A portable perimetry dynamic threshold detection method and system based on interactive response

Abstract

The application discloses a portable perimetry dynamic threshold detection method and system based on interactive response. In view of the defects of existing perimetry, such as redundant dense sampling, misjudgment caused by subject cognitive fatigue, and interference of the pose of a portable device, the application first constructs an asymmetric topological weight matrix based on the arcuate trend of retinal neuroanatomy, and performs active learning dynamic target optimization by fusing Shannon information entropy, so that dimension reduction and rapid sampling are realized. Secondly, the patient's transient response delay is quantified as a cognitive load index, the cursor stimulation brightness step is closed-loop nonlinearly adjusted, and false positives caused by visual hesitation are avoided. Finally, the spatial six-axis pose variance and the first derivative of the key time of the device are extracted at a high frequency, the true and false are decoupled through a cross-modal confidence scoring network, and dirty data is physically peeled off. Under the premise of ensuring the clinical detection rate, the application greatly compresses the detection time, and gives the portable device high anti-interference and environmental adaptability.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and ophthalmic medical device processing technology, specifically relating to a method and system for dynamic threshold detection of visual field by quantifying patient interaction response characteristics in primary healthcare scenarios, relying on portable devices such as virtual reality (VR). Background Technology

[0002] Perimetry, as the clinical "gold standard" for assessing the functional status of the optic nerve and retina, plays an irreplaceable role in the early screening and follow-up of blinding neuro-ophthalmic diseases such as glaucoma. Currently, the mainstream clinical standard automated perimetry (SAP) and its derivative portable devices, when deployed in primary healthcare settings (such as township health centers and community elderly care institutions), reveal technical bottlenecks in their underlying architecture and algorithm mechanisms.

[0003] First, the static traversal strategy based on the isotropic assumption leads to system redundancy and long detection time. The underlying detection algorithm of existing mainstream perimeters is essentially a quasi-exhaustive traversal based on a fixed spatial grid. This control logic has inherent limitations: it assumes that adjacent test points have "isotropic correlation" in spatial distribution, failing to fully incorporate the arcuate anatomical topological isomorphism of the human retinal nerve fiber layer (RNFL). This results in redundancy in computational power and time in healthy areas, while lacking adaptive high-density tracking at nerve bundle damage boundaries. This equal-density sampling typically makes single-eye detection take 10 to 15 minutes, which is insufficient to meet the efficient screening needs of large populations at the grassroots level.

[0004] Second, traditional open-loop control lacks temporal closed-loop feedback on psychophysical states, leading to reduced data confidence in noisy environments. Existing systems' threshold approximation mechanisms assume subjects maintain constant cognitive load and attention during darkroom testing. However, when targeting elderly individuals with cognitive decline, prolonged, high-frequency visual stimuli can easily induce cognitive exhaustion and fatigue stress. Current technologies lack the ability to quantify and compensate for real-time response delays, resulting in random responses (false positives) or missed presses due to fatigue in the threshold region. Existing devices struggle to clean these aberrations from the underlying data stream, easily creating false lesions in the output visual field map.

[0005] Third, existing portable VR devices lack multimodal anti-disturbance mechanisms for non-standard degrees of freedom environments. In recent years, most portable vision meters based on virtual reality (VR) have only focused on optical carrier migration, without reconstructing their core control flow for portable edge computing scenarios. In unrestricted environments lacking professional darkrooms and chin rest fixation, involuntary micro-movements of the subject's head can disrupt the mapping coordinates of the virtual cursor on the retina. Existing portable devices lack cross-modal temporal joint alignment and decoupling of spatial pose sensors (IMUs) and interactive response signals, easily misjudging "physiological visual field misses" caused by head movements as "pathological blind spots," affecting the detection specificity of the device in non-standard environments at the grassroots level.

[0006] In summary, there is a need in this field for a portable field of view detection system with full-element perception and adaptive control capabilities. This system needs to overcome the limitations of fixed grid scanning and fixed step size, introduce an active learning mechanism based on medical anatomical topology, and construct a noise decoupling model based on cross-modal temporal dynamics to achieve high-precision, highly robust, and fast closed-loop determination of the field of view threshold on edge devices with limited computing power. Summary of the Invention

[0007] To address the fundamental shortcomings of existing static field-of-view detection technologies, such as "redundant and time-consuming detection due to the isotropic assumption," "low data confidence due to the lack of cognitive load decoupling in open-loop control," and "lack of spatial pose anti-disturbance mechanisms in existing portable devices," the primary objective of this invention is to provide a dynamic threshold detection method for portable field-of-view meters based on interactive response.

[0008] Specifically, this invention aims to overcome the bottlenecks of existing technologies through the following multi-dimensional technological restructuring:

[0009] 1. Introducing an active learning dimensionality reduction mechanism to achieve adaptive detection. This invention aims to overcome the logical limitations of traditional fixed-grid scanning by introducing the anatomically isomorphic topological prior of the optic nerve fiber layer (RNFL) in glaucoma and integrating an active learning dynamic target optimization algorithm based on Shannon information entropy. By constructing a joint objective cost function, it endows portable edge computing systems with nonlinear dimensionality reduction detection capabilities, aiming to effectively shorten the time for monocular visual field threshold determination while ensuring clinical-grade detection rates, thus meeting the needs of primary care screening.

[0010] 2. Constructing a closed-loop feedback based on psychophysical temporal dynamics to reduce data distortion caused by cognitive exhaustion. Addressing the cognitive decline of primary care subjects and the interactive noise caused by long-term testing, this invention aims to establish a threshold compensation mechanism based on interactive response characteristics. By extracting the patient's transient response delay sequence in real time and quantifying it as a "cognitive load index," the stimulus step size of the cursor brightness is dynamically and adaptively scaled non-linearly. This aims to flexibly resolve threshold misjudgments caused by visual hesitation at the algorithmic level, effectively reducing false positive and false negative rates.

[0011] 3. Overcoming spatial pose drift in portable, non-standard environments and achieving cross-modal anti-disturbance feature decoupling. For screening environments lacking professional darkrooms and chin-down restrictions, this invention aims to provide a highly robust data cleaning model. By performing deep multimodal cross-domain fusion of the portable device's spatial inertial measurement unit (IMU) matrix and the first-order evolution gradient of the interactive time series, non-visual artifacts caused by involuntary head movements and fatigue-induced distraction are accurately removed. This effectively solves the defect of existing VR perimeters that easily misjudge "physiological visual field misses" as "pathological blind spots," improving the anti-interference capability of edge devices. To achieve the above objectives, this invention novelly provides the following technical solution: a portable perimeter dynamic threshold detection method and system based on interactive response, including...

[0012] The overall hardware execution base of this invention includes:

[0013] Portable control host equipped with an edge computing microprocessor (such as an NPU or ARM architecture core).

[0014] Virtual reality (VR) head-mounted display device with built-in inertial measurement unit (IMU, supporting six-axis spatial pose capture)

[0015] And a wireless interactive controller with millisecond-level response latency

[0016] Based on the aforementioned underlying hardware platform, this invention provides a portable field of view dynamic threshold detection method based on interactive response. Its core technical solution includes the following sequentially executed and closed-loop coupled method steps:

[0017] Step S1: Multidimensional Spatiotemporal Interaction Baseline Anchoring and Asymmetric Prior Topology Graph Construction

[0018] S11 (Topology layer initialization):

[0019] At the system's underlying layer, within the two-dimensional polar coordinate mapping space of the global field of view, an anisotropic arcuate anatomical orientation of the human retinal nerve fiber layer (RNFL) is deeply embedded, instantiating an asymmetric "neural topological prior weight matrix." This matrix is ​​used to constrain the geodesic spatial correlation between any two discrete test coordinate points within the visual field space on the same physical trajectory of the same nerve bundle, based on underlying mathematical relationships.

[0020] S12 (Psychophysical Baseline Anchoring):

[0021] During the initial test phase, the system applies high-contrast threshold cursor stimulation to high-confidence anchor point regions such as the physiological blind spot and central visual field. Simultaneously, the system captures the patient's transient response time series in a fatigue-free state and extracts the individual-specific baseline motor key delay time. Based on the basic physiological variance, a personalized psychophysical time reference window is constructed.

[0022] Step S2: Dynamic target dimensionality reduction optimization mechanism based on active learning and information entropy

[0023] S21 (Posterior probability update):

[0024] The system abandons the traditional preset grid polling scanning sequence. A Bayesian inference engine running within the microprocessor dynamically updates the posterior probability distribution of pathological scotomas (visual defects) at all untested spatial discrete points in real time based on historical interaction feedback. Specifically, when a missing response is detected at any coordinate point (i.e., a lesion is confirmed), the system uses that point as observational evidence, combined with the "neural topology prior weight matrix" generated in step S11, and utilizes Bayesian rules to simultaneously increase the posterior probability of defects at all adjacent untested points located on the same nerve bundle trajectory as the lesion. This establishes the medical probability evolution rule of "lesion spreading along nerve fibers" at the algorithm's underlying layer.

[0025] S22 (Joint Target Optimization):

[0026] The system constructs a joint objective cost function that integrates Shannon entropy (representing the information gain limit prediction) and neural topological prior weight matrix (representing lesion spread traction). In each forward projection iteration, the optimal solution of this function is solved in real time, prioritizing the next round of target point projection to the spatial coordinate system that represents the "current uncertainty peak of the system" and is "adjacent to the known suspected lesion neural bundle". This enables extremely sparse dimensionality reduction sampling of low-risk healthy retinal regions and adaptive high-density tracking of high-risk edge zones.

[0027] Step S3: Nonlinear stepped stimulus closed-loop compensation integrating endogenous cognitive load index

[0028] S31 (Load Quantification Extraction):

[0029] When projecting cursor stimulation onto the optimal target point selected in step S2, the system accurately measures and extracts the real-time response time of this interaction. ), calculate its relative reference time window ( The time delay deviates from the gradient;

[0030] S32 (Step Size Adaptive Decay):

[0031] The time delay deviation gradient is mapped to a "cognitive load index" representing the patient's threshold of light sensitivity. This index is hard-coded as a negative decay factor into the Markov decision iteration equation for stimulus cursor brightness (dB). When the index suddenly increases (indicating the patient is in the threshold region of visual hesitation or cognitive overload), the system automatically and seamlessly switches from a large step size to a minimum probe microstep size mode using a non-linear smooth scaling mechanism. This avoids threshold oscillations and misjudgments caused by excessive brightness stimulus span.

[0032] Step S4: Dirty data stripping and true / false decoupling driven by cross-modal temporal dynamics

[0033] S41 (Heterogeneous Feature High-Frequency Latch):

[0034] In response to abnormal triggering events that may occur during the detection process, such as "no response timeout" or "rapid pressing", the system's underlying scheduler uses high-frequency synchronous latching of the six-axis spatial pose flow of the VR headset's built-in IMU (focusing on extracting the transient fluctuation variance of the yaw and pitch velocity matrices), as well as the first-order dynamic derivative of the button time series of the interactive controller.

[0035] S42 (Multimodal Hard Decoupling and State Reset):

[0036] The aforementioned heterogeneous spatial modalities (pose micro-motion features) and temporal modalities (interactive rhythmic abrupt change features) are fed into an embedded multimodal confidence scoring network. Strict true / false decoupling logic is implemented.

[0037] Pathological record: If the joint assessment determines that the head is rigid and stable and the key rhythm is stable, then the missing response of this coordinate is anchored as "real optic nerve injury lesion";

[0038] Physiological stripping: If the IMU pose drift exceeds the limit or the response rhythm shows typical fatigue decay characteristics, it is judged as "non-visual motion artifact / mental distraction". The system immediately triggers the physical isolation mechanism to strip the dirty data of the frame and activates the preset audio-visual wake-up protocol to suspend the subject's state, and resets the test task of the spatial sector at an opportune time.

[0039] Furthermore, the present invention also provides a portable perimeter dynamic threshold detection system for performing the above method, comprising at least:

[0040] Heterogeneous multi-source data synchronization sensing and benchmark anchoring module: used to establish a two-dimensional polar coordinate mapping space and RNFL weight matrix, and extract the underlying hardware platform's... data;

[0041] The topological target adaptive optimization module based on active learning performs Bayesian posterior probability updates and joint target optimization calculations, and outputs the spatial three-dimensional coordinates of the projected target in the next frame.

[0042] Cognitive load-driven nonlinear brightness compensation module: extracts real-time response delay to calculate cognitive load index, and controls the light source brightness gradient step size of VR rendering engine in a closed loop;

[0043] Cross-modal dynamics true / false decoupling and data cleaning module: integrates IMU six-axis data stream and first-order derivative of key press time, performs multimodal confidence scoring and physical isolation of dirty data. Attached Figure Description

[0044] Figure 1 System physical architecture and interaction flow diagram based on portable field of view meter

[0045] Figure 2 Flowchart of Dynamic Threshold Detection Algorithm

[0046] Figure 3 Visual neural topology and active learning spatial optimization principle diagram

[0047] Figure 4 : Logic block diagram for multimodal dirty data confidence cleaning Detailed Implementation

[0048] This invention relies on a system that includes an edge computing microprocessor (with computing power supporting INT8 / FP16 hardware acceleration) and a high refresh rate VR headset (refresh rate... Built-in six-axis IMU, polling frequency ) and low-latency Bluetooth interactive controller (button anti-shake time) The hardware foundation of this system. The core dynamics model and algorithm steps executed at the underlying level are as follows:

[0049] 1. Core Implementation of Phase S1: RNFL Topology Weight Matrix ( Mathematical initialization of )

[0050] To overcome the technical bias of "isotropy", the system pre-defined a two-dimensional polar coordinate system in the global view. A standard retinal nerve fiber layer (RNFL) arcuate anatomical model was imported into China.

[0051] For any two discrete test coordinate points and The system does not calculate its straight-line Euclidean distance, but instead calculates its geodesic distance along the physiological nerve bundle trajectory. Based on this, the system's underlying layer initializes the asymmetric topological correlation weight matrix:

[0052]

[0053] in, This represents the physiological topological diffusion coefficient. The physical meaning of this matrix is: if point... Visual field defects were detected. Due to the conduction and damage characteristics of nerve fibers, points located on the same nerve bundle trajectory... The probability of developing the disease will increase exponentially, while adjacent points that cross the horizontal suture zone (not in the same nerve bundle) will remain unaffected.

[0054] 2. Core Implementation of Phase S2: Joint Objective Optimization Dynamics Based on Shannon Information Entropy

[0055] Before each frame is projected, the Bayesian inference engine within the microprocessor updates the posterior probability of the lesion for the set U of all untested sites in real time. The execution logic of its update rule is as follows: if the points have been measured... Once confirmed as a genuine field-of-view defect, the system will then convert the topology matrix. Substituting the prior conditional probability into Bayes' formula, we calculate and amplify the unmeasured points that have a high topological correlation with it. The post-mortem probability of the disease.

[0056] Based on the updated probability matrix, the system calculates the unmeasured points. Independent Shannon information entropy :

[0057]

[0058]

[0059]

[0060]

[0061] 3. Core Implementation of Phase S3: Nonlinear Brightness (dB) Compensation Model Driven by Cognitive Load Closed-Loop

[0062] In targeting When performing step threshold measurement, the system needs to dynamically determine the brightness of the stimulus cursor in the next frame. (Unit: dB)

[0063] The system extracts the patient's current actual response time via high frequency using a Bluetooth handset. And its underlying key delay anchored in the S1 phase. By comparison, an intrinsic brightness compensation equation is constructed:

[0064]

[0065] in, The system's preset basic width approximation step size; To ensure the minimum probe microstep size for test progression; This is the cognitive depletion decay constant.

[0066] Engineering execution logic: When the subject is in the threshold region and hesitates due to visual blur (i.e. When the exponential decay term converges sharply to near zero, the system automatically and smoothly scales the adjustment step size of the next frame from a coarse, wide step size to an extremely small, high-precision micro-step size. This avoids "patients blindly pressing buttons" and "threshold oscillations" caused by step overshoot at the system level, while preventing the algorithm from getting stuck in a dead loop with zero step size in an extreme state of hesitation, thus achieving a two-way closed loop of test accuracy and patient psychological compliance.

[0067] 4. Core Implementation of Phase S4: Multidimensional Confidence Cleaning Network with Cross-Modal Temporal Decoupling

[0068] To eliminate dirty data from portable, non-standard environments, the system constructs a confidence scoring function that integrates spatial attitude and temporal rhythm. .

[0069] When the system detects a single response event (including missed presses due to timeout), it extracts the variance of the six-axis angular velocity of the head-mounted display IMU within the current time window T in real time. (Indicating whether the head is rigid and stable), and the absolute value of the first-order evolutionary kinetic derivative of the button response time. (This characterizes whether there are sudden changes in the rhythm of movement, such as fatigue, sluggishness, or rapid, random pressing). The scoring model is as follows:

[0070]

[0071] in, Based on the confidence bias term, For cross-modal penalty weights.

[0072] Execution logic: This function is a non-linear inverse logic gate based on a Sigmoid variant. It will only function if and only if there is no significant head offset. When the value is extremely small and the key rhythm conforms to normal physiological conditions (the absolute value of the derivative tends to be stable), Only when the value exceeds the hard threshold (preset to 0.75) will the missing response be confirmed as a "true blind spot". Conversely, if high-frequency head shaking or an exponential change in time response is detected (whether due to positive fatigue or negative random pressing), the system determines... The data in the frame is immediately physically overwritten and destroyed, and the front-end GUI rendering interface is reset and the sound and light wake-up protocol is triggered.

Claims

1. A portable perimeter dynamic threshold detection method based on interactive response, characterized in that, The method is completely independent of fixed spatial grids and open-loop blind scan logic, and includes the following sequentially executed and closed-loop coupled steps: S1. Multidimensional Spatiotemporal Interaction Baseline Anchoring and Asymmetric Prior Topology Construction: In a two-dimensional polar coordinate mapping space, the arcuate anatomical orientation of the human retinal nerve fiber layer is deeply embedded to generate an asymmetric neural topology prior weight matrix; at the same time, the transient response time series of the subject in a fatigue-free state during the test initialization phase is extracted to anchor the basic key delay time to construct a psychophysical benchmark window. S2. Dynamic target dimensionality reduction optimization based on active learning and information entropy: Through the Bayesian inference engine running in the microprocessor, the posterior probability distribution of visual defects of all unmeasured spatial discrete points in the world is updated in real time; and a joint objective cost function that integrates Shannon information entropy and the neural topology prior weight matrix is ​​constructed, and the next round of threshold detection is initiated in real time to the spatial coordinate point of the neural bundle with the highest uncertainty in the system and adjacent to the known lesion. S3. Nonlinear stepwise stimulus closed-loop compensation with endogenous cognitive load index: The real-time response time of the target stimulus projected in step S2 is measured at high frequency, and its time delay deviation gradient relative to the psychophysical reference window is calculated and quantified as a cognitive load index; the cognitive load index is used as a negative attenuation factor to nonlinearly and smoothly scale the stepwise stimulus step size of the cursor brightness in the next frame. S4. Dirty data stripping and true / false decoupling driven by cross-modal temporal dynamics: For abnormal triggering events during the detection process, the fluctuation variance of the spatial pose of the virtual reality head-mounted display device and the first-order dynamic derivative of the interaction button time series are synchronously latched; the above heterogeneous modal data are fed into the multimodal confidence scoring network for true / false decoupling, and physical isolation and state reset are performed on dirty data that are judged as non-visual artifacts.

2. The detection method according to claim 1, characterized in that, The specific method for constructing the asymmetric neural topology prior weight matrix in step S1 is as follows: for any two discrete test coordinate points in the view mapping space... and The system does not calculate its straight-line Euclidean distance, but instead calculates its geodesic distance along the physiological nerve bundle trajectory. Based on the geodesic distance, an exponential decay model is constructed, and the neural topology prior weight matrix is ​​output. It is used to establish probabilistic constraints on the spread of lesions along the same nerve bundle trajectory from the underlying mathematical relationships.

3. The detection method according to claim 1, characterized in that, The specific logic for solving the joint objective cost function in step S2 is as follows: Before each round of projection, the optimal projection target point is solved and output using the following joint objective cost function formula. : Where U is the set of unmeasured points and K is the set of measured points; For the measured points The detection indicator is 1 when a defect is detected and 0 when the defect is healthy. This represents the independent Shannon information entropy of unmeasured points updated based on Bayesian posterior probability. The entropy gain weighting coefficient is... This is the topology traction weighting coefficient.

4. The detection method according to claim 1, characterized in that, The specific execution formula for nonlinear step-stress closed-loop compensation in step S3 is as follows: The system constructs an intrinsic brightness compensation equation to calculate the brightness of the stimulus cursor in the next frame. : in, For the extracted real-time response time, The base key delay time for anchoring; The system's preset basic width approximation step size, To ensure the minimum probe microstep size for test progression; This represents the cognitive depletion decay constant. When a patient experiences visual hesitation in the threshold region, it causes... At this point, the exponential decay term converges rapidly, and the system automatically scales the brightness adjustment step size to the minimum probe microstep size. .

5. The detection method according to claim 1, characterized in that, The multimodal confidence scoring network for true / false decoupling in step S4 is specifically as follows: Extract the variance of the six-axis angular velocity of the head-mounted display within the current time window T. And the absolute value of the first-order evolutionary dynamic derivative of the key response time Construct a nonlinear confidence scoring function based on a Sigmoid variant. : in, Based on the confidence bias term, For cross-modal penalty weights.

6. The detection method according to claim 5, characterized in that, The specific logic for determining physical isolation and state reset in step S4 is as follows: If and only if there is no significant head deviation and the key-pressing rhythm conforms to normal physiological norms. If the response exceeds the preset threshold, the system will identify the corresponding missing response as a real blind spot and record it. When high-frequency head-shaking posture drift exceeds the limit or the response rhythm exhibits an exponential abrupt change, If the current interaction is below the preset judgment threshold, the system determines that the current interaction is a physiological visual miss or fatigue artifact, immediately physically overwrites and destroys the frame data, and triggers the reset of the front-end rendering interface and the audio-visual wake-up protocol.

7. A portable perimeter dynamic threshold detection system, used to perform the detection method according to any one of claims 1 to 6, characterized in that, The system includes: The physical hardware base includes a portable control host equipped with an edge computing microprocessor, a virtual reality head-mounted display device with a built-in inertial measurement unit, and a low-latency wireless interactive controller; A heterogeneous multi-source data synchronization sensing and benchmark anchoring module is deployed on the control host to establish a polar coordinate mapping space and topology matrix, and to extract the basic key delay time. The topological target adaptive optimization module based on active learning is used to perform Bayesian posterior probability update and joint objective cost function solution, and output the spatial coordinates of the projected target in the next frame. A cognitive load-driven nonlinear brightness compensation module is used to extract the real-time response delay to calculate the cognitive load index and to control the light source brightness gradient step size of the rendering engine in a closed loop. The cross-modal dynamics true / false decoupling and data cleaning module is used to fuse pose fluctuation variance and key timing derivative, and to perform multimodal confidence scoring and physical isolation of dirty data.

Images