resource environment where high-quality labeled data is scarce, existing solutions generally adopt generative data augmentation technology to expand training samples in order to improve the detection sensitivity of the model for rare early lesions such as microaneurysms. However, such approaches have significant limitations when facing real clinical environments. Due to the lack of an effective mechanism for measuring feature domain shift in existing models, when the input image contains strong optical interference such as lens reflections and thermal noise from photosensitive elements introduced by portable devices, the model has difficulty distinguishing biological lesion features from non-biological artifacts, which can easily lead to model illusions and result in a high false positive rate. In addition, simple data augmentation fails to make full use of the anatomical topological constraints of the retina, causing the model to sacrifice specificity in pursuit of high detection rates and failing to ensure the anatomical rationality of the segmentation results under complex lighting conditions. Therefore, how to effectively curb misjudgments caused by optical noise while retaining the high sensitivity advantage of generative data augmentation for small lesions, and resolving the contradiction between high sensitivity and high anatomical specificity, has become an urgent technical problem to be solved. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a data-enhanced fundus image recognition lesion segmentation and analysis system. Specifically, the technical solution of this invention includes: The image acquisition module is used to acquire the target fundus image to be analyzed and extract the texture distribution features of the target fundus image; The drift measurement module is used to call the preset synthetic training domain feature distribution, calculate the homology distance between the texture distribution features and the synthetic training domain feature distribution, and generate the feature domain offset index from the homology distance. The feature decoupling module is used to dynamically adjust the activation weights of the enhancement-sensitive feature channels and the topological feature channels in the segmentation network based on the feature domain offset index, and generate an adaptive feature tensor. The lesion segmentation module is used to perform pixel-level classification of target fundus images based on adaptive feature tensors and output a lesion segmentation mask that conforms to anatomical constraints.

[0004] Preferably, a preset synthetic training domain feature distribution is invoked, the homology distance between the texture distribution features and the synthetic training domain feature distribution is calculated, and the homology distance is used to generate a feature domain offset index, including: The target fundus image is input into a pre-trained texture feature extraction network to obtain the current texture feature vector; Calculate the Mahalanobis distance between the current texture feature vector and the centroid of the synthetic training domain feature distribution; The Mahalanobis distance is mapped to a predetermined numerical range using a preset normalization function to generate a feature domain offset index. The feature domain offset index is used to numerically characterize the probability that the target fundus image contains non-biological artifacts or optical noise.

[0005] Preferably, based on the feature domain offset index, the activation weights of the enhanced sensitive feature channel and the topological feature channel in the segmentation network are dynamically adjusted to generate an adaptive feature tensor, including: setting a drift safety threshold; If the feature domain offset index is greater than the drift safety threshold, the activation weight of the enhanced sensitive feature channel is reduced and the activation weight of the topological feature channel is increased to generate an adaptive feature tensor biased towards structural consistency. If the feature domain offset index is less than or equal to the drift safety threshold, the activation weights of the enhanced sensitive feature channels are maintained or increased to generate an adaptive feature tensor that is biased towards capturing small lesions.

[0006] Preferably, the enhanced sensitive feature channel consists of convolutional layers trained based on generative data augmentation samples, used to extract nonlinear high-dimensional lesion features; The topological feature channel consists of convolutional layers trained based on anatomical morphology constraints, used to extract retinal vascular connectivity features and optic disc-optic cup geometric position features.

[0007] Preferably, the process of constructing the pre-defined synthetic training domain feature distribution includes: A synthetic fundus image set containing preset lesion types was generated using a probability diffusion model. Extract texture features from each image in the synthetic fundus image set; Cluster analysis is performed on all extracted texture features to construct a synthetic training domain feature distribution that characterizes the synthetic data augmentation domain.

[0008] Preferably, the state where the feature domain offset index is greater than the drift safety threshold corresponds to a scenario where there is adversarial optical noise in the target fundus image; Among them, adversarial optical noise includes lens reflection texture or photosensitive element thermal noise texture introduced by the acquisition equipment.

[0009] Preferably, based on adaptive feature tensors, pixel-level classification is performed on the target fundus image to output a lesion segmentation mask that conforms to anatomical constraints, including: The adaptive feature tensor is input into the decoder network to calculate the lesion probability map for each pixel. Morphological filters based on connected component analysis are applied to post-process the lesion probability map; Calculate the geometric features of each connected region, remove connected regions whose geometric features do not meet the preset retinal vessel topological continuity condition, and generate the final lesion segmentation mask.

[0010] Preferred options also include: The feedback correction module is used to analyze the historical distribution data of the feature domain offset index. Based on historical distribution data, the statistical boundary parameters of the synthetic training domain feature distribution are adjusted according to a preset update cycle to calibrate the sensitivity of the drift measurement module.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention calculates the homology distance between texture distribution features and the synthetic training domain through a drift measurement module, quantifies the degree of feature domain offset, and dynamically adjusts the activation weights of the enhanced sensitive channel and the topological structure channel accordingly. When strong optical interference is detected, the system automatically reduces the sensitivity to high-frequency noise and enhances structural constraints, effectively suppressing model illusions caused by lens reflection or thermal noise, and significantly reducing the false positive rate under complex lighting conditions. 2. This invention utilizes a feature decoupling mechanism to resolve the contradiction between the high sensitivity brought by generative data augmentation and the high specificity brought by anatomical topological constraints. The system maintains a high detection advantage for rare lesions such as microaneurysms in clean images and automatically switches to an analysis mode that favors structural consistency in noisy images. Thus, in a low-resource environment where high-quality labeled data is lacking, it balances the comprehensiveness of lesion identification with the accuracy of diagnostic results. 3. This invention introduces a post-processing verification mechanism based on anatomical morphology constraints. By calculating the geometric morphological features of connected regions and their topological continuity with retinal vessels, artifacts with abnormal shapes or positions that do not conform to the vascular symbiosis rules are automatically eliminated. This ensures that the output lesion segmentation mask strictly meets biological rationality in terms of topological structure, eliminates interference from non-biological features, and improves the clinical credibility of the auxiliary diagnostic results. 4. This invention uses a feedback correction module to statistically analyze historical distribution data, which can dynamically adjust the statistical boundary parameters of the synthetic training domain according to a preset period. This self-evolutionary mechanism enables the system to automatically adapt to fluctuations in basic noise caused by equipment aging or changes in the acquisition environment, avoiding long-term decay or failure of drift measurement sensitivity, and effectively solving the problems of model aging and environmental adaptability during long-term operation. Attached Figure Description

[0012] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] Example 1:

[0015] Please see Figure 1 A data-enhanced fundus image recognition lesion segmentation and analysis system includes: an image acquisition module for acquiring the target fundus image to be analyzed and extracting the texture distribution features of the target fundus image; The drift measurement module is used to call the preset synthetic training domain feature distribution, calculate the homology distance between the texture distribution features and the synthetic training domain feature distribution, and generate the feature domain offset index from the homology distance. The feature decoupling module is used to dynamically adjust the activation weights of the enhancement-sensitive feature channels and the topological feature channels in the segmentation network based on the feature domain offset index, and generate an adaptive feature tensor. The lesion segmentation module is used to perform pixel-level classification of target fundus images based on adaptive feature tensors and output a lesion segmentation mask that conforms to anatomical constraints.

[0016] This embodiment provides a data-augmented fundus image recognition lesion segmentation and analysis system. The system aims to resolve the contradiction between the high sensitivity brought by generative data augmentation and the high specificity brought by anatomical topological constraints in a low-resource environment where high-quality labeled data is scarce. The system acquires the target fundus image to be analyzed through the image acquisition module, such as the retinal RGB image acquired by a portable fundus camera. The core of this module integrates a texture distribution feature extraction unit, which is used to extract the statistical expression of the image in the feature space, that is, to focus on the underlying signal patterns such as high-frequency noise distribution and illumination gradient continuity. The drift measurement module, acting as the system's gatekeeper, calls upon a preset synthetic training domain feature distribution, calculates the homology distance of the aforementioned texture distribution features relative to this distribution, and defines this distance as the feature domain offset index. In the specific implementation of this embodiment, in order to adapt to the subsequent dynamic weight adjustment logic, the index is specifically defined as the probability metric value after nonlinear normalization of the original distance, with a value range of [0,1], thereby eliminating the computational uncertainty caused by the original distance value range of [0,+\infty]. Its source is the drift metric module calculation; its physical meaning is the degree to which the feature pattern of the current input image deviates from the model training experience, and it is dimensionless. The feature decoupling module acts as the decision-making center, based on the feature domain offset index. The dynamic weight redistribution is performed to adjust the activation weights of the enhancement sensitive feature channel and the topological feature channel in the segmentation network, generating an adaptive feature tensor. Based on the recombined adaptive feature tensor, the lesion segmentation module performs pixel-level classification on the target fundus image and outputs a lesion segmentation mask that not only marks the location of the lesion but also passes the anatomical rationality check. This embodiment introduces a drift metric and feature decoupling mechanism, which retains the advantage of generative data augmentation in high detection rate of rare lesions such as early microaneurysms, while effectively curbing model illusions and false positives caused by noise from acquisition equipment such as lens reflections. To verify the above technical effects, quantitative experiments were conducted on a test set containing 500 fundus images, 200 of which contained strong lens reflections and thermal noise. Experimental results show that, compared with the baseline model that does not incorporate drift measurement and dynamic weighting mechanism, the performance of this system is significantly improved in the strong optical interference test set: the Dice coefficient for lesion segmentation increased from 0.74 to 0.86, and the IoU value increased from 0.68 to 0.79; especially in terms of misdiagnosis rate control, the false positive rate decreased from 15.3% to 2.1%, and the data shows that this system significantly reduced the misdiagnosis rate in the strong optical interference test set.

[0017] Example 2:

[0018] In this embodiment, a preset synthetic training domain feature distribution is invoked, the homology distance between the texture distribution features and the synthetic training domain feature distribution is calculated, and the homology distance is used to generate a feature domain offset index. This includes: inputting the target fundus image into a pre-trained texture feature extraction network to obtain the current texture feature vector; calculating the Mahalanobis distance between the current texture feature vector and the centroid of the synthetic training domain feature distribution; and mapping the Mahalanobis distance to a predetermined numerical range using a preset normalization function to generate a feature domain offset index. The feature domain offset index is used to numerically characterize the probability that the target fundus image contains non-biological artifacts or optical noise.

[0019] This embodiment details the specific implementation process of calculating the feature domain offset index in the drift measurement module; the system inputs the target fundus image into a pre-trained texture feature extraction network, which is a lightweight variant with the fully connected layers removed, focusing on extracting texture rather than semantics, thereby obtaining the current texture feature vector V; Specifically, this texture feature extraction network is built on the ResNet-18 architecture. The standard output channel count of its Layer 3 (conv4_x) is defined as 256. To reduce computational redundancy and focus on mid-to-low-level texture features, the network removes Layer 4 (conv5_x) and subsequent fully connected layers from the original structure, retaining only Layer 3 (conv4_x). The input image size is adjusted to 224x224, and after convolutional processing, the output feature map has a size of 14x14x256. A global average pooling layer compresses the feature map into a current texture feature vector V with a dimension of 256. The system calculates the Mahalanobis distance d between the current texture feature vector V and the centroid of the synthetic training domain feature distribution. The Mahalanobis distance is chosen because it considers the correlation and scale differences between the dimensions of the features. The calculation formula is as follows:

[0020] Where V is the current texture feature vector, sourced from the real-time output of the texture feature extraction network; m is the centroid vector of the feature distribution in the synthetic training domain, sourced from the statistical calculations during the offline training phase of the system, i.e., the global mean obtained by aggregating the clustering results according to Example 5, to match the single vector input requirement; s is the covariance matrix of the feature distribution in the synthetic training domain, sourced from the statistical calculations during the offline training phase of the system, i.e., the global covariance matrix obtained by aggregating the clustering results; the superscript -1 in the formula indicates matrix inversion, and the superscript T indicates matrix transpose; the Mahalanobis distance d is mapped to a predetermined numerical range using a preset normalization function. Generate feature domain offset index The calculation formula is as follows:

[0021] Where e is a natural base constant, with a value of approximately 2.71828; The slope adjustment parameter is derived from ROC curve analysis on the validation set, and its physical meaning is to control the response speed of the exponent to changes in distance; although the Mahalanobis distance d is statistically defined as a dimensionless value, here... It is still used as a dimensionless scaling factor to adjust the activation steepness of the Sigmoid function, ensuring that the mapping curve has appropriate discriminative power within the effective distance range; The distance baseline bias is derived from the 95th percentile value of the Mahalanobis distance of normal samples in the synthetic training domain; This embodiment, by employing Mahalanobis distance and nonlinear mapping, can accurately transform outliers in a high-dimensional feature space into an intuitive probability index, quantifying the likelihood that the current image contains non-biological features that the model has not learned, and providing a reliable numerical basis for subsequent weight adjustments.

[0022] Example 3:

[0023] In this embodiment, based on the feature domain offset index, the activation weights of the enhancement-sensitive feature channel and the topology feature channel in the segmentation network are dynamically adjusted to generate an adaptive feature tensor. This includes: setting a drift safety threshold; if the feature domain offset index is greater than the drift safety threshold, then the activation weights of the enhancement-sensitive feature channel are reduced and the activation weights of the topology feature channel are increased to generate an adaptive feature tensor biased towards structural consistency; if the feature domain offset index is less than or equal to the drift safety threshold, then the activation weights of the enhancement-sensitive feature channel are maintained or increased to generate an adaptive feature tensor biased towards the ability to capture small lesions; the enhancement-sensitive feature channel is composed of convolutional layers trained based on generative data augmentation samples, used to extract nonlinear high-dimensional lesion features; The topological feature channel consists of convolutional layers trained based on anatomical morphology constraints, used to extract retinal vascular connectivity features and optic disc-optic cup geometric position features; the state where the feature domain offset index is greater than the drift safety threshold corresponds to a scenario where there is adversarial optical noise in the target fundus image; among which, adversarial optical noise includes lens reflection texture or photosensitive element thermal noise texture introduced by the acquisition device.

[0024] This embodiment describes in detail how to use the feature domain offset index. The weights of different channels in the segmentation network are dynamically adjusted. The encoder part of the segmentation network adopts a two-stream architecture, which includes two types of parallel feature channels: Enhanced sensitive feature channel stream A: composed of 3 stacked residual convolutional blocks, used to extract nonlinear high-dimensional lesion features; its training method is as follows: construct an enhanced dataset containing synthetic lesions such as microaneurysms and bleeding points generated by a probability diffusion model, and train it using weighted cross-entropy loss, where the weight of lesion pixels is set to 10 times that of background pixels, thereby giving this channel high sensitivity to small, nonlinear lesion features; The topological feature channel B consists of three structurally identical but parameter-independent residual convolutional blocks used to extract retinal vascular connectivity features and optic disc-optic cup geometric location features. Its training method involves using only a noise-free real dataset with rigorously annotated anatomical structures by physicians, and jointly optimizing it using the clDice loss function and the Hausdorff distance loss function to force this channel to learn the connectivity of retinal vessels and the geometric location features of the optic disc. Specifically, the system constructs the total loss function using a weighted summation method. To address the inconsistency in the dimensions of probability and distance metrics, the calculation formula is as follows:

[0025] in, For the calculation Loss value; The calculated Hausdorff distance loss value; Represents the natural logarithm function; The topological continuity weight is set to 1.0; The geometric distance weight is set to 0.05; The numerical stability constant has a value of 1e-6; a logarithmic function is introduced here. and small weight The aim is to suppress the excessive dominance of Hausdorff distance on gradients and ensure that the model can learn the connectivity and geometric location features of blood vessels in a balanced way. Based on this, the system sets a drift safety threshold. Specifically, The value range is [0.35, 0.45], with a preferred value of 0.4. This threshold is obtained by: statistically analyzing the probability distribution of the feature domain offset index on a validation set of 1000 clean fundus images confirmed by doctors to be free of obvious artifacts, and selecting the 95th percentile of its cumulative distribution function (CDF) as... This ensures that 95% of high-quality images are classified as safe mode, avoiding accidental triggering of defense mechanisms; in response to the feature domain offset index. Greater than the drift safety threshold The system determines that adversarial optical noise exists in the target fundus image, including lens reflection textures or thermal noise textures introduced by the acquisition device. In this high-risk mode, the system implements a conservative defense strategy, reducing the activation weights of the enhanced sensitive feature channels. And increase the activation weights of the topological feature channels. The specific weight calculation is as follows:

[0026] Where K is the adjustment gain, defined as a dimensionless scalar with a value range of [1.8, 2.2], and a value of 2.0 is chosen to match... and The dimensionless property of the weights ensures that the calculated weights are accurate. The value is dimensionless; this parameter is determined through grid search; in order to follow the general evaluation principles of machine learning and prevent data leakage, the search process is strictly performed on a noisy validation set, not a test set; the noisy validation set is constructed by adding simulated optical noise to the original validation set samples; the search objective is to maximize the second derivative of the false positive rate reduction curve in the noisy validation set, that is, the inflection point value that achieves the best balance between suppressing noise and preserving lesions; It should be noted that, given that The theoretical upper limit is 1.0, when the input image undergoes severe drift, such as... Directly substituting into the above formula may lead to... For example, 1.7, and For example, -0.7; to avoid the collapse of physical meaning and numerical calculation anomalies caused by negative weights, the system introduces a numerical truncation mechanism; after calculating the initial... Then, apply the following constraint function, the calculation formula of which is:

[0027] Based on the truncated renew This step ensures that the weights are always constrained within the closed interval [0,1], meaning that under extreme noise conditions, the system will completely suppress the enhancement-sensitive feature channels. Only the topological feature channels are retained, i.e. ; When generating the adaptive feature tensor, the feature decoupling module uses a channel attention mechanism to apply the calculated weights... and The feature maps that are applied to the outputs of streams A and B respectively, i.e. Regarding the issue of unclear skip connection construction in the U-Net architecture pointed out during the review, this embodiment explicitly adopts a homologous dynamic weighted skip connection strategy: for the Kth layer skip connection between the encoder and decoder, the system also extracts the Kth layer features of the enhanced sensitive feature channel. With the Kth layer features of the topological structure feature channel And reuse the global weights calculated in the current frame. and Perform linear weighted fusion to generate skip connection features This feature is concatenated into the upsampled feature map of the Kth layer of the decoder. This design ensures that the network also has the ability to dynamically suppress optical noise on shallow texture transmission paths, and constructs a complete full-scale adaptive segmentation topology. Response to feature domain offset index Less than or equal to the drift safety threshold The system determines that the image quality conforms to the training domain distribution and maintains or increases the activation weights of the enhanced sensitive feature channels. For this branch, the system uses a deterministic linear recovery formula to calculate the weights to avoid logical uncertainty. The formula is as follows:

[0028] in, The linear restoration coefficient has a value range of [1.0, 1.5], with a preferred value of 1.25. This controls the image quality as it moves from the critical point towards a perfect state. During the transition, the weights are biased towards increasing the rate of the sensitive channel; when That is, when the image quality perfectly matches the training domain. The system is in full enhancement sensitivity mode to maximize the detection rate of small lesions; when That is, when the image quality approaches the critical point, The high-drift branch achieves numerical smoothing at the threshold boundary, ensuring the stability of the system during mode switching; This embodiment implements a soft switching mechanism. When faced with clean images, the system fully utilizes the powerful learning capabilities of AI to detect early lesions. However, when faced with dirty images containing reflections or noise, the system automatically degenerates into a more conservative, anatomically based analysis mode. This dynamic game significantly improves the robustness of the system in complex clinical environments, especially in primary healthcare institutions.

[0029] Example 4: In this embodiment, the process of constructing the preset synthetic training domain feature distribution includes: generating a set of synthetic fundus images containing preset lesion types using a probability diffusion model; extracting texture features from each image in the synthetic fundus image set; performing cluster analysis on all extracted texture features to construct a synthetic training domain feature distribution that characterizes the synthetic data augmentation domain.

[0030] This embodiment describes the construction process of the baseline source of the drift measurement module, namely the synthetic training domain feature distribution; a synthetic fundus image set containing preset lesion types is generated using a probability diffusion model. This step artificially creates long-tailed distribution data that is difficult to collect in the real world, thus enriching the diversity of training samples. Using the same texture feature extraction network as the system in operation, texture feature vectors are extracted from each image in the synthetic fundus image set. Cluster analysis is performed on all extracted texture features, for example, Gaussian Mixture Model (GMM) is used to identify the multimodal distribution structure in the feature space. In response to the issue pointed out in the review that direct aggregation of multimodal features would cause the mean to fall into the low-density region, this step abandons full covariance aggregation and instead adopts a principal mode locking strategy: among the multiple data clusters identified by GMM, the unique main peak representing the main synthetic data distribution is identified according to the sample size weight, and other low-weight small clusters are regarded as outliers in the generation process and are removed. The constructed synthetic training domain feature distribution is ultimately uniquely determined by the single mean vector m and single covariance matrix s corresponding to the main peak. This process mathematically guarantees that the unimodal Gaussian assumption of Mahalanobis distance calculation holds, ensuring that the metric benchmark is located at the geometric center of the effective data cluster, rather than an invalid gap between multiple clusters, thereby achieving accurate quantification of feature offset. This embodiment constructs a baseline distribution based on synthetic data, which in effect defines an AI cognitive boundary. Any image falling outside this boundary is considered a potential adversarial attack or an unknown domain, thus triggering a defense mechanism. This solves the problem that traditional methods rely solely on limited real data to establish boundaries, resulting in weak generalization ability.

[0031] Example 5: In this embodiment, the target fundus image is classified at the pixel level based on the adaptive feature tensor, and a lesion segmentation mask that conforms to anatomical constraints is output. This includes: inputting the adaptive feature tensor into the decoder network to calculate the lesion probability map for each pixel; applying a morphological filter based on connected component analysis to post-process the lesion probability map; calculating the geometric morphological features of each connected component, removing connected components whose geometric morphological features do not meet the preset retinal vessel topological continuity condition, and generating the final lesion segmentation mask.

[0032] This embodiment details the process of generating the final lesion segmentation mask based on adaptive feature tensors; the adaptive feature tensors are input into the decoder network, spatial resolution is restored through upsampling, the probability of each pixel belonging to the lesion category is calculated, and a lesion probability map is generated; a morphological filter based on connected component analysis is applied to post-process the lesion probability map, including binarization and identification of all independent connected regions. Based on this, the geometric morphological features and topological invariant features of each connected component are calculated, and connected components that do not meet the preset retinal vessel topological continuity condition are removed. The retinal vessel topological continuity condition is specifically defined in this embodiment as follows: the connected component must be spatially located within the blood supply topological neighborhood of the vessel, constrained by a distance threshold, and topologically should be a simply connected closure, constrained by the Euler number. The verification criteria are as follows, and their calculation formula is:

[0033] Where c is the validity indicator of the connected component, 1 for retention and 0 for rejection; p is the perimeter of the connected component, which is derived from the boundary pixel statistics; a is the area of ​​the connected component, which is derived from the total number of internal pixels. The shape factor threshold is 1.2; this value is derived from the statistical analysis of the shape factors of connected components labeled as microaneurysms in the training set. (The formula is missing from the original text.) The item is the roundness index, where, Pi is a constant, and its physical meaning is to measure how close the shape of the boundary of a connected domain is to a standard circle; it is obtained by selecting the 99th percentile value of its distribution and rounding it down with a margin of redundancy. This threshold is used to strictly distinguish approximately circular microaneurysms, i.e., shape factors. This is in contrast to elongated scratches or irregular optical noise, i.e., shape factor. This ensures the morphological accuracy of lesion screening; The distance between this connected component and its nearest neighbor vessel is calculated using distance transformation. The threshold for blood vessel companion distance is physically defined as the anatomical rule that lesions tend to appear near blood vessels. In this embodiment, it is set to 3 times the average blood vessel diameter. The average blood vessel diameter is a global average width constant obtained by offline statistical analysis of all blood vessel labeled samples in the training domain dataset. The unit is pixels, for example, it is set to 5 pixels. Using offline statistical constants instead of real-time measurements is to avoid the risk of errors in real-time blood vessel diameter calculation due to lesion occlusion or poor image quality, thereby ensuring the executability and stability of the threshold logic; E is the Euler number of the connected domain, used to characterize the topological properties, that is, to describe its structure by the difference between the number of connected components and the number of holes of the calculated object. For a single connected region without holes, E=1 to ensure that there are no unnatural topological holes inside the lesion; finally, the region with c=1 is retained to generate the final lesion segmentation mask; This embodiment constitutes a line of defense for the system. Even if the feature decoupling at the front end fails to completely filter out noise, the topology verification at the back end can forcibly remove artifacts that do not look like diseases, are located in impossible positions, or have abnormal topological structures based on the anatomical common sense of shape, position, and topology, thereby ensuring the medical rationality of the output results.

[0034] Example 6: In this embodiment, it further includes: a feedback correction module, used to collect historical distribution data of the statistical feature domain offset index; and based on the historical distribution data, to adjust the statistical boundary parameters of the synthetic training domain feature distribution according to a preset update cycle, so as to calibrate the sensitivity of the drift measurement module.

[0035] This embodiment describes the system's feedback correction module, which aims to achieve adaptive evolution of the model. During system operation, the feedback correction module continuously records and calculates the feature domain offset index. Historical distribution data is used, such as the index values ​​of all screened images recorded in past periods. Based on this historical distribution data, the statistical boundary parameters of the synthetic training domain feature distribution are adjusted according to a preset update cycle, i.e., the distribution centroid m and covariance matrix s are updated to calibrate the sensitivity of the drift measurement module. The parameter update formula is as follows:

[0036] in, The updated centroid of the distribution; The centroid of the current distribution; The mean of the real sample features observed recently; K is the dynamic momentum update factor, which aims to dynamically adjust the update step size according to the sample size of the current observation data, and its calculation formula is:

[0037] in, Let N be the number of samples in the current observation batch. This represents the total number of samples processed historically by the system. To prevent update stagnation after long-term operation, a threshold of 0.01 is set for the minimum value of K, meaning the final value is [value to be filled in]. ; This dynamic setting ensures that the update weights of the covariance matrix are proportional to the reliability of the observed samples, thus avoiding drastic disturbances to the global statistical distribution caused by small batches of noisy data. Furthermore, to fully calibrate the Mahalanobis distance metric space, the system synchronously updates the covariance matrix s online. Since Mahalanobis distance is highly dependent on the discrete shape of the feature distribution, updating only the centroid cannot adapt to changes in environmental noise variance. Therefore, the following modified momentum-based covariance update formula is adopted, explicitly introducing the variance component within the observation batch to correct the logical defect in the original formula that lacks intra-batch variance. The calculation formula is as follows:

[0038] in, This is the updated covariance matrix, used for Mahalanobis distance calculation in the next period; This is the covariance matrix currently in use; The covariance matrix of the current batch of observed sample features is calculated using the following formula:

[0039] Where the superscript T denotes the matrix transpose operation, Used to capture the data dispersion in the current environment; This represents the total number of samples in the current observation batch, used for normalization to perform unbiased estimation of the covariance calculation; The texture feature vector of the i-th sample in the current observation batch is used as the basic data unit for covariance calculation; the term in the formula This is the additional variance term generated by mean shift; the formula fully preserves the second-order statistical properties of the data, preventing the covariance matrix from numerically collapsing as the number of iterations increases; This embodiment updates the statistical boundary slowly and synchronously, enabling the system to automatically adapt to the overall increase in the baseline noise level caused by equipment aging or changes in lighting conditions in different hospitals. This avoids excessive sensitivity decay caused by long-term artificially high exponents due to environmental drift, thus solving the problems of model aging and environmental adaptability.

[0040] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A data-augmented fundus image recognition lesion segmentation and analysis system, characterized in that, include: The image acquisition module is used to acquire the target fundus image to be analyzed and extract the texture distribution features of the target fundus image; The drift measurement module is used to call a preset synthetic training domain feature distribution, calculate the homology distance between the texture distribution features and the synthetic training domain feature distribution, and generate a feature domain offset index based on the homology distance. The feature decoupling module is used to dynamically adjust the activation weights of the enhancement sensitive feature channel and the topology feature channel in the segmentation network according to the feature domain offset index, and generate an adaptive feature tensor. The lesion segmentation module is used to perform pixel-level classification on the target fundus image based on the adaptive feature tensor and output a lesion segmentation mask that conforms to anatomical constraints.

2. The fundus image recognition lesion segmentation and analysis system based on data augmentation according to claim 1, characterized in that, Calling a preset synthetic training domain feature distribution, calculating the homology distance between the texture distribution features and the synthetic training domain feature distribution, and generating a feature domain offset index based on the homology distance, including: The target fundus image is input into a pre-trained texture feature extraction network to obtain the current texture feature vector; Calculate the Mahalanobis distance between the current texture feature vector and the centroid of the synthetic training domain feature distribution; The Mahalanobis distance is mapped to a predetermined numerical range using a preset normalization function to generate the feature domain offset index, wherein the feature domain offset index is used to numerically characterize the probability that the target fundus image contains non-biological artifacts or optical noise.

3. The fundus image recognition lesion segmentation and analysis system based on data augmentation according to claim 2, characterized in that, Based on the feature domain offset index, the activation weights of the enhancement-sensitive feature channels and the topological feature channels in the segmentation network are dynamically adjusted to generate an adaptive feature tensor, including: Set a drift safety threshold; If the feature domain offset index is greater than the drift safety threshold, then the activation weight of the enhanced sensitive feature channel is reduced and the activation weight of the topology feature channel is increased to generate the adaptive feature tensor biased towards structural consistency. If the feature domain offset index is less than or equal to the drift safety threshold, the activation weight of the enhanced sensitive feature channel is maintained or increased to generate the adaptive feature tensor that is biased towards the ability to capture small lesions.

4. The fundus image recognition lesion segmentation and analysis system based on data augmentation according to claim 3, characterized in that, The enhanced sensitive feature channel consists of convolutional layers trained based on generative data augmentation samples, used to extract nonlinear high-dimensional lesion features; The topological feature channel consists of convolutional layers trained based on anatomical morphology constraints, used to extract retinal vascular connectivity features and optic disc-optic cup geometric position features.

5. The fundus image recognition lesion segmentation and analysis system based on data augmentation according to claim 4, characterized in that, The process of constructing the preset synthetic training domain feature distribution includes: A synthetic fundus image set containing preset lesion types was generated using a probability diffusion model. Extract the texture features of each image in the synthesized fundus image set; Cluster analysis is performed on all extracted texture features to construct the feature distribution of the synthetic training domain that characterizes the synthetic data augmentation domain.

6. The fundus image recognition lesion segmentation and analysis system based on data augmentation according to claim 3, characterized in that, The state in which the feature domain offset index is greater than the drift safety threshold corresponds to a scenario in which there is adversarial optical noise in the target fundus image; The aforementioned adversarial optical noise includes lens reflection textures or photosensitive element thermal noise textures introduced by the acquisition device.

7. The fundus image recognition lesion segmentation and analysis system based on data augmentation according to claim 1, characterized in that, Based on the adaptive feature tensor, pixel-level classification is performed on the target fundus image, outputting a lesion segmentation mask that conforms to anatomical constraints, including: The adaptive feature tensor is input into the decoder network to calculate the lesion probability map for each pixel. The lesion probability map is post-processed using a morphological filter based on connected component analysis. Calculate the geometric features of each connected region, remove connected regions whose geometric features do not meet the preset retinal vessel topological continuity condition, and generate the final lesion segmentation mask.

8. The fundus image recognition lesion segmentation and analysis system based on data augmentation according to claim 1, characterized in that, Also includes: The feedback correction module is used to statistically analyze the historical distribution data of the feature domain offset index; Based on the historical distribution data, the statistical boundary parameters of the synthetic training domain feature distribution are adjusted according to a preset update cycle to calibrate the sensitivity of the drift measurement module.

Images