Strabismus detection system based on artificial intelligence

Abstract

The invention relates to the technical field of artificial intelligence, provides an artificial intelligence-based strabismus detection system, and aims to improve the accuracy and intelligent level of strabismus recognition. The system comprises an eye image acquisition module, a feature extraction module, a strabismus identification module, a neural feedback control module and a report generation module. The method comprises the following steps: constructing a coupling state driving prediction network, fusing CNN feature extraction, a Hopfield network, a Kuramoto oscillator, an attention mechanism and a multi-layer perceptron, and carrying out deep modeling on eye derived structure features under multiple visual angles; further, a reverse under-approximation analysis algorithm is introduced into the system, and strategy feasibility evaluation and optimization selection based on the state space are achieved; according to the method, the classification precision and model robustness of strabismus recognition are effectively improved, the adaptive capacity of the system to individual differences is enhanced, and the method has the remarkable advantages of intelligence, data driving and personalized regulation and control.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, in particular to an artificial intelligence-based strabismus detection system. Background Art

[0002] Strabismus is a common eye disease characterized by the inability of the visual axes of both eyes to focus on the same target at the same time. Common types include esotropia, exotropia, hypertropia, and hypotropia. Strabismus not only affects visual function, but may also cause diplopia, visual fatigue, and even amblyopia, seriously affecting the patient's quality of life. Some existing strabismus detection systems have introduced artificial intelligence technology, using convolutional neural networks (CNN) or posture estimation models to classify and judge eye images, and have a certain degree of automatic recognition capability. Such systems mainly use static image features such as eye positioning, pupil deviation detection, and palpebral fissure symmetry analysis to complete the preliminary judgment of strabismus type. Initial results have been achieved in improving detection efficiency and reducing human errors; however, these existing strabismus detection systems still have the following shortcomings: on the one hand, existing systems mainly rely on single-frame image analysis, lack the ability to dynamically model multiple perspectives (such as up, down, left, and right) or continuous gaze behaviors, and have difficulty capturing stability patterns and relative coupling behaviors during eye movement, which limits the accuracy of recognition and the ability to refine strabismus types; on the other hand, existing models mostly remain at the classification judgment level, lack deep integration with the user's individual eye movement mechanism or physiological parameters, and are unable to further output executable regulation suggestions or intervention paths. Summary of the Invention

[0003] The present invention provides an artificial intelligence-based strabismus detection system designed to improve the accuracy and robustness of strabismus recognition and automatically generate personalized intervention strategies. First, the system innovatively constructs a coupled-state driven prediction network. By integrating CNN feature extraction, a Hopfield network, a Kuramoto oscillator, an attention mechanism, and a multi-layer perceptron (MLP), it models and fuses structural features derived from eye images from multiple viewpoints. This enhances the ability to model the coupling characteristics between eye structures, significantly improving recognition accuracy and classification detail compared to existing technologies. Second, the system significantly enhances the intelligence of the strabismus detection system through the collaborative design of a coupled-driven recognition mechanism and a reverse analysis intervention optimization method. This mechanism not only strengthens the closed-loop feedback capability of the model but also enables personalized, data-driven intelligent intervention capabilities. The synergistic effect of these two key innovations enables the present invention to not only improve the refinement and intelligence of strabismus recognition but also enhance the system's ability to generate personalized control solutions. The system demonstrates strong technical adaptability and application potential in intelligent application fields such as visual behavior analysis and functional control suggestion generation.

[0004] The present invention provides an artificial intelligence-based strabismus detection system, which includes an eye image acquisition module, a feature extraction module, a strabismus recognition module, a neural feedback control module, and a report generation module;

[0005] The eye image acquisition module collects multi-directional gaze images of the human eye, including frontal, left, right, and up and down gaze, and performs size normalization and color space standardization on the gaze images to generate eye image data;

[0006] A feature extraction module extracts derived structural features of eye image data, including eye deviation angle features, left and right eye symmetry features, pupil position change trend features, and gaze trajectory stability features;

[0007] The strabismus recognition module combines CNN feature extraction, Hopfield networks, Kuramoto oscillators, and an attention mechanism with an MLP fusion strategy to construct a coupled state-driven prediction network. This coupled state-driven prediction network processes derived structural features and outputs strabismus recognition results, including normal, esotropia, exotropia, hypertropia, and hypotropia categories, as well as confidence scores and derived structural feature analysis results for each category.

[0008] The neural feedback control module constructs a reverse under-approximation analysis algorithm, combines the derived structural features, and uses the reverse under-approximation analysis algorithm to generate the optimal intervention strategy;

[0009] The report generation module integrates the strabismus identification results and the optimal intervention strategy to formulate personalized eye function regulation recommendations and generate a test report; personalized eye function regulation recommendations include intervention methods, intensity, duration and expected improvement path.

[0010] Furthermore, the process of generating the squint recognition result by the squint recognition module specifically includes the following steps:

[0011] Step S1: Establish a ResNet-18 model to extract low-level features, mid-level features, and high-level features of derived structural features;

[0012] Step S2: Process low-level features to generate coupled enhanced local feature vectors;

[0013] Step S3: Perform multi-scale feature extraction on the intermediate features through multi-scale convolution kernels to generate multi-scale feature vectors;

[0014] Step S4: Use the channel attention mechanism to process high-level features and generate a global feature vector;

[0015] Step S5: Enhance the attention weights of local feature vectors, multi-scale feature vectors, and global feature vectors through MLP learning coupling, perform weighted fusion, and generate a comprehensive feature vector;

[0016] Step S6: Map the comprehensive feature vector to the category space dimension through a linear layer and output the strabismus recognition result.

[0017] Furthermore, step S2 specifically includes: processing low-level features through the spatial attention mechanism to generate local feature vectors, using the Hopfield network and Kuramoto oscillator to construct a Hopfield-Kuramoto coupling model, enhancing the local feature vectors through the Hopfield-Kuramoto coupling model, simulating the coupling behavior between local neurons of the local feature vectors, improving the expressiveness and robustness of edge areas and local shapes, and generating coupled enhanced local feature vectors.

[0018] Furthermore, the process of generating the optimal intervention strategy by the neural feedback control module specifically includes the following steps:

[0019] Step B1: Collect individual physiological parameter sets, including eye muscle tension parameters and nerve conduction velocity; construct a nonlinear eye movement state space model based on derived structural features; define the target state set, safe state set, and alternative intervention strategy set;

[0020] Step B2: For each intervention strategy in the candidate intervention strategy set, introduce under-approximation parameters, perform reverse dynamic propagation on the nonlinear eye movement state space model through the distributed interval decomposition method, and construct an under-approximation backward reachable set;

[0021] Step B3: Combine the underapproximate backward reachable set and the safe state set, perform an intersection operation, and construct the feasible region of the strategy; calculate the feasibility score of the feasible region of the strategy in the safe state set, and use the stability index function to evaluate and obtain the average stability score;

[0022] Step B4: Construct a strategy performance indicator set based on the feasibility score and average stability score. Introduce user preference parameters to assign weights to the indicators of the strategy performance indicator set, integrate them to form a comprehensive performance indicator, and use a weighted multi-objective optimization method to maximize the comprehensive performance indicator to screen the optimal intervention strategy. User preference parameters include intervention acceptability and comfort priority.

[0023] By adopting the above scheme, the beneficial effects achieved by the present invention are as follows:

[0024] By constructing a coupled-state driven prediction network, the present invention achieves multi-level deep modeling and fusion of derived structural features of eye images, improving the feature expressiveness and classification accuracy during strabismus recognition. This network integrates the image encoding capabilities of CNN, the stable state discrimination capabilities of the Hopfield network, and the modeling capabilities of the Kuramoto oscillator for local phase synchronization relationships. It also introduces an attention mechanism and an MLP fusion strategy to weightedly integrate local and global features, thereby achieving more discriminative judgment of different strabismus types (esotropia, exotropia, hypertropia, and hypotropia). This structure effectively solves the problem of inaccurate classification in existing strabismus recognition models under complex sight angles and multi-viewing conditions, enhancing the system's robustness and adaptability in actual use.

[0025] In terms of generating intervention strategies, the present invention introduces a reverse underapproximation analysis algorithm, which realizes dynamic modeling and strategy evaluation of individual eye movement state space. By constructing an underapproximation backward reachable set, the intervention strategy is predicted and screened for effect, thereby improving the accuracy and scientific nature of the intervention suggestions. This process combines multiple source parameters such as the user's eye muscle tension, physiological characteristics, and derived image features, and designs a dynamic feasibility evaluation mechanism between the target state set and the safe state set. This makes the generated intervention suggestions not only theoretically feasible but also highly individual adaptable. This method solves the problems of the traditional model's lack of intervention effect prediction and single control scheme, enhances the system's feedback control capability, and provides decision support for personalized visual function assistance.

[0026] Overall, the present invention deeply integrates intelligent recognition and feedback intervention modules, realizing an intelligent process from eye image acquisition and strabismus recognition to the generation of control suggestions. The introduction and integration of various algorithms not only enriches the system's modeling methods for structural features, but also makes the generation of control suggestions more adaptable and scientific, improving the system's overall performance and intelligence level in strabismus detection tasks. The above improvements fully leverage the capabilities of artificial intelligence in visual structure modeling and complex system optimization, ensuring that the present invention has good application value and broad promotion potential in practical scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a module diagram of an artificial intelligence-based strabismus detection system proposed in the present invention;

[0028] Figure 2 This is the dynamic evolution diagram of neuron phase synchronization proposed in Example 3. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0030] Example 1, according to Figure 1 , the present invention provides a strabismus detection system based on artificial intelligence, which includes an eye image acquisition module, a feature extraction module, a strabismus recognition module, a neural feedback control module and a report generation module;

[0031] The eye image acquisition module collects multi-directional gaze images of the human eye, including frontal, left, right, and up and down gaze, and performs size normalization and color space standardization on the gaze images to generate eye image data;

[0032] A feature extraction module extracts derived structural features of eye image data, including eye deviation angle features, left and right eye symmetry features, pupil position change trend features, and gaze trajectory stability features;

[0033] The strabismus recognition module combines CNN feature extraction, Hopfield networks, Kuramoto oscillators, and an attention mechanism with an MLP fusion strategy to construct a coupled state-driven prediction network. This coupled state-driven prediction network processes derived structural features and outputs strabismus recognition results, including normal, esotropia, exotropia, hypertropia, and hypotropia categories, as well as confidence scores and derived structural feature analysis results for each category.

[0034] The neural feedback control module constructs a reverse under-approximation analysis algorithm, combines the derived structural features, and uses the reverse under-approximation analysis algorithm to generate the optimal intervention strategy;

[0035] The report generation module integrates the strabismus identification results and the optimal intervention strategy to formulate personalized eye function regulation recommendations and generate a test report; personalized eye function regulation recommendations include intervention methods, intensity, duration and expected improvement path.

[0036] Example 2: This example is based on Example 1. In this example, the process of generating a squint recognition result by the squint recognition module specifically includes the following steps:

[0037] Step S1: Establish a ResNet-18 model to extract low-level features, mid-level features, and high-level features of derived structural features;

[0038] Low-level features include:

[0039] Texture: The delicate texture of the eyelids, eyelashes, and pupil edges;

[0040] Edges: the outline of the palpebral fissure, the edge of the cornea, and the dividing line between the white of the eye and the pupil;

[0041] Local contrast: reflects the changes in brightness and darkness in the image, including the distribution of highlights and shadows caused by lighting;

[0042] These features help detect pupil position, edge drift, and small geometric differences;

[0043] Intermediate features include:

[0044] Local structural contours: spatial structural relationships of the canthus, palpebral fissure, and pupil;

[0045] Eye deviation structure: including whether the left and right eyes are looking in the same direction and whether one eye is obviously offset from the center position;

[0046] Symmetry pattern: This is an important basis for judging whether the left and right eyes are structurally symmetrical, and whether it is esotropia or exotropia;

[0047] These features help the model understand the relative positions and dynamic trends between eyes;

[0048] Advanced features include:

[0049] Global semantic structure: overall eye posture characteristics, including the spatial configuration of frontal / upward / downward gaze states;

[0050] Gaze offset mode;

[0051] Stability patterns of gaze behavior: changing trends during long-term gaze;

[0052] These features facilitate high-level discrimination of strabismus types;

[0053] Step S2: Process low-level features to generate coupled enhanced local feature vectors;

[0054] Step S3: Perform multi-scale feature extraction on the intermediate features through multi-scale convolution kernels to generate multi-scale feature vectors;

[0055] Step S4: Use the channel attention mechanism to process high-level features and generate a global feature vector;

[0056] Step S5: Enhance the attention weights of local feature vectors, multi-scale feature vectors, and global feature vectors through MLP learning coupling, perform weighted fusion, and generate a comprehensive feature vector;

[0057] Step S6: Map the comprehensive feature vector to the category space dimension through a linear layer and output the strabismus recognition result.

[0058] Example 3, according to Figure 2 This embodiment is based on the second embodiment. In this embodiment, step S2 specifically includes: processing low-level features through a spatial attention mechanism to generate local feature vectors, using a Hopfield network and a Kuramoto oscillator to construct a Hopfield-Kuramoto coupling model, enhancing the local feature vectors through the Hopfield-Kuramoto coupling model, simulating the coupling behavior between local neurons of the local feature vectors, performing neuronal phase synchronization, improving the expressiveness and robustness of edge regions and local shapes, and generating coupled-enhanced local feature vectors; the evolution formula used in the Hopfield-Kuramoto coupling model is as follows:

[0059] The evolution formula of the characteristic state term:

[0060] ;

[0061] in, Indicates the current center position index, Represents the neighbor node position index, Indicates the The local feature vector at each position, Indicates the The local feature vector at each position, Represents the local feature vector The rate of change over time; Indicates that the activation function acts on the local feature vector , increase nonlinear expression capabilities, Indicates that the activation function acts on the local feature vector ; Represents the connection weight in the Hopfield network; represents the noise term, represents the adjustment coefficient of the Hopfield-Kuramoto coupling term; represents the derivative of the coupling strength function; and Indicates location and The phase state vector of represents the transpose symbol; represents the coupling modulation function;

[0062] The evolution formula of the oscillatory state term is:

[0063] ;

[0064] in, express The rate of change over time, Indicates the The intrinsic rotation frequency matrix of the positions, represents the identity matrix, represents the outer product matrix, represents the projection matrix; represents the coupling strength function; represents the derivative of the coupling modulation function.

[0065] Embodiment 4: This embodiment is based on embodiment 2. In this embodiment, step S2 specifically includes: processing low-level features through a spatial attention mechanism to generate a local feature vector.

[0066] Example 5: This example is based on Example 3. In this example, the process of generating the optimal intervention strategy by the neural feedback control module specifically includes the following steps:

[0067] Step B1: Collect individual physiological parameter sets, including eye muscle tension parameters and nerve conduction velocity; construct a nonlinear eye movement state space model based on derived structural features; define the target state set, safe state set, and alternative intervention strategy set;

[0068] Step B2: For each intervention strategy in the set of candidate intervention strategies, introduce an underapproximated parameter and perform reverse dynamic propagation on the nonlinear eye movement state space model through the distributed interval decomposition method to construct an underapproximated backward reachable set. The formula used is as follows:

[0069] Construct the under-approximated backward reachable set formula:

[0070] ;

[0071] in, represents the expansion tolerance of the target state set, represents the intervention strategy number index, Indicates the intervention strategies, Indicates intervention strategy The under-approximated backward reachable set constructed below is Indicates backward reachability; represents the state variable, Indicates time In the time interval middle, Indicates the upper limit of the back propagation time; Representation parameter sets, including eye muscle tone parameters, nerve conduction velocity, and derived structural features; Indicates status In intervention strategies With parameter set Push down and back Results per unit time; Represents the target state set of expansion, Indicates expansion operation;

[0072] Step B3: Combine the underapproximate backward reachable set and the safe state set, perform an intersection operation, and construct the feasible region of the strategy; calculate the feasibility score of the feasible region of the strategy in the safe state set, and use the stability index function to evaluate and obtain the average stability score;

[0073] Step B4: Construct a strategy performance indicator set based on the feasibility score and average stability score. Introduce user preference parameters to assign weights to the indicators of the strategy performance indicator set, integrate them to form a comprehensive performance indicator, and use a weighted multi-objective optimization method to maximize the comprehensive performance indicator to screen the optimal intervention strategy. User preference parameters include intervention acceptability and comfort priority.

[0074] Example 6. This example is based on Example 5. In this example, the report generation module integrates the strabismus recognition results and the optimal intervention strategy to formulate personalized eye function regulation suggestions and generate a test report; the personalized eye function regulation suggestions include the intervention method, intensity, duration, and expected improvement path;

[0075] Strabismus recognition results:

[0076] Test subject information: Male, 24 years old, no history of eye surgery, recently experienced frequent visual fatigue;

[0077] Image acquisition direction: five directions: front view, left view, right view, top view, and bottom view;

[0078] Derived structural feature analysis results:

[0079] Left and right eye gaze trajectory deviation index: 0.62 (high);

[0080] Eye position symmetry score: 0.38 (asymmetric);

[0081] The mean pupil center drift angle was 5.3°.

[0082] Gaze stability score: 0.74 (low normal);

[0083] Confidence score:

[0084] Normal: 0.07;

[0085] Esotropia: 0.82;

[0086] Exotropia: 0.06;

[0087] Hypertropia: 0.03;

[0088] Hypostropia: 0.02;

[0089] Recognition result: esotropia (confidence 0.82);

[0090] Optimal intervention strategy:

[0091] Candidate intervention strategy set:

[0092] A: Strong visual stimulation + passive training;

[0093] B: Mild muscle training + visual feedback tracking;

[0094] C: Dynamic gaze point transfer method + stroboscopic adjustment;

[0095] D: combined intervention (weak stimulation + feedback + active exercise);

[0096] Results of reverse under-approximated reachable set analysis:

[0097] Strategy B had the highest security and stability scores (average score 0.87) and a high individual comfort acceptance rate;

[0098] Although strategy D has a rapid improvement, the individual feedback risk is relatively high (safety score 0.66);

[0099] Optimal intervention strategy selection: Strategy B: mild muscle training + visual feedback tracking.

[0100] The present invention and its embodiments are described above. Such description is not restrictive. What is shown in the accompanying drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in this field are inspired by it and do not depart from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based strabismus detection system, comprising an eye image acquisition module, wherein the eye image acquisition module acquires eye image data; characterized in that: The system also includes a feature extraction module, a strabismus recognition module, a neural feedback control module and a report generation module; The feature extraction module extracts derived structural features of the eye image data; The strabismus recognition module constructs a coupled state driven prediction network, processes the derived structural features through the coupled state driven prediction network, and outputs a strabismus recognition result; The neural feedback control module constructs a reverse under-approximation analysis algorithm, combines the derived structural features, and uses the reverse under-approximation analysis algorithm to generate the optimal intervention strategy; The report generation module integrates the strabismus identification results and the optimal intervention strategy, formulates personalized eye function regulation suggestions, and generates a test report.

2. The artificial intelligence-based strabismus detection system according to claim 1, characterized in that: The method for constructing a coupled state-driven prediction network is to combine CNN feature extraction, Hopfield network, Kuramoto oscillator, and the fusion strategy of attention mechanism and MLP.

3. The artificial intelligence-based strabismus detection system according to claim 2, characterized in that: The process of generating the strabismus recognition result by the strabismus recognition module specifically includes the following steps: Step S1: Establish a ResNet-18 model to extract low-level features, mid-level features, and high-level features of derived structural features; Step S2: Process low-level features to generate coupled enhanced local feature vectors; Step S3: Perform multi-scale feature extraction on the intermediate features through multi-scale convolution kernels to generate multi-scale feature vectors; Step S4: Use the channel attention mechanism to process high-level features and generate a global feature vector; Step S5: performing weighted fusion on the coupled enhanced local feature vector, the multi-scale feature vector and the global feature vector to generate a comprehensive feature vector; Step S6: Process the comprehensive feature vector to generate a strabismus recognition result.

4. The artificial intelligence-based strabismus detection system according to claim 3, characterized in that: Step S2 specifically includes: processing low-level features through the spatial attention mechanism to generate local feature vectors, using the Hopfield network and Kuramoto oscillator to build a Hopfield-Kuramoto coupling model, enhancing the local feature vectors through the Hopfield-Kuramoto coupling model, simulating the coupling behavior between local neurons of the local feature vectors, and generating coupled-enhanced local feature vectors.

5. The artificial intelligence-based strabismus detection system according to claim 1, characterized in that: The process of generating the optimal intervention strategy by the neural feedback control module specifically includes the following steps: Step B1: Collect individual physiological parameter sets and construct a nonlinear eye movement state space model based on derived structural features; define a safe state set and a set of alternative intervention strategies; Step B2: For each intervention strategy in the candidate intervention strategy set, introduce under-approximation parameters, perform reverse dynamic propagation on the nonlinear eye movement state space model through the distributed interval decomposition method, and construct an under-approximation backward reachable set; Step B3: Combine the underapproximate backward reachable set to construct the strategy feasible region; calculate the feasibility score of the strategy feasible region in the safe state set, and use the stability index function to evaluate and obtain the average stability score; Step B4: Construct a strategy performance indicator set based on the feasibility score and the average stability score, introduce user preference parameters to assign weights to the indicators of the strategy performance indicator set, and screen the optimal intervention strategy.

6. The artificial intelligence-based strabismus detection system according to claim 5, characterized in that: User preference parameters include intervention acceptability and comfort priority.

Images