High-precision visual function intelligent diagnosis method and system and electronic equipment

Abstract

The invention relates to the technical field of visual function diagnosis, and discloses a high-precision visual function intelligent diagnosis method and system and electronic device.The high-precision visual function intelligent diagnosis method comprises the following steps that eye multi-modal data are synchronously collected, a space-time tensor is constructed, and optical collection parameters are dynamically adjusted based on real-time environment illumination; performing multi-scale feature enhancement and tensor decomposition under manifold constraint on the space-time tensor, and extracting low-dimensional core features; according to the core feature dynamic optimization processing parameters, balancing feature reconstruction, classification precision and system energy consumption through multi-target joint optimization; the distributed cooperative processing of the multi-node detection task and the generation of a visual diagnosis report, wherein the node distribution is constrained by the spatial distance. According to the visual dysfunction diagnosis system and method, the precision and practicability of the visual dysfunction diagnosis system in a complex environment are improved by introducing technical means such as multi-modal fusion, illumination self-adaption, energy consumption optimization, D-S evidence collaboration and three-dimensional visualization.

Description

Technical Field

[0001] The present invention relates to the technical field of visual function diagnosis, and in particular to a high-precision visual function intelligent diagnosis method, system and electronic equipment. Background Art

[0002] Currently, early identification and intervention of visual dysfunction are key to improving the efficiency of diagnosis and treatment of visual diseases. To achieve more efficient and accurate auxiliary diagnosis capabilities, a multimodal fusion system that integrates structural imaging, functional analysis, and physiological signal perception is urgently needed.

[0003] Current mainstream technical solutions typically employ single-modal image processing algorithms to analyze retinal structure or rely on eye tracking data for behavioral pattern recognition. For example, lesion area segmentation is performed using OCT (optical coherence tomography) images, which are then combined with classification models to identify lesion types. Alternatively, statistical models are constructed using fixation points and saccade patterns from eye movement trajectories to aid in the diagnosis of neurological visual abnormalities. Furthermore, some systems incorporate temporal feature analysis methods to model pupil response characteristics or eye micromovements.

[0004] However, existing technologies still have some shortcomings: First, since existing methods are mostly based on single-modal input, they are unable to fully capture the structural-functional collaborative characteristics of retinal lesions. Especially in early lesions, due to weak signal expression or discrete distribution, a single modality often cannot provide sufficient basis for discrimination. Secondly, image acquisition is highly dependent on lighting conditions and lacks a dynamic adjustment mechanism, resulting in significant fluctuations in image quality under natural light interference conditions, seriously affecting the accuracy of downstream analysis. In addition, most existing systems use static parameter settings and centralized task processing methods, and are unable to flexibly schedule resources based on node status or task load, making it difficult for the system to operate efficiently on mobile devices or in resource-constrained scenarios. In terms of collaborative diagnosis, the lack of spatial constraint optimization mechanisms and robust data fusion strategies can easily lead to a lack of effective collaboration between distributed nodes and the existence of single-point failure risks. In terms of visualization, traditional two-dimensional result displays cannot accurately express the spatial location of the lesion and the dynamic characteristics of eye movement abnormalities, limiting the doctor's comprehensive judgment on the development trend of the disease. Summary of the Invention

[0005] The purpose of the present invention is to provide a high-precision intelligent visual function diagnosis method, system and electronic equipment, which solves the problems of existing visual function diagnosis systems in multimodal data fusion, environmental adaptability, resource optimization and scheduling, distributed collaborative computing and three-dimensional visualization expression.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a high-precision intelligent diagnosis method for visual function, comprising the following steps: S1. Synchronously collect multimodal eye data and construct a spatiotemporal tensor, dynamically adjusting optical acquisition parameters based on real-time ambient lighting; S2. Perform multi-scale feature enhancement and tensor decomposition under manifold constraints on the spatiotemporal tensor to extract low-dimensional core features; S3. Dynamically optimize processing parameters based on the core features, and balance feature reconstruction, classification accuracy, and system energy consumption through multi-objective joint optimization; S4, distributed collaborative processing of multi-node detection tasks and generation of visual diagnosis reports, where node allocation is constrained by spatial distance.

[0007] Preferably, the present invention also provides a high-precision intelligent visual function diagnosis system, comprising: The data acquisition and adjustment module is used to achieve synchronous acquisition and dynamic preprocessing of multimodal information, and combines multispectral sensing and real-time exposure control technology to perceive and adjust lighting conditions; The feature processing and decomposition module is used to construct the associated feature tensor including retinal structure, eye movement behavior and pupil physiological data, and perform feature extraction and compression processing through the manifold constrained tensor decomposition algorithm; Parameter optimization and control module, which is used to adaptively update system parameters based on Lyapunov stability theory and schedule computing resources and operation control under preset strategies; The collaborative processing and diagnosis module is used to aggregate the diagnostic information output by distributed edge nodes, combine the DS evidence theory to perform data fusion, and complete the calculation coordination and result synthesis based on the spatial location information and task allocation rules.

[0008] Preferably, the present invention also provides a computer device, comprising a memory, a processor and a communication module, wherein the memory stores executable instructions, and the processor implements the above-mentioned high-precision visual function intelligent diagnosis method when executing the instructions.

[0009] In summary, the present invention includes at least one of the following beneficial technical effects: 1. This method constructs a spatiotemporal correlation feature tensor by synchronously collecting retinal structure, eye movement function, and pupillary physiological data, effectively overcoming the limitations of single-modality data. By employing manifold-constrained tensor decomposition technology, the essential correlation features of multimodal data are retained during dimensionality reduction, significantly improving the specificity and sensitivity of lesion identification.

[0010] 2. This invention's dynamic ambient light adjustment mechanism eliminates the impact of external light interference on image quality through multispectral sensing and real-time exposure control. Combined with hardware-level clock synchronization and data buffering technology, it ensures the spatiotemporal consistency of multi-source heterogeneous data in complex environments, providing reliable input for subsequent analysis.

[0011] 3. This invention designs a parameter update law based on Lyapunov stability theory to achieve a dynamic balance between computational accuracy and system energy consumption. Using an improved multi-objective optimization algorithm to generate a Pareto-optimal solution set, this approach not only ensures diagnostic performance but also meets the stringent low-power and real-time requirements of mobile medical devices.

[0012] 4. This invention uses edge computing node clustering and UWB precise positioning technology to build a spatially constrained task allocation model. By integrating multi-node diagnostic results through DS evidence theory, it effectively reduces the risk of single-point failures and improves the system's fault tolerance and diagnostic reliability in complex deployment environments.

[0013] 5. This invention integrates raycast volume rendering and dynamic trajectory annotation technology to generate a visual report that includes the spatial distribution of lesions and abnormal eye movement patterns. This intuitive presentation of multidimensional data helps doctors quickly locate lesions, shortening diagnostic decision-making time and making it particularly suitable for screening and follow-up observation of early-stage lesions. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Schematic diagram of the method flow of the present invention; Figure 2 Schematic diagram of the system structure of the present invention; Figure 3 Schematic diagram of the computer device structure of the present invention. DETAILED DESCRIPTION

[0015] The following is combined with Figure 1 -Attached Figure 3 , the present invention is described in further detail.

[0016] This invention provides a high-precision intelligent visual function diagnosis method, system, and electronic device. By incorporating technologies such as multimodal fusion, adaptive illumination, energy consumption optimization, DS evidence collaboration, and three-dimensional visualization, it improves the accuracy and practicality of the visual dysfunction diagnosis system in complex environments. like Figure 1 As shown, the high-precision visual function intelligent diagnosis method may include the following steps: S1. Synchronously collect multimodal eye data and construct a spatiotemporal tensor, dynamically adjusting optical acquisition parameters based on real-time ambient lighting; In this embodiment, the synchronous acquisition and dynamic adjustment of multimodal data are achieved through the collaboration of multi-source heterogeneous sensing units. Specifically, an optical coherence tomography imaging unit is used to acquire retinal tomographic structural data. This imaging unit preferably uses a near-infrared broadband light source and extracts micron-level resolution cross-sectional images of biological tissue through the principle of interferometry. Simultaneously, eye movement trajectories are captured using a high-frame-rate eye tracking unit. This unit is based on the principle of pupil-corneal reflection and uses an arrayed infrared LED light source and a high-speed CMOS sensor to achieve sub-pixel tracking accuracy.

[0017] Dynamic adjustment of ambient light is based on a multi-spectral sensing unit, which includes a spectrophotometer and a light intensity sensor array. When a sudden change occurs in the visible light band, according to the preset exposure control model ;in, is the central wavelength; is the device constant; To prevent the elimination of zero factors; is the camera exposure time; The central wavelength The light intensity at the location; real-time adjustment of the imaging unit integration time, where It is preferably set to 550nm to match the spectral sensitivity of the human eye and prevent zero constant The introduction of effectively avoids the calculation overflow problem under extreme low illumination conditions.

[0018] In the data fusion stage, the collected multimodal data are constructed into a spatiotemporal tensor , where the spatial dimension , Determined by the imaging unit resolution, the time dimension Corresponding to the preset acquisition time window, channel dimension This data consists of three types of heterogeneous data: structured light intensity, eye movement displacement vector, and pupil diameter. During tensor construction, a timestamp alignment module eliminates clock drift between sensor units, and cubic spline interpolation is used to unify the timing benchmarks for data at different sampling rates.

[0019] The dynamic adjustment mechanism further includes an adaptive weight allocation strategy: when the ambient light intensity Exceeding the threshold When the threshold is set, the weight of structured light imaging is automatically reduced and the sampling rate of eye tracking data is increased. This strategy is implemented through the hardware abstraction layer instruction set to ensure the balance of multimodal data quality under complex lighting conditions. It is preferred to dynamically update based on the statistical characteristics of historical data, and the update cycle is synchronized with the equipment calibration cycle.

[0020] In this embodiment, the exposure time control model and the parameter calibration of the multi-spectral sensing unit are realized through an offline calibration process. In the calibration stage, a standard light source is used to illuminate the integrating sphere to generate a known light intensity distribution, and the device sensitivity coefficient is obtained by least squares fitting. The calibration data is stored in the device’s encrypted storage area and is called through a digital signature verification mechanism during the online working phase.

[0021] The time synchronization accuracy of multimodal data is guaranteed by the following technical means: a precise clock source is deployed in the main control unit, and the IEEE1588v2 precision time protocol is used to achieve clock synchronization of each submodule; priority queue scheduling is implemented in the data transmission layer, and the highest real-time priority is given to eye movement trajectory data; a data buffer pool is set up in the application layer, and the timing deviation caused by instantaneous jitter is eliminated through the sliding window mechanism.

[0022] S2. Perform multi-scale feature enhancement and tensor decomposition under manifold constraints on the spatiotemporal tensor to extract low-dimensional core features; In this embodiment, multi-scale feature enhancement and manifold constraint decomposition are achieved by constructing a hierarchical differential operator and coordinating it with manifold space projection. Specifically, anisotropic multi-scale differential operations are used to extract multi-level edge features of retinal images, where the differential operator is designed as a scale-weighted fusion form: ; in, Represents the image gradient features after multi-scale enhancement; is the input eye image matrix; express Directional Order partial derivative operator; express Directional Order partial derivative operator; is the differential scale parameter; the attention matrix is ​​used Enhance features in key areas; is the scale weight coefficient. The introduction of makes large-scale features have a higher contribution during fusion and effectively suppresses high-frequency noise interference.

[0023] In key areas such as the macula, a dynamic attention weight adjustment mechanism is implemented: ; in, is the local mean of the image; is the standard deviation; is the Sigmoid function; is the pixel intensity value at the coordinate in the image. This weight map is combined with the multi-scale gradient feature to perform a Hadamard product operation, which increases the response intensity of the pathological feature-significant area by 30%-50%.

[0024] The Tucker decomposition of manifold constraints is achieved by the following technical means: First, the eye movement trajectory manifold is constructed based on the phase space reconstruction method , preferably, the time delay embedding method is used to map the one-dimensional eye movement signal into the three-dimensional phase space, and its fractal dimension is calculated as the manifold dimension constraint. Then the core tensor is initialized rank parameter to satisfy .

[0025] Perform rank-constrained Tucker decomposition: ; in, is the original space-time tensor; is the core tensor; For the modular factor matrix; express Modulo tensor product; constraints satisfied ;in, is the eye movement phase space manifold; is the Hausdorff dimension of the eye movement phase space manifold.

[0026] The manifold projection operation is implemented by the following steps: At each factor matrix After updating, its column vector is projected onto the manifold tangent space constructed by historical eye movement data, which is calculated by principal component analysis.

[0027] In this embodiment, the multi-scale differential operator is implemented by using separable convolution technology to optimize computational efficiency. Specifically, the two-dimensional convolution kernel is decomposed into one-dimensional kernels in the horizontal and vertical directions for separate operations, which reduces the computational complexity from Reduce to The computation process is accelerated on the GPU side through texture memory, leveraging hardware features to improve data access locality.

[0028] S3, dynamically optimize processing parameters based on core features, and balance feature reconstruction, classification accuracy and system energy consumption through multi-objective joint optimization; In this embodiment, dynamic parameter optimization and control are achieved by cooperating with Lyapunov stability theory and multi-objective Pareto optimization. Specifically, the system state vector is constructed. Characterize the real-time status of the computing unit's temperature, memory usage, and communication bandwidth, and design the Lyapunov function based on this status ,in, is the system state vector; is a positive definite weight matrix; the weight matrix A diagonal matrix is ​​preferably used to decouple the coupling effects of different state variables, and its diagonal elements are set according to the thermodynamic characteristics of the equipment and the level of mission criticality.

[0029] Design parameter update equation based on Lyapunov stability theory: ; in, for System parameter vector at time ; is the system state vector, LL is the composite objective function including reconstruction loss and classification loss; is the gradient learning rate; is the stability coefficient. , Through the dynamic adjustment of the online learning mechanism, preferably, the sliding window variance analysis method is used to evaluate the system stability. Automatically increase when >0.1 Weight.

[0030] The multi-objective optimization problem is modeled as: Formulate the Pareto optimization problem: ; in, is the optimizable parameter vector; is the original space-time tensor; To reconstruct the tensor; Cross entropy loss for disease classification; is the system energy consumption; And impose the gradient norm constraint ;in, is the gradient vector of the total loss function; represents the L2 norm; is the gradient threshold.

[0031] The normalization of the weight coefficients is achieved through the soft maximization function: ; in, is the weight coefficient; is the standardized score of each objective function, the temperature parameter Control the sharpness of weight distribution; is the target function index; is the base of natural logarithms.

[0032] In this embodiment, the system status monitoring module achieves real-time data collection through the following technical means: embedded sensors are deployed in the computing unit to acquire temperature and power consumption data at a 100Hz sampling rate; a probe program is embedded in the communication interface layer to calculate bandwidth utilization and transmission latency. Data is transmitted to the optimization decision module via a shared memory area, using a double buffering mechanism to avoid read and write conflicts.

[0033] Dynamic parameter adjustment is achieved through a separate control-decision-making architecture: the control thread performs parameter update calculations at a fixed frequency, while the decision thread asynchronously runs the optimization algorithm to generate candidate solutions. Data is exchanged between the two threads via a thread-safe queue, and semaphores are used to ensure state synchronization.

[0034] S4, distributed collaborative processing of multi-node detection tasks and generation of visual diagnosis reports, where node allocation is constrained by spatial distance; In this embodiment, distributed collaborative diagnosis is achieved through the collaboration of edge computing node clustering and spatial constraint optimization. Specifically, multiple edge computing nodes are deployed to form a distributed processing network, and each node obtains three-dimensional coordinates in real time through an ultra-wideband (UWB) positioning system. ,Preferably, the positioning system operating frequency is set in the 6.5-8.5GHz frequency band to ensure centimeter-level positioning accuracy. Node spacing constraints Implemented through a distributed consensus algorithm, when it is detected that node movement causes the distance to be less than the safety threshold, the node migration or task reallocation mechanism is automatically triggered.

[0035] The task allocation model is constructed as a delay minimization problem: ; in, To calculate the delay; is the communication delay; is the total number of distributed nodes; is the three-dimensional coordinate of the node; is the minimum anti-interference distance; For the The three-dimensional coordinates of the nodes; For the The three-dimensional coordinates of the nodes; represents the Euclidean distance.

[0036] Preferably, the sliding window method is used to statistically analyze the distribution characteristics of historical task computational loads, and exponential smoothing is used to predict future computational demands. Dynamic evaluation is achieved through probe packet injection, which periodically sends test packets to the central node to measure the actual bandwidth.

[0037] In the optimization solution stage, a hybrid strategy of branch and bound method and genetic algorithm is adopted: first, an initial feasible solution set is generated, and candidate solutions are screened by calculating the load balance degree of each node; then, the crossover and mutation operation of the genetic algorithm is performed in the neighborhood of feasible solutions. Preferably, the crossover probability is set to 0.7, and the mutation probability decays linearly with the number of iterations.

[0038] The diagnostic report generation module implements multi-source data fusion: it receives local diagnostic results from each node (including the coordinates of the retinal lesion area, abnormal point marks of the eye movement trajectory, and quantitative indicators), and performs confidence fusion through DS evidence theory. Specifically, it defines the basic probability distribution function and uses the orthogonal sum formula: ; in, For the first piece of evidence, The basic probability distribution of For the second piece of evidence, the hypothesis The basic probability distribution of Assume the target to be calculated; is the intersection operation of the hypothesis set; The independent evidence of each node is synthesized, and the final decision is made based on the maximum confidence hypothesis.

[0039] In this embodiment, the 3D visualization engine uses a ray casting algorithm to render retinal tomography images, converts OCT volume data into RGBA transfer function mapping, and generates a 3D structural view through a GPU-accelerated volume rendering pipeline. The eye movement trajectory data is fitted with a B-spline curve to generate a smooth path, and abnormal point marks are highlighted using a sphere bounding box with a sphere radius. and anomaly confidence satisfy The mapping relationship.

[0040] The implementation of spatial constraints is guaranteed by the following technologies: deploying Kalman filter in the positioning data stream processing layer to predict the node motion trajectory and calculate the collision risk in advance; using directional antenna to adjust the signal radiation pattern in the physical layer. The transmission power is automatically reduced to below -10dBm to avoid electromagnetic interference. Clock synchronization between nodes is achieved through the WhiteRabbit protocol, and the time deviation is controlled within ±10ns.

[0041] The high-precision visual function intelligent diagnosis system described below and the high-precision visual function intelligent diagnosis method described above can be referred to in correspondence with each other.

[0042] Please see the attached Figure 2 The present invention also provides a high-precision visual function intelligent diagnosis system, comprising: The data acquisition and adjustment module is used to achieve synchronous acquisition and dynamic preprocessing of multimodal information, and combines multispectral sensing and real-time exposure control technology to perceive and adjust lighting conditions; The feature processing and decomposition module is used to construct the associated feature tensor including retinal structure, eye movement behavior and pupil physiological data, and perform feature extraction and compression processing through the manifold constrained tensor decomposition algorithm; Parameter optimization and control module, which is used to adaptively update system parameters based on Lyapunov stability theory and schedule computing resources and operation control under preset strategies; The collaborative processing and diagnosis module is used to aggregate the diagnostic information output by distributed edge nodes, combine the DS evidence theory to perform data fusion, and complete the calculation coordination and result synthesis based on the spatial location information and task allocation rules.

[0043] The system of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, so they will not be repeated here.

[0044] Please see the attached Figure 3 The present invention also provides a computer device, including a memory, a processor and a communication module. The memory stores executable instructions, and the processor implements the above-mentioned high-precision visual function intelligent diagnosis method when executing the instructions.

[0045] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A high-precision intelligent diagnostic method for visual function, characterized in that: The following steps are involved: S1. Synchronously collect multimodal eye data and construct a spatiotemporal tensor, dynamically adjusting optical acquisition parameters based on real-time ambient lighting; S2. Perform multi-scale feature enhancement and tensor decomposition under manifold constraints on the spatiotemporal tensor to extract low-dimensional core features; S3. Dynamically optimize processing parameters based on the core features, and balance feature reconstruction, classification accuracy, and system energy consumption through multi-objective joint optimization; S4, distributed collaborative processing of multi-node detection tasks and generation of visual diagnosis reports, where node allocation is constrained by spatial distance.

2. The high-precision intelligent diagnosis method for visual function according to claim 1, characterized in that: The steps of synchronously collecting multimodal eye data and constructing a spatiotemporal tensor include: Synchronously collect structural, functional, and physiological data through OCT imaging modules, eye trackers, and pupil detectors; Construct the collected data into a four-dimensional space-time tensor ,in 、 is the spatial dimension; is the length of the time series; is the number of modal channels.

3. The high-precision intelligent diagnosis method for visual function according to claim 2, characterized in that: The dynamic adjustment of ambient light includes: Acquire ambient light wavelength distribution through multispectral sensor ; Dynamically adjust the camera exposure time according to the following formula: ； in, is the central wavelength; is the device constant; To prevent the elimination of zero factors; is the camera exposure time; The central wavelength The light intensity at.

4. The high-precision intelligent diagnosis method for visual function according to claim 1, characterized in that: The step of multi-scale feature enhancement includes: Construct anisotropic multiscale differential operators: ； in, Represents the image gradient features after multi-scale enhancement; is the input eye image matrix; express Directional Order partial derivative operator; express Directional Order partial derivative operator; is the differential scale parameter; the attention matrix is ​​used Enhance features in key areas; is the scale weight coefficient.

5. The high-precision intelligent diagnosis method for visual function according to claim 4, characterized in that: The tensor decomposition step under the manifold constraint includes: Perform rank-constrained Tucker decomposition: ； in, is the original space-time tensor; is the core tensor; For the modular factor matrix; express Modulo tensor product; constraints satisfied ;in, is the eye movement phase space manifold; is the Hausdorff dimension of the eye movement phase space manifold.

6. The high-precision intelligent diagnosis method for visual function according to claim 1, characterized in that: The step of dynamically optimizing processing parameters according to the core features comprises: Design parameter update equation based on Lyapunov stability theory: ； in, for System parameter vector at time ; is the system state vector, LL is the composite objective function including reconstruction loss and classification loss; is the gradient learning rate; is the stability coefficient.

7. The high-precision intelligent diagnosis method for visual function according to claim 6, characterized in that: The steps of the multi-objective joint optimization include: Formulate the Pareto optimization problem: ； in, is the optimizable parameter vector; is the original space-time tensor; To reconstruct the tensor; Cross entropy loss for disease classification; is the system energy consumption; And impose the gradient norm constraint ;in, is the gradient vector of the total loss function; represents the L2 norm; is the gradient threshold.

8. The high-precision intelligent diagnosis method for visual function according to claim 1, characterized in that: The distributed collaborative processing in step 4 includes: Establish a node task allocation model: ； in, To calculate the delay; is the communication delay; is the total number of distributed nodes and satisfies the node spacing constraint ;in, is the three-dimensional coordinate of the node; is the minimum anti-interference distance; For the The three-dimensional coordinates of the nodes; For the The three-dimensional coordinates of the nodes; represents the Euclidean distance; In the spatial distance constraint: Minimum node spacing , and the node coordinates Real-time updates via UWB positioning system.

9. High-precision visual function intelligent diagnosis system, characterized by: The high-precision intelligent diagnosis method for visual function according to any one of claims 1 to 8 comprises: The data acquisition and adjustment module is used to achieve synchronous acquisition and dynamic preprocessing of multimodal information, and combines multispectral sensing and real-time exposure control technology to perceive and adjust lighting conditions; The feature processing and decomposition module is used to construct the associated feature tensor including retinal structure, eye movement behavior and pupil physiological data, and perform feature extraction and compression processing through the manifold constrained tensor decomposition algorithm; Parameter optimization and control module, which is used to adaptively update system parameters based on Lyapunov stability theory and schedule computing resources and operation control under preset strategies; The collaborative processing and diagnosis module is used to aggregate the diagnostic information output by distributed edge nodes, combine the DS evidence theory to perform data fusion, and complete the calculation coordination and result synthesis based on the spatial location information and task allocation rules.

10. A computer device comprising a memory, a processor and a communication module, characterized in that: The memory stores executable instructions, and when the processor executes the instructions, the high-precision visual function intelligent diagnosis method according to any one of claims 1 to 8 is implemented.

Images