Retinal image abnormality intelligent recognition system based on ophthalmic auxiliary diagnosis

Abstract

The application discloses an abnormal intelligent identification system for retinal images based on auxiliary diagnosis of ophthalmology, belongs to the technical field of medical image processing and analysis, and is characterized in that: a plurality of groups of images of the same eye to be examined are collected to form an analysis sequence, an anatomical structure feature map and a tissue texture feature map of each frame of image in the sequence are extracted in parallel, conditional normal texture construction and intra-group comparison, multi-frame texture consistency verification and anatomical structure deviation quantitative analysis are performed on the anatomical structure feature map and the tissue texture feature map, a texture abnormality intensity map, a texture space-time variation map, a texture region stability classification map and a structure deviation degree map are generated, and the above multiple results are adaptively weighted and fused and graded. The application is characterized in that: a collaborative technical route of multi-modal imaging, double-feature decoupling extraction, conditional normal comparison, space-time consistency verification and adaptive fusion decision is adopted, the dependence on labeled data is reduced, and high-robustness intelligent identification of structure and texture abnormalities in retinal images is realized.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and analysis technology, and more specifically, to an intelligent recognition system for retinal image abnormalities based on ophthalmic auxiliary diagnosis. Background Technology

[0002] With the in-depth application of artificial intelligence technology in the field of medical image analysis, automatic analysis of retinal images (such as fundus color photos and optical coherence tomography images) based on deep learning has become a research hotspot in ophthalmic auxiliary diagnosis. The goal is to automatically identify pathological changes in images, such as diabetic retinopathy, glaucoma, and age-related macular degeneration, in order to improve screening efficiency and assist clinical decision-making.

[0003] Conventional technical solutions mainly follow the following two paradigms:

[0004] Supervised end-to-end disease classification models: These methods rely on large-scale, finely labeled disease image datasets to train deep convolutional neural networks to directly output specific disease labels or severity ratings;

[0005] The core limitations of this paradigm are: first, performance is severely constrained by the scale and quality of labeled data, and obtaining medical expert-level labels is costly and time-consuming; second, the model is essentially learning the distribution of existing disease labels, and lacks the ability to identify unknown abnormality types that are not present or rare in the training data, thus limiting its generalization ability.

[0006] Unsupervised or self-supervised anomaly detection models: To reduce reliance on anomaly-labeled data, these methods attempt to define a “normal” pattern by learning the feature distribution of a large number of normal retinal images, and then classify samples that deviate from this pattern as anomalies.

[0007] While this paradigm reduces reliance on anomaly annotations, it still has the following shortcomings: general image reconstruction models struggle to incorporate prior knowledge of retinal anatomy (such as vascular tree morphology and the anatomical location of the optic disc and macula), leading to insensitivity to structural aberrations (such as enlarged optic disc cupping) or false positives for normal anatomical variations. This is because there are significant individual physiological differences in retinal anatomy, and such retrieval schemes may misjudge normal structural variations as texture abnormalities, resulting in low specificity.

[0008] Furthermore, the two paradigms mentioned above are mostly used for analyzing single static images, and cannot utilize the dynamic information and consistency constraints contained in multi-view image sequences collected under different gaze directions. Therefore, they are easily affected by noise and artifacts in single-frame images, and their specificity needs to be improved.

[0009] In summary, the core challenge facing current retinal image anomaly recognition technology lies in how to build an intelligent recognition system that can simultaneously and sensitively perceive structural distortions and subtle texture changes, and provide anatomically interpretable results, while reducing reliance on large-scale anomaly annotation data. Summary of the Invention

[0010] The purpose of this invention is to solve the existing problems and provide an intelligent recognition system for retinal image abnormalities based on ophthalmological auxiliary diagnosis, compared with existing technologies.

[0011] The objective of this invention can be achieved through the following technical solution: an intelligent retinal image abnormality recognition system based on ophthalmic auxiliary diagnosis, comprising: an image acquisition and processing module, used to control the imaging device to acquire multiple sets of retinal images of the same examined eye when performing at least two different preset eye movements, and to perform standardized preprocessing on the multiple sets of retinal images to form an analysis sequence;

[0012] The dual-path feature decoupling extraction module is used to receive the analysis sequence, synchronously and in parallel input the pre-trained and optimized anatomical structure feature extraction branch and tissue texture feature extraction branch, and output the anatomical structure feature map and tissue texture feature map of each frame of the analysis sequence, forming the anatomical structure feature map sequence and the tissue texture feature map sequence.

[0013] The multiple alignment analysis module is used to receive the anatomical structure feature map sequence and the original tissue texture feature map sequence, and perform conditional normal texture construction and intra-group alignment analysis, inter-frame texture consistency verification analysis and anatomical structure deviation quantification analysis. Based on the results of the multiple alignment analysis, it generates intra-frame texture anomaly intensity map, inter-frame texture runaway variation map, texture region stability classification map and structural deviation map respectively.

[0014] The information fusion and verification module is used to receive multi-class analysis result data sent from the multi-comparison analysis module, perform adaptive weighted fusion to obtain a comprehensive abnormal intensity map, and perform collaborative analysis based on the comprehensive abnormal intensity map, texture region stability classification map, and structural deviation map to classify and locate lesions.

[0015] Furthermore, the anatomical structure feature extraction branch is constructed based on a pre-trained retinal image segmentation network to extract anatomical structure features including blood vessel morphology, optic disc geometric parameters, and foveal position. The tissue texture feature extraction branch is trained through self-supervised learning to extract texture features that are decoupled from anatomical structures and represent tissue texture, pigment, reflection, hemorrhage, and pigmentation.

[0016] Furthermore, the specific process of conditional normal texture construction and intra-group comparison analysis includes:

[0017] Using the anatomical structure feature map of the t-th frame in the anatomical structure feature map sequence as conditional input, a pre-trained conditional generation network is used to reconstruct the predicted normal tissue texture feature map corresponding to the current anatomical structure state. The pixel-wise difference between the tissue texture feature map of this frame and the predicted normal tissue texture feature map is calculated to obtain the original difference map.

[0018] The original difference map is passed through a 1×1 convolutional layer and a sigmoid activation function to map it into an intra-frame texture anomaly intensity map with pixel values ​​in the range [0,1].

[0019] Furthermore, the specific process of multi-frame texture consistency verification analysis includes:

[0020] Cross-frame collaborative alignment of the tissue texture feature map sequence is performed, and the texture feature similarity between any two frames in the sequence at corresponding spatial locations is calculated to obtain multiple inter-frame similarity matrices.

[0021] Take the median, mean, or other aggregated statistics from multiple inter-frame similarity matrices to generate an inter-frame texture runaway variation map that characterizes the overall variation intensity of texture at each location.

[0022] Based on multiple inter-frame similarity matrices, and according to preset thresholds and consecutive frame number determination rules, each spatial location is classified and labeled as a stable texture region or a transient texture region, generating a texture region stability classification map.

[0023] Furthermore, the process of obtaining transient and stable texture regions specifically includes:

[0024] In a multi-frame similarity matrix, when the texture feature similarity at the corresponding spatial location is greater than a first preset threshold for more than N consecutive frames (N≥2), the corresponding spatial location is marked as a stable texture region; when the texture feature similarity at the corresponding spatial location is less than a second preset threshold for less than N frames, the corresponding spatial location is marked as a transient texture region; wherein, the first preset threshold is greater than the second preset threshold.

[0025] Furthermore, the anatomical structure feature map of each frame image is compared with the pre-stored standard normal anatomical structure feature template through multiple channels and pixel by pixel. The quantization deviation values ​​of each channel are linearly weighted and fused, and then mapped to the structure deviation map of the pixel value in the interval [0,1] through a non-linear function.

[0026] Furthermore, the specific process of adaptive weighted fusion performed by the information fusion verification module includes:

[0027] S1. Feature Tensor Construction: The intra-frame texture anomaly intensity map, inter-frame texture runaway variation map, texture region stability classification map, and structural deviation map are concatenated along the channel dimension to form a multi-source evidence feature tensor with a four-dimensional evidence vector.

[0028] S2, Adaptive Weight Generation: Input the multi-source evidence feature tensor into the adaptive weight generation network based on the spatial attention mechanism to dynamically generate an adaptive weight tensor that corresponds one-to-one with each pixel position.

[0029] S3. Weighted Fusion and Initial Confidence Calculation: Adaptive weight tensors are used to perform pixel-level weighted summation on the four types of evidence to generate a comprehensive anomaly intensity map;

[0030] S4. Stability-based lesion grading and localization: Based on preset threshold rules, the comprehensive abnormal intensity map, texture region stability classification map, and structural deviation map are jointly judged to mark and locate lesions of different grades.

[0031] Furthermore, the specific process of "stability-based lesion grading and localization" in S4 includes:

[0032] S4.1 High-confidence lesion determination: For a pixel marked as "stable texture region" in the texture region stability classification map, if its corresponding value in the comprehensive abnormal intensity map is greater than the preset first determination threshold, the pixel is determined as a high-confidence lesion.

[0033] S4.2 Determination of Lesions Requiring Review: For a pixel marked as "transient texture region" in the texture region stability classification map, if its corresponding value in the comprehensive abnormality intensity map is greater than the preset second determination threshold, and its corresponding value in the structural deviation map is greater than the preset third determination threshold, then the pixel is determined to be a lesion requiring review; wherein, the second determination threshold is less than the first determination threshold.

[0034] S4.3 Suspected Artifact Identification: For a pixel marked as "transient texture region" in the texture region stability classification map, if its corresponding value in the comprehensive anomaly intensity map is greater than the preset second judgment threshold, but its corresponding value in the structure deviation map is less than the preset third judgment threshold, then the pixel is identified as a suspected artifact.

[0035] Compared with the prior art, the advantages of this invention are:

[0036] This invention acquires multiple sets of images from the same examined eye and constructs an analysis sequence. By constraining spatiotemporal consistency, it greatly enhances the system's anti-interference capability and diagnostic robustness. For the analysis sequence, it employs parallel and decoupled "anatomical structure feature extraction branch" and "tissue texture feature extraction branch" to obtain anatomical structure feature maps and tissue texture feature maps. This achieves explicit separation and deep encoding of anatomical morphology information and tissue texture information in retinal images. For the dual feature maps, it performs conditional normal texture construction and intra-group comparison, multi-frame texture consistency verification, and anatomical structure deviation quantitative analysis. Its core lies in learning the "normal" anatomy-texture mapping relationship and standard anatomical morphology distribution of the retina, rather than memorizing specific disease morphologies. The system training only requires a large number of easily obtainable unlabeled normal retinal images, eliminating the heavy dependence of traditional supervised classification models on scarce lesion labeling data.

[0037] After the above multiple comparison analysis, an intra-frame texture anomaly intensity map, an inter-frame texture runaway variation map, a texture region stability classification map, and a structural deviation map are generated. The multi-dimensional evidence maps are adaptively weighted and fused, and a hierarchical decision is made based on texture stability and fusion confidence. Based on rule-based hierarchical localization, the comprehensive anomaly intensity of continuous values ​​and the texture stability classification and structural deviation of discrete values ​​are logically coordinated to formulate clear hierarchical judgment rules (high confidence lesions, lesions requiring review, and suspected artifacts). This process not only outputs the location of the abnormality, but also assigns different clinical meanings and action priorities to the abnormal regions. The generated structured report can be directly embedded into the doctor's workflow, significantly improving the practicality of the auxiliary diagnostic results and the action guidance value. Attached Figure Description

[0038] Figure 1 This is a system principle block diagram of the present invention;

[0039] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] Example 1: This invention discloses an intelligent retinal image abnormality recognition system based on ophthalmological auxiliary diagnosis. Please refer to [link to relevant documentation]. Figures 1-2 It includes an image acquisition and processing module, a dual-path feature decoupling and extraction module, a multiple comparison and analysis module, and an information fusion and verification module.

[0042] The image acquisition and processing module is used to control the imaging device to acquire multiple sets of retinal images of the same eye when it performs at least two different preset eye movements, and to perform standardized preprocessing on the multiple sets of retinal images to form an analysis sequence.

[0043] Specifically, the imaging device is a fundus camera or an optical coherence tomography scanner. The module issues voice or visual commands to the subject through a built-in program, guiding the eyes to perform at least two different preset actions, such as "upward gaze," "downward gaze," "nasal gaze," and "temporal gaze." During each action, the imaging device is simultaneously triggered to perform high-speed continuous shooting or scanning to acquire multiple frames of retinal images from the corresponding viewpoint. Subsequently, the module performs standardized preprocessing on all acquired images, including: unifying the image size to a preset size (e.g., 512x512 pixels), normalizing the pixel grayscale values ​​(e.g., Z-score normalization), and using image enhancement algorithms (e.g., contrast-limited adaptive histogram equalization) to highlight image details. Finally, all processed images are arranged in chronological order of acquisition time to form a multi-frame analysis sequence.

[0044] The dual-path feature decoupling extraction module is used to receive the analysis sequence, synchronously and in parallel input the pre-trained and optimized anatomical structure feature extraction branch and tissue texture feature extraction branch, and output the anatomical structure feature map and tissue texture feature map of each frame of the analysis sequence, forming the anatomical structure feature map sequence and the tissue texture feature map sequence.

[0045] The anatomical structure feature extraction branch is constructed based on a pre-trained retinal image segmentation network, preferably an anatomical structure encoder based on U-Net, to extract anatomical structure features including blood vessel morphology, optic disc geometric parameters, and fovea position.

[0046] The tissue texture feature extraction branch is trained through self-supervised learning, preferably using a texture encoder pre-trained based on the SimCLR framework, to extract texture features that are decoupled from anatomical structures and represent tissue texture, pigment, reflection, hemorrhage, and pigmentation.

[0047] By explicitly separating anatomical structures and tissue texture at the feature level, and quantifying anomalies through two independent paths—"comparison with a standard template" and "comparison with conditionalized normal texture"—this design ensures that structural distortions such as optic cup enlargement in glaucoma can be accurately quantified, while texture changes such as microaneurysms and hemorrhages are unaffected by individual anatomical differences, thus significantly improving specificity and interpretability. Furthermore, by acquiring multi-view image sequences and performing "multi-frame texture consistency verification," the system can effectively distinguish between persistent stable lesions and transient artifacts. Utilizing the intrinsic constraint of dynamic information from multi-image analysis sequences, the system's anti-interference capability and diagnostic robustness are significantly enhanced.

[0048] It should be added that this invention does not limit the specific network structure used for the extraction branches. Any deep learning model or traditional image processing algorithm that can achieve the corresponding feature extraction function falls within the protection scope of this invention.

[0049] The multiple alignment analysis module is used to receive the anatomical structure feature map sequence and the original tissue texture feature map sequence, and perform conditional normal texture construction and intra-group alignment analysis, inter-frame texture consistency verification analysis and anatomical structure deviation quantification analysis. Based on the results of the multiple alignment analysis, it generates intra-frame texture anomaly intensity map, inter-frame texture runaway variation map, texture region stability classification map and structural deviation map respectively.

[0050] The specific process of conditional normal texture construction and intra-group comparison analysis includes:

[0051] Using the anatomical structure feature map of the t-th frame in the anatomical structure feature map sequence as conditional input, a pre-trained conditional generation network is used to reconstruct the predicted normal tissue texture feature map corresponding to the current anatomical structure state. The pixel-wise difference between the tissue texture feature map of this frame and the predicted normal tissue texture feature map is calculated to obtain the original difference map.

[0052] The original difference map is passed through a 1×1 convolutional layer and a Sigmoid activation function to map it into an intra-frame texture anomaly intensity map with pixel values ​​ranging from [0,1], where the pixel value represents the probability that a texture anomaly exists at that pixel location;

[0053] The conditional generative network here is preferably a conditional generative neural network model. This model is trained on a massive dataset of normal retinal images, using the anatomical structure feature map of each image as a condition and the corresponding real normal tissue texture feature map as the learning target. It is trained in a supervised manner, so that it learns the mapping relationship from anatomical structure to normal texture. That is, it is trained on a large amount of normal structure-normal texture pairing data. Based on the input current anatomical structure, the model outputs a predicted normal tissue texture feature map. This feature map represents the normal tissue texture feature map that should appear without pathological changes under the current anatomical structure state. For example, when the input anatomical structure feature map represents a tilted optic disc, a feature map with healthy texture and natural fit to the tilted optic disc boundary will be generated, rather than the texture of a standard upright optic disc. This mechanism ensures that subsequent texture difference comparison can isolate the influence of pure morphological variation and focus on detecting real tissue pathological changes.

[0054] This module performs intra-group comparisons, comparing the current texture with "the normal texture it should have" and marking the probability of texture abnormality at the pixel level;

[0055] By generating comparisons under conditions, on the one hand, the comparisons are ensured to be performed within the same anatomical framework, eliminating interference from individual differences; on the other hand, the dissection analysis of structural and textural abnormality signals is realized, providing orthogonal and high-quality feature inputs for subsequent multi-evidence fusion.

[0056] The specific process of multi-frame texture consistency verification analysis includes:

[0057] Cross-frame collaborative alignment is performed on the tissue texture feature map sequence. The texture feature similarity between any two frames in the sequence at corresponding spatial locations is calculated. The commonly used calculation method is cosine similarity. The above calculation is performed on all locations in the entire image to obtain an inter-frame similarity matrix of the same size as the input image. This calculation is repeated for all pairs of frames in the sequence to obtain multiple inter-frame similarity matrices.

[0058] Take the median, mean, or other aggregated statistics from multiple inter-frame similarity matrices to generate an inter-frame texture runaway variation map that characterizes the overall variation intensity of texture at each location.

[0059] Based on multiple inter-frame similarity matrices, according to preset thresholds and consecutive frame number determination rules, each spatial location is classified and labeled as a stable texture region or a transient texture region, generating a texture region stability classification map.

[0060] This module first performs inter-group comparison of texture feature maps, analyzes the variation of texture under different actions, generates a texture spatiotemporal variation map, marks the degree of texture instability at the pixel level, then performs inter-group comparison logic judgment, applies rule classification based on the texture spatiotemporal variation map, generates a texture stability classification map, and marks whether the pixel belongs to the category label of "continuous anomaly" or "transient anomaly".

[0061] The process of obtaining transient and stable texture regions specifically includes:

[0062] In a multi-frame similarity matrix, when the texture feature similarity at the corresponding spatial location is greater than the first preset threshold for more than N consecutive frames (N≥2), the corresponding spatial location is marked as a stable texture region. Here, the first preset threshold is used to determine "high similarity". When the texture feature similarity between two frames is greater than the first preset threshold, it means that the two frames are highly similar, that is, the texture at this location hardly changes between different frames.

[0063] When the texture feature similarity of less than N frames at the corresponding spatial location is less than the second preset threshold, the corresponding spatial location is marked as a transient texture region. Here, the second preset threshold is used to determine "low similarity". When the texture feature similarity of two frames is less than the second preset threshold, it means that the two frames are significantly different, that is, the texture at this location has changed drastically between different frames.

[0064] Among them, the first preset threshold is greater than the second preset threshold. That is, in the judgment rule, there is a "high similarity" interval and a "low similarity" interval. The value in between may be regarded as "neutral" or needs to be further judged in combination with other evidence. The value of the preset threshold is based on a combination of large model analysis of sample data and human experience.

[0065] By generating continuous inter-frame texture runaway mutation maps and discrete texture region stability classification maps respectively, we achieved decoupled analysis and characterization of texture anomaly 'mutation intensity' and 'change pattern', providing richer and more accurate spatiotemporal evidence for subsequent fusion decisions;

[0066] The specific process of quantifying anatomical deviations includes:

[0067] The anatomical structure feature map of each frame in the anatomical structure feature map sequence is compared with the pre-stored standard normal anatomical structure feature template through multi-channel pixel-by-pixel comparison. The quantization deviation values ​​of each channel are linearly weighted and fused, and then mapped to the structure deviation map with pixel values ​​in the range of [0,1] through a non-linear function. The pixel value represents the degree of deviation of the anatomical structure from the normal, so as to quantify the degree of deviation of its shape, position or topological relationship.

[0068] The multi-channel comparison includes: for the feature channel representing the vascular skeleton, calculating its Hausdorff distance with the standard template; for the feature channel representing the optic disc cup concavity, calculating the complement (1-DSC) of its Dessian similarity coefficient with the standard template.

[0069] The core of the "conditional normal texture construction" and "anatomical deviation quantification" analysis paths proposed in this module lies in learning the "normal" anatomy-texture mapping relationship and standard anatomical morphology distribution of the retina, rather than memorizing specific disease morphologies. Therefore, system training only requires a large number of easily accessible unlabeled normal retinal images, freeing it from the heavy dependence of traditional supervised classification models on expensive and scarce lesion labeling data. At the same time, since the model establishes a cognition of a broad range of "normal" patterns, rare lesions, new lesions, or complex comorbidities that have not appeared in the training data can be effectively perceived and indicated as long as they deviate from the learned normal patterns. This fundamentally solves the problems of poor generalization and detection blind spots in supervised methods.

[0070] The information fusion and verification module is used to receive multi-class analysis result data sent from the multi-comparison analysis module, perform adaptive weighted fusion analysis, obtain a comprehensive abnormal intensity map, and perform graded judgment and localization of lesions based on the comprehensive abnormal intensity map, texture region stability classification map and structural deviation map in collaborative analysis.

[0071] The specific process of adaptive weighted fusion analysis performed by the information fusion verification module includes:

[0072] S1. Feature Tensor Construction: The intra-frame texture anomaly intensity map, inter-frame texture runaway variation map, texture region stability classification map, and structural deviation map are concatenated along the channel dimension to form a multi-source evidence feature tensor with a four-dimensional evidence vector. It should be noted that "concatenation along the channel dimension" is a standard operation in deep learning data processing. In the deep learning framework, the image data of each image is usually represented as a three-dimensional tensor with shape (H, W, C), where H is the height, W is the width, and C is the number of channels. "Concatenation along the channel dimension" means merging the channel dimension C of multiple input tensors while keeping H and W unchanged. The purpose is to combine multiple two-dimensional data with the same spatial size into a new data block along the "channel" dimension. This concatenation method preserves the spatial alignment of all original information and forms a unified data structure, integrating four heterogeneous information (texture anomaly intensity, spatiotemporal variation pattern, stability classification, and structural deviation) into a normalized tensor, which can be directly used as the standard input for subsequent spatial attention networks.

[0073] S2. Adaptive weight generation: The multi-source evidence feature tensor is input into the adaptive weight generation network based on the spatial attention mechanism. The network takes the multi-source evidence feature tensor, which is formed by stitching together the four types of feature maps generated by the aforementioned multiple comparison analysis, as input. By analyzing the inherent correlation and spatial context information between the texture anomaly intensity, spatiotemporal variation pattern, stability classification and structural deviation contained therein, the network dynamically generates an adaptive weight tensor that corresponds one-to-one with each pixel position.

[0074] S3. Weighted Fusion and Initial Confidence Calculation: An adaptive weight tensor is used to perform pixel-level weighted summation on four types of evidence (texture, spatiotemporal, stability, and structure) to generate a comprehensive anomaly intensity map. This is a quantitative synthesis step, which combines the intelligently evaluated weights with the original evidence, compressing and fusing the information from the four dimensions into a single, continuous confidence value. This value comprehensively reflects all the evidence and their relative importance, and is the system's preliminary quantitative score for "whether there is an anomaly at this point," providing a direct and comparable numerical basis for the final decision.

[0075] S4. Stability-based lesion grading and localization: Based on preset threshold rules, the comprehensive abnormal intensity map, texture region stability classification map, and structural deviation map are jointly judged to mark and locate lesions of different grades.

[0076] The specific process of "stability-based lesion grading and localization" in S4 includes:

[0077] S4.1 High-Confidence Lesion Determination: For pixels marked as "Stable Texture Region" in the texture region stability classification map, if their corresponding value in the comprehensive anomaly intensity map is greater than the preset first determination threshold, then the pixel is determined as a high-confidence lesion, directly classified as a high-confidence lesion, when combined with structural deviation. Figure 1 During the analysis, a low structure value indicates that this abnormality may primarily be a pure texture lesion;

[0078] S4.2 Determination of Lesions Requiring Review: For pixels marked as "transient texture regions" in the texture region stability classification map, if their corresponding value in the comprehensive abnormality intensity map is greater than the preset second judgment threshold, and their corresponding value in the structural deviation map is greater than the preset third judgment threshold, then the pixel is determined to be a lesion requiring review; wherein, the second judgment threshold is less than the first judgment threshold, indicating that although the point is unstable, it is supported by significant structural abnormalities, so it is considered to be highly likely to be a real lesion and requires the doctor to review it carefully;

[0079] S4.3 Suspected artifact identification: For a pixel marked as "transient texture region" in the texture region stability classification map, if its corresponding value in the comprehensive anomaly intensity map is greater than the preset second judgment threshold, but its corresponding value in the structural deviation map is less than the preset third judgment threshold, then the pixel is identified as a suspected artifact, indicating that the point abnormally lacks anatomical structure support and is unstable itself, so it is judged as a suspected artifact (such as reflection).

[0080] The output of high-confidence lesion regions / lesions requiring review, along with their rich multi-dimensional features (texture, structure, spatiotemporal features), provides excellent and clean input for subsequent connection to a dedicated disease classification module.

[0081] It should be added that the article involves comparisons of various thresholds. Thresholds, preset values, preset ranges, etc., are set for result comparison and analysis to determine good or bad. The magnitude of these thresholds is determined by a combination of large-scale model analysis of sample data and human experience. They can also be appropriately adjusted based on seasonal or common-sense influences.

[0082] In summary, the analysis involves controlling the imaging device to acquire multimodal image sequences of the same eye under different gaze directions, providing a dynamic information basis for analysis. Subsequently, the dual-path feature decoupling extraction module separates the anatomical structural features and tissue texture features of the images. The multiple comparison anomaly analysis module performs three core analyses in parallel: generating "predicted normal texture" based on its own structure and performing intra-group comparison to detect texture anomalies; performing cross-frame consistency verification on multi-frame texture features to distinguish between stable lesions and transient artifacts; and comparing the current anatomical structure with a standard template to quantify structural deviations. Finally, the information fusion verification module adaptively weights and fuses the multi-dimensional evidence maps generated by the above analyses, and makes lesion grading decisions based on texture stability and fusion confidence.

[0083] It employs a collaborative technical approach that combines multimodal imaging, feature decoupling, conditional normal comparison, spatiotemporal consistency verification, and adaptive decision fusion. This approach enables highly sensitive, specific, and robust detection of structural and textural abnormalities in retinal images without relying on large-scale lesion annotation data. It also provides interpretable, gradeable, and localizable clinical auxiliary diagnostic information, effectively addressing the core issues of existing technologies, such as strong dependence on annotation data, insensitivity to unknown abnormalities, susceptibility to artifact interference, and insufficient specificity.

[0084] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto; any equivalent substitutions or modifications made by those skilled in the art within the technical scope disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent retinal image abnormality recognition system based on ophthalmic auxiliary diagnosis, characterized in that: include: The image acquisition and processing module is used to control the imaging device to acquire multiple sets of retinal images of the same eye when it performs at least two different preset eye movements, and to perform standardized preprocessing on the multiple sets of retinal images to form an analysis sequence. The dual-path feature decoupling extraction module is used to receive the analysis sequence, synchronously and in parallel input the pre-trained anatomical structure feature extraction branch and tissue texture feature extraction branch, and output the anatomical structure feature map and tissue texture feature map of each frame of the analysis sequence, forming the anatomical structure feature map sequence and the tissue texture feature map sequence. The multiple alignment analysis module is used to receive the anatomical structure feature map sequence and the original tissue texture feature map sequence, and perform conditional normal texture construction and intra-group alignment analysis, multi-frame texture consistency verification analysis and anatomical structure deviation quantification analysis. Based on the results of the multiple alignment analysis, it generates intra-frame texture anomaly intensity map, inter-frame texture runaway variation map, texture region stability classification map and structural deviation map. The information fusion and verification module is used to receive multi-class image data, perform adaptive weighted fusion to obtain a comprehensive abnormal intensity map, and perform collaborative analysis based on the comprehensive abnormal intensity map, texture region stability classification map, and structural deviation map to classify and locate lesions.

2. The intelligent retinal image abnormality recognition system based on ophthalmic auxiliary diagnosis according to claim 1, characterized in that: The anatomical structure feature extraction branch is built based on a pre-trained retinal image segmentation network and is used to extract anatomical structure features including blood vessel morphology, optic disc geometric parameters, and foveal position. The tissue texture feature extraction branch is trained through self-supervised learning and is used to extract texture features that are decoupled from anatomical structures and represent tissue texture, pigment, reflection, hemorrhage, and pigmentation.

3. The intelligent retinal image abnormality recognition system based on ophthalmic auxiliary diagnosis according to claim 2, characterized in that: The specific process of conditional normal texture construction and intra-group comparison analysis includes: Using the anatomical structure feature map of the t-th frame in the anatomical structure feature map sequence as conditional input, a pre-trained conditional generation network is used to reconstruct the predicted normal tissue texture feature map corresponding to the current anatomical structure state. The pixel-wise difference between the tissue texture feature map of this frame and the predicted normal tissue texture feature map is calculated to obtain the original difference map. The original difference map is passed through a 1×1 convolutional layer and a sigmoid activation function to map it into an intra-frame texture anomaly intensity map with pixel values ​​in the range [0,1].

4. The intelligent retinal image abnormality recognition system based on ophthalmic auxiliary diagnosis according to claim 3, characterized in that: The specific process of multi-frame texture consistency verification analysis includes: Cross-frame collaborative alignment of the tissue texture feature map sequence is performed, and the texture feature similarity between any two frames in the sequence at corresponding spatial locations is calculated to obtain multiple inter-frame similarity matrices. Take the median, mean, or other aggregated statistics from multiple inter-frame similarity matrices to generate an inter-frame texture runaway variation map that characterizes the overall variation intensity of texture at each location. Based on multiple inter-frame similarity matrices, and according to preset thresholds and consecutive frame number determination rules, each spatial location is classified and labeled as a stable texture region and a transient texture region, generating a texture region stability classification map.

5. The intelligent retinal image abnormality recognition system based on ophthalmic auxiliary diagnosis according to claim 4, characterized in that: The process of obtaining transient and stable texture regions specifically includes: In a multi-frame similarity matrix, when the texture feature similarity at the corresponding spatial location is greater than a first preset threshold for more than N consecutive frames (N≥2), the corresponding spatial location is marked as a stable texture region; when the texture feature similarity at the corresponding spatial location is less than a second preset threshold for less than N frames, the corresponding spatial location is marked as a transient texture region; wherein, the first preset threshold is greater than the second preset threshold.

6. The intelligent retinal image abnormality recognition system based on ophthalmic auxiliary diagnosis according to claim 5, characterized in that: The specific process of quantifying anatomical deviations includes: The anatomical structure feature map of each frame image is compared with the pre-stored standard normal anatomical structure feature template through multiple channels and pixel by pixel. The quantization deviation values ​​of each channel are linearly weighted and fused, and then mapped to the structure deviation map of pixel values ​​in the interval [0,1] through a non-linear function.

7. The intelligent retinal image abnormality recognition system based on ophthalmic auxiliary diagnosis according to claim 6, characterized in that: The specific process of adaptive weighted fusion performed by the information fusion verification module includes: S1. Feature Tensor Construction: The intra-frame texture anomaly intensity map, inter-frame texture runaway variation map, texture region stability classification map, and structural deviation map are concatenated along the channel dimension to form a multi-source evidence feature tensor with a four-dimensional evidence vector. S2, Adaptive Weight Generation: Input the multi-source evidence feature tensor into the adaptive weight generation network based on the spatial attention mechanism to dynamically generate an adaptive weight tensor that corresponds one-to-one with each pixel position. S3. Weighted Fusion and Initial Confidence Calculation: Adaptive weight tensors are used to perform pixel-level weighted summation on the four types of evidence to generate a comprehensive anomaly intensity map; S4. Stability-based lesion grading and localization: Based on preset threshold rules, the comprehensive abnormal intensity map, texture region stability classification map, and structural deviation map are jointly judged to mark and locate lesions of different grades.

8. The intelligent retinal image abnormality recognition system based on ophthalmic auxiliary diagnosis according to claim 7, characterized in that: The specific process of "stability-based lesion grading and localization" in S4 includes: S4.1 High-confidence lesion determination: For a pixel marked as "stable texture region" in the texture region stability classification map, if its corresponding value in the comprehensive abnormal intensity map is greater than the preset first determination threshold, then the pixel is determined as a high-confidence lesion. S4.2 Determination of Lesions Requiring Review: For a pixel marked as "transient texture region" in the texture region stability classification map, if its corresponding value in the comprehensive abnormality intensity map is greater than the preset second determination threshold, and its corresponding value in the structural deviation map is greater than the preset third determination threshold, then the pixel is determined to be a lesion requiring review; wherein, the second determination threshold is less than the first determination threshold. S4.3 Suspected Artifact Identification: For a pixel marked as "transient texture region" in the texture region stability classification map, if its corresponding value in the comprehensive anomaly intensity map is greater than the preset second judgment threshold, but its corresponding value in the structure deviation map is less than the preset third judgment threshold, then the pixel is identified as a suspected artifact.

Images