Non-contact parkinson's disease preventive detection system and method based on retinal image eye movement characteristics

Abstract

The application relates to the technical field of non-contact physiological parameter detection, and discloses a non-contact Parkinson's disease preventive detection system and method based on retinal image eyeball movement characteristics, which extracts the space-time characteristics of eye images, establishes an individualized eye movement baseline, calculates the standardized deviation of the measured characteristics relative to the baseline, and realizes disease risk grading evaluation by relying on time sequence consistency joint inference; and solves the problem that the existing non-contact eye movement detection is distorted due to the lack of individualized baseline.

Description

Technical Field

[0001] This invention relates to the field of non-contact physiological parameter detection technology, specifically to a non-contact Parkinson's disease prevention detection system and method based on retinal imaging eye movement characteristics. Background Technology

[0002] Eye-tracking technology, by capturing the trajectory of eye movements and dynamic changes in pupil size, has significant application value in the fields of physiological state monitoring and neurological function assessment. Currently, eye-tracking feature detection technologies are mainly divided into two categories: contact-based and non-contact-based.

[0003] Contact-based solutions rely on specialized hardware such as head-mounted eye trackers and corneal electrodes, which have drawbacks such as poor wearing comfort, high equipment costs, and difficulty in widespread adoption on mobile devices.

[0004] Non-contact solutions, which rely on cameras to capture images of the eyes, offer advantages such as being painless, non-invasive, and easy to operate. However, existing technologies still have the following shortcomings: First, the multi-task processing architecture is fragmented and lacks real-time performance. Existing systems process eye semantic segmentation, image quality assessment, and pupil localization as independent modules in a sequential manner. Each module repeatedly extracts features, resulting in high computational redundancy and large inference latency, which makes it difficult to meet the real-time processing requirements of mobile terminals.

[0005] Second, the lack of an image quality adaptive mechanism leads to low-quality frames contaminating the features. Existing solutions do not introduce frame-level quality assessment during the decoding stage, resulting in low-quality frames such as blinks, occlusions, and blurs being included in eye-tracking feature calculations, introducing artifacts and jumps, and reducing feature reliability.

[0006] Third, the pupil localization accuracy and robustness are insufficient. Existing methods are prone to localization drift under complex conditions such as strong ambient light interference and overlapping corneal reflections, and they do not effectively utilize the pupil-corneal reflection vector to eliminate head micro-motion interference.

[0007] Fourth, ignoring individual differences and using group-based standards leads to distorted assessments. Existing solutions often use the statistical mean of the group as the judgment benchmark, without establishing an individualized baseline for eye movement function, resulting in normal physiological variations being misjudged as abnormal.

[0008] Fifth, the lack of temporal constraints in single-test results leads to poor stability. Existing solutions do not incorporate temporal consistency constraints based on historical test records, making it easy for adjacent test results to exhibit non-physiological extreme jumps, thus reducing the reliability of long-term tracking data. Summary of the Invention

[0009] The present invention aims to provide a non-contact Parkinson's disease prevention detection system and method based on retinal imaging eye movement characteristics, in order to solve the problem of assessment distortion caused by the lack of individualized baseline in existing non-contact eye movement detection systems.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: a non-contact Parkinson's disease prevention detection system based on retinal imaging eye movement characteristics, comprising: The multi-task eye image perception module is used to receive the raw eye video sequence, extract high-dimensional spatiotemporal features through a shared encoder, and output the eye semantic mask sequence, frame-level quality vector sequence and pupil coarse localization heatmap in parallel. The eye-tracking feature parsing and quality constraint module is used to perform precise localization of pupil and corneal reflection based on the eye semantic mask sequence and pupil coarse localization heatmap, and to perform artifact removal and trajectory interpolation in combination with the frame-level quality vector sequence to calculate a multi-dimensional eye-tracking feature set; The individualized eye movement baseline construction and deviation measurement module is used to construct an individualized eye movement baseline based on the eye movement features collected by the standard calibration task during the first detection, and to map the measured eye movement features into a standardized deviation relative to the individualized eye movement baseline in subsequent detections; The temporal consistency joint inference and risk assessment module is used to perform risk classification based on the standardized deviation, historical detection records and feature quality adaptive weighting, and output risk probability, risk level and structured detection report.

[0011] The principles and advantages of this scheme are as follows: In practical applications, by using a shared encoder and a multi-task parallel decoding architecture, eye semantic segmentation, image quality assessment, and coarse pupil localization are integrated into a unified feature extraction process, significantly reducing computational redundancy and meeting the real-time processing requirements of mobile terminals; through frame-level quality vector-driven artifact removal and trajectory interpolation mechanisms, the contamination of eye-tracking feature calculations by low-quality frames is effectively suppressed, improving feature reliability; through the construction of individualized eye-tracking baselines, measured features are mapped to standardized deviations relative to the user's normal range, avoiding misjudgments caused by group statistical bias; through temporal consistency constraints and adaptive weighted joint inference of feature quality, occasional noise and extreme jumps in a single detection are filtered out, improving the stability and consistency of long-term tracking and evaluation.

[0012] Preferably, as an improvement, the multi-task eye image perception module includes: The shared feature encoding unit is configured to take the original eye video sequence as input, use depthwise separable convolution to replace the standard convolution operation, and extract the shared feature map by stacking multiple downsampling layers. The multi-task parallel decoding unit is configured to connect in parallel to the eye semantic segmentation head, the image quality assessment head, and the pupil coarse localization head based on the shared feature map, and output the eye semantic mask sequence, the frame-level quality vector sequence, and the pupil coarse localization heatmap, respectively. The multi-task collaborative training unit is configured to perform end-to-end collaborative optimization of the shared feature encoding unit and the multi-task parallel decoding unit through a weighted multi-task loss function.

[0013] The beneficial effects of this improvement are as follows: The use of a shared feature encoding unit, replacing standard convolution operations with depthwise separable convolution, decouples spatial convolution from channel convolution, significantly reducing the number of model parameters and floating-point operations, enabling the system to achieve real-time inference on mainstream mobile processors; the multi-task parallel decoding unit synchronously outputs semantic masks, quality vectors, and coarse localization heatmaps based on the shared feature map, avoiding redundant computations caused by independent encoding of each task, and achieving multi-dimensional information extraction in a single forward propagation; the multi-task collaborative training unit optimizes the weighted loss function end-to-end, enabling each decoding task to mutually promote each other at the gradient level, improving the overall performance of segmentation accuracy, quality assessment accuracy, and localization speed.

[0014] Preferably, as an improvement, the eye-tracking feature analysis and quality constraint module includes: The pupil and corneal reflection fine localization unit is configured to extract the pupil mask channel and the corneal reflection mask channel from the semantic mask sequence, filter the largest connected component through connected component analysis and calculate the geometric center to obtain the pupil center coordinates and the corneal reflection bright spot center coordinates, and then calculate the pupil-corneal reflection vector. The eye-track temporal reconstruction and artifact removal unit is configured to establish eye-track based on the pupil-corneal reflection vector calculated frame by frame, mark invalid frames according to the eyelid occlusion and pupil visibility in the frame-level quality vector, and perform smooth reconstruction of the continuous interval marked as invalid frames using cubic spline interpolation. The three-dimensional eye movement feature calculation unit is configured to map the pupil-corneal reflection vector to the visual coordinate system through the device calibration matrix, and calculate the gaze stability feature, saccade velocity feature and microsaccade frequency feature.

[0015] The beneficial effects of this improvement are as follows: The pupil and corneal reflection precision localization unit obtains high-precision pupil center and corneal reflection bright spot center through connected component analysis and geometric center calculation, and constructs a pupil-corneal reflection vector to effectively encode the pupil offset relative to a fixed corneal reflection point, eliminate the interference of minor head movements on gaze direction estimation, and improve localization robustness; The eye movement trajectory temporal reconstruction and artifact removal unit automatically marks invalid frames based on eyelid occlusion and pupil visibility in the frame-level quality vector, and uses cubic spline interpolation to smoothly reconstruct the missing intervals, ensuring the continuity and physical rationality of the eye movement trajectory in the temporal domain and avoiding feature jumps introduced by artifacts; The three-dimensional eye movement feature calculation unit maps two-dimensional image features to the viewpoint coordinate system through the device calibration matrix, and comprehensively extracts multi-dimensional features such as gaze stability, saccade velocity, and microsaccade frequency to comprehensively characterize the eye movement function state and provide high-information input for subsequent individualized assessment.

[0016] Preferably, as an improvement, the time-series consistency joint inference and risk assessment module includes: The feature quality adaptive weighting unit is configured to calculate feature-level reliability weights based on the acquisition process quality of each feature, and construct a weighted feature vector based on the reliability weights and the standardized deviation. The temporal consistency constraint unit is configured to incorporate user historical detection records. When the deviation of a certain dimension of the current detection changes abruptly compared with the most recent historical detection and the magnitude of the change exceeds the physiologically reasonable range, temporal smoothing correction is triggered. The classification and risk assessment unit is configured to use a weighted support vector machine as a risk classifier, map the weighted feature vectors to risk probabilities, and perform three-level risk quantification. The personalized feedback generation unit is configured to automatically generate a structured detection report based on the risk assessment level and specific deviation indicators.

[0017] The beneficial effects of this improvement are as follows: The feature quality adaptive weighting unit calculates reliability weights based on the image quality during each feature acquisition process, giving greater influence to high-quality features and suppressing low-quality features, making the risk classification results more robust and reducing the interference of single acquisition environment fluctuations on the evaluation conclusions; The temporal consistency constraint unit introduces users' historical detection records and uses exponential weighted averaging to smooth out jumps that exceed the physiologically reasonable range, assuming that the real changes in eye movement function are a gradual process, effectively filtering out occasional noise and temporary physiological fluctuations, and improving the credibility of long-term monitoring data; The classification and risk assessment unit uses a weighted support vector machine, combined with feature quality weights and class balance factors, to solve the class imbalance problem where healthy samples far outnumber abnormal samples, outputting more reliable risk probabilities and three-level quantitative grades; The personalized feedback generation unit automatically generates a structured detection report based on the evaluation grade and specific deviation indicators, presenting the eye movement function status in a visual way, making it easy for users to intuitively understand the trend of their own eye movement feature changes.

[0018] Non-contact Parkinson's disease prevention detection methods based on retinal imaging ocular motion characteristics include: S1. Multimodal eye image acquisition: The system acquires raw eye video sequences at a fixed frame rate using a visible light camera, and optionally triggers a near-infrared light source to acquire auxiliary frame sequences. The system automatically selects single-modal or dual-modal working mode according to the hardware configuration. S2. Shared Feature Encoding and Multi-Task Parallel Decoding: The original eye video sequence is input into a shared encoder, and a shared feature map is extracted by replacing the standard convolution operation with a depthwise separable convolution. In parallel, a pixel-wise semantic mask is output through an eye semantic segmentation head, a frame-level quality vector is output through an image quality assessment head, and a pupil coarse localization heatmap is output through a pupil coarse localization head. The shared encoder and the three decoding heads are then optimized end-to-end through a weighted multi-task loss function. S3. Precise localization of pupil and corneal reflection and reconstruction of eye movement trajectory: Extract the pupil mask channel and corneal reflection mask channel from the semantic mask, calculate the geometric center through connected component analysis to obtain the coordinates of the pupil center and the center of the corneal reflection spot, and then calculate the pupil-corneal reflection vector; establish the eye movement trajectory based on the pupil-corneal reflection vector frame by frame, mark invalid frames according to the eyelid occlusion degree and pupil visibility in the frame-level quality vector, and perform smooth reconstruction of the continuous interval of invalid frames using cubic spline interpolation; S4. 3D eye movement feature calculation: The pupil-corneal reflection vector is mapped to the visual coordinate system through the device calibration matrix, and the gaze stability feature, saccade velocity feature and microsaccade frequency feature are calculated; wherein, the saccade velocity feature includes detecting saccade events and fitting the saccade master sequence relationship, and the microsaccade frequency feature is obtained by detecting energy peaks through continuous wavelet transform in the 1 to 3 Hz frequency band. S5. Personalized Eye-Tracking Baseline Construction: When users use the service for the first time, they are guided to complete a standardized four-degree-of-freedom eye-tracking calibration task, collect multi-dimensional eye-tracking feature data, and construct an individual baseline vector; physiological constraints are introduced to optimize the solution, including the main sequence index falling within the range of [0.3, 0.5], the micro-saccade frequency falling within the range of [0.5, 2.0] Hz, and prior adjustments to the fixation stability parameters based on the user's age; S6. Deviation Measurement and Risk Assessment: In subsequent periodic tests, the measured eye movement features are mapped to a standardized deviation relative to the individual baseline. A weighted feature vector is constructed by combining feature quality adaptive weights and temporal consistency constraints. A weighted support vector machine is used for risk classification and three-level risk quantification is performed. A structured test report is automatically generated based on the risk level and specific deviation indicators.

[0019] The beneficial effects of this improvement are as follows: Through multimodal eye image acquisition, the system can automatically switch between single-modal and dual-modal working modes according to hardware configuration; near-infrared auxiliary frames enhance pupil contrast under low light and strong ambient light interference conditions, improving adaptability to complex scenes; shared feature encoding and multi-task parallel decoding, through depthwise separable convolution and weighted collaborative training, significantly compress the model size while ensuring accuracy, enabling real-time inference at the edge; precise localization of pupil and corneal reflexes, combined with PCCR vectors and cubic spline interpolation, accurately locates and repairs missing trajectories, ensuring the integrity and accuracy of the eye movement feature set; personalized eye movement baseline construction introduces physiological constraint optimization, constraining parameters such as the main sequence index and microsaccade frequency to within ranges conforming to human physiological laws, and making prior adjustments based on age, so that the baseline estimation has both goodness of fit and physiological rationality, avoiding overfitting distortion; deviation measurement and risk assessment, combined with adaptive weighting of feature quality and temporal consistency constraints, construct high-quality, low-noise weighted feature vectors, and achieve stable and reliable risk classification through weighted support vector machines, and the final generated structured detection report can provide quantitative reference for subsequent clinical evaluation. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the system structure according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of a multi-task eye image perception module.

[0022] Figure 3 This is a schematic diagram of the eye-tracking feature analysis and quality constraint module.

[0023] Figure 4 This is a schematic diagram of the structure of the time-series consistency joint inference and risk assessment module. Detailed Implementation

[0024] The following detailed description illustrates the specific implementation methods: The basic implementation examples are as follows: Figure 1 As shown: Non-contact Parkinson's disease prevention detection systems based on retinal imaging and eye movement characteristics include: The multi-task eye image perception module is used to receive raw eye video sequences, extract high-dimensional spatiotemporal features through a shared encoder, and output eye semantic mask sequences, frame-level quality vector sequences, and pupil coarse localization heatmaps in parallel.

[0025] As attached Figure 2 As shown, the multi-task eye image perception module includes a shared feature encoding unit, a multi-task parallel decoding unit, and a multi-task collaborative training unit.

[0026] Shared feature coding units, configured to use the original eye video sequence As input, depthwise separable convolution (DSC) is used instead of standard convolution operations, and shared feature maps are extracted by stacking multiple downsampling layers. Where T is the total number of video frames, H×W is the spatial resolution of a single frame, and C=128 is the number of feature channels.

[0027] The shared feature encoding unit uses depthwise separable convolution, which decouples the convolution operation in the spatial dimension from the pointwise convolution in the channel dimension, significantly reducing the number of model parameters and floating-point operations, thus meeting the real-time inference requirements of mobile terminals.

[0028] The encoder E, which shares feature coding units, consists of multiple stacked downsampling layers. Each layer contains depthwise separable convolutions, batch normalization, and ReLU activation functions, achieving spatial resolution downsampling through stride convolutions. This downsampling rate is optimized and balanced to retain sufficient spatial details to support accurate pupil localization while reducing the computational overhead of subsequent decoding heads, ensuring the system achieves real-time processing speeds on mainstream mobile processors.

[0029] In addition, the system can optionally receive near-infrared auxiliary frames. Specifically, a near-infrared auxiliary acquisition unit is configured to optionally receive near-infrared auxiliary frame sequences synchronously triggered by a near-infrared light source with a wavelength of 850 nm or 940 nm. Together with the original eye video sequence, it constitutes a multimodal input, and the system automatically selects a single-modal or dual-modal working mode according to the hardware configuration.

[0030] Near-infrared auxiliary frames are mainly used to enhance the contrast between the pupil and iris under low light or strong ambient light interference conditions, thereby improving the robustness of subsequent segmentation and localization.

[0031] The multi-task parallel decoding unit is configured to access three lightweight decoding heads in parallel based on the shared feature map F, including: Eye semantic segmentation head An upsampling strategy combining deconvolution and bilinear interpolation is used to restore the shared feature map F to its original input resolution.

[0032] Specifically, First, the 128-dimensional feature channels are mapped to a 4-dimensional output using a 1×1 convolution. Then, 8x bilinear upsampling is performed to obtain a pixel-wise semantic mask M, the mathematical expression of which is: .

[0033] The mask has four channels corresponding to the background, pupil region, iris region, and corneal reflective spot categories, respectively. The mask output is normalized pixel-by-pixel using the Softmax function to obtain the category probability distribution, and then converted into discrete category labels using the argmax operation. The accurate segmentation of the pupil region and corneal reflective spot provides the foundation for subsequent eye-tracking feature analysis.

[0034] Image quality assessment head Global Average Pooling (GAP) is used to compress the spatial dimension of the shared feature map F into channel-level statistics, which are then mapped to a frame-level quality vector Q via two fully connected layers. The mathematical expression for Q is: .

[0035] The four components of this vector are image sharpness indices. Pupil visibility index Eyelid occlusion index With light uniformity index .in, The degree of image blurring is assessed by analyzing the proportion of high-frequency components. Calculated based on the proportion and contrast of the pupil region in the semantic mask; The degree of occlusion is quantified by detecting the overlap ratio between the eyelid edge and the pupil area; Evaluate the standard deviation of illumination distribution within the pupil region. This quality vector is used for subsequent feature weighting and artifact removal, ensuring that eye-tracking features are computed only on high-quality frames.

[0036] Pupil coarse positioning head Two layers of 3×3 convolution are used to extract local features, and then a single-channel heatmap is output through 1×1 convolution. Its mathematical expression is .

[0037] This heatmap represents the probability distribution of the pupil center in the low-resolution space of the shared feature map, and the peak positions of the heatmap correspond to the coarse localization coordinates of the pupil center. This coarse localization result is used to initialize the fine localization algorithm in Module 2, reducing the number of iterations and improving the system response speed.

[0038] The multi-task collaborative training unit achieves collaborative optimization of the three decoders through a weighted multi-task loss function. Total loss function. Defined as the weighted sum of the losses from the three tasks, its mathematical expression is:

[0039] in, We employ a cross-entropy loss function with class weights. Using the mean squared error loss function, The weighted mean square error loss function is adopted. , , The hyperparameter weights for each loss are determined through optimization using the validation set.

[0040] The cross-entropy loss function with class weights is used to address the class imbalance problem in eye semantic segmentation. Since the pupil region and corneal reflection bright spots account for a small proportion of the overall image, while background pixels dominate, the system assigns higher loss weights to the pupil and corneal reflection categories to ensure segmentation accuracy for small target regions. The class weights are calculated based on the reciprocal of the pixel frequency of each category in the training set, allowing rarer categories to receive a larger gradient contribution in the loss function.

[0041] The Mean Squared Error (MSE) loss function is used to monitor manually annotated image quality ground truth values ​​as a supervisory signal. These quality ground truth values ​​are generated through an automated annotation process. The sharpness metric is calculated from the Laplacian operator response. Pupil visibility is determined by the detection rate of the pupil detection algorithm. The degree of eyelid occlusion is automatically labeled by the eyelid segmentation model. Illumination uniformity is calculated from the standard deviation of gray levels within the pupil area. This automated annotation strategy avoids the need for large-scale manual quality scoring while maintaining annotation consistency and scalability.

[0042] The weighted mean squared error loss function is used to supervise the model's output of the ground truth heatmap of the pupil center generated by the Gaussian kernel. Specifically, a two-dimensional Gaussian distribution heatmap is generated centered on the manually labeled or automatically detected pupil center coordinates as the ground truth label, and the standard deviation of the Gaussian kernel is adaptively adjusted according to the pupil diameter. The weighting strategy gives higher weight to the central region of the heatmap, encouraging the model to output a sharper peak distribution and improving localization accuracy. Through the joint optimization of the above three losses, the system achieves a balance between segmentation accuracy, quality assessment accuracy, and localization speed.

[0043] The eye-tracking feature parsing and quality constraint module is used to perform precise localization of pupil and corneal reflection based on the eye semantic mask sequence and pupil coarse localization heatmap, and to perform artifact removal and trajectory interpolation in combination with the frame-level quality vector sequence to calculate a multi-dimensional eye-tracking feature set.

[0044] As attached Figure 3 As shown, the eye movement feature analysis and quality constraint module includes a pupil and corneal reflection fine localization unit, an eye movement trajectory temporal reconstruction and artifact removal unit, and a three-dimensional eye movement feature calculation unit.

[0045] Pupil and corneal reflex precision localization unit, configured to, for each frame t in the video sequence, from the semantic mask Extracting the pupil mask channel By filtering the largest connected component through connected component analysis and calculating the geometric center, the coordinates of the pupil center are obtained. Its expression is;

[0046]

[0047] The summation operation iterates through the coordinates (u, v) of all pixels in the pupil mask that have a value of 1. This is based on the pixel area of ​​the pupil mask. Physical size of a single pixel obtained from calibration Calculate the equivalent diameter of the pupil The output is then sent to subsequent modules.

[0048] Extracting corneal reflection mask channels Calculate the coordinates of the center of the corneal reflective bright spot. ; Calculate the pupil-corneal reflection vector based on the two center coordinates mentioned above: ; The pupil-corneal reflectance vector (PCCR vector) is a two-dimensional vector that encodes the direction and distance of the pupil center's offset relative to the corneal reflectance point. It is a standard feature representation in the field of eye tracking. Since the corneal reflectance point is fixed on the corneal surface, while the pupil center shifts relative to eye movement, the PCCR vector can effectively eliminate the interference of minute head movements on eye movement measurements, providing robust gaze direction estimation.

[0049] The temporal reconstruction and artifact removal unit for eye movement trajectory is configured to establish eye movement trajectories based on pupil-corneal reflection vectors calculated frame by frame. For each frame, a judgment is performed based on the quality vector: If eyelid occlusion Exceeding the preset threshold or pupil visibility Below the preset threshold If so, the frame is marked as an invalid frame; in this embodiment, the following is set: It is 0.3; It is 0.7.

[0050] For consecutive intervals marked as invalid frames, cubic spline interpolation is used, with the PCCR vectors of the valid frames at both ends of the interval and their first-order (or second-order) temporal differences as boundary conditions, to smoothly reconstruct the missing values ​​within the interval and simultaneously generate a valid frame mask. .

[0051] The three-dimensional eye-tracking feature calculation unit is configured to map the pupil-corneal reflection vector on the two-dimensional image plane to the viewpoint coordinate system through the device calibration matrix to obtain the viewpoint coordinate sequence. And calculate multidimensional features for different eye-tracking subtasks, including: Fixation stability analysis calculates the two-dimensional covariance matrix of the effective frame viewpoint coordinate sequence within a continuous fixation task segment. Extracting feature values and , Calculate the discrete standard deviation of the fixation point. and 1 Area of ​​an ellipse with equal probability density ; and where is the eigenvalue of the covariance matrix.

[0052] Eye saccade speed analysis calculates the visual angular velocity between adjacent valid frames. ; This is the inter-frame time interval. This refers to the video sampling rate.

[0053] Based on velocity threshold Detect saccade events and extract saccade parameter sets. In this embodiment, a speed threshold is set. =30 degrees / second. When the angular velocity exceeds this threshold for three or more consecutive frames, it is considered a saccade event. The latency period is the time from when the target appears until eye saccharification begins. The duration is the total duration of the eye saccade. Peak velocity, which is the maximum angular velocity during the eye saccade. The saccade amplitude is the visual distance between the start and end points of the saccade.

[0054] And fit the main sequence relationship of eye saccades Where k is the proportionality coefficient and α is the power exponent. This power law relationship is a fundamental law in eye movement physiology, and deviations from the parameters (k, α) can reflect the functional state of the brainstem eye movement control pathway.

[0055] Microsaccade frequency analysis involves performing linear detrending processing on the visual signal after excluding saccade events within the fixation segment, detecting energy peaks in the 1-3 Hz frequency band using continuous wavelet transform, and counting the number of microsaccades per unit time. and amplitude distribution skewness .

[0056] The individualized eye-tracking baseline construction and deviation measurement module is used to construct an individualized eye-tracking baseline based on the eye-tracking features collected during the first detection using the standard calibration task, and to map the measured eye-tracking features into a standardized deviation relative to the individualized eye-tracking baseline in subsequent detections.

[0057] The individualized eye-tracking baseline construction and deviation measurement module includes a standard calibration task acquisition unit, an individual baseline parameter solving unit, and a deviation calculation unit.

[0058] The standard calibration task acquisition unit is configured to guide users to complete a standardized four-degree-of-freedom eye-tracking calibration task when they use the system for the first time. This task includes central fixation, four-corner saccades, resting tracking, and light reflection tasks, and acquires multidimensional eye-tracking feature data. Specifically, the central fixation task requires the user to fixate on the central crosshair on the screen for 30 seconds, during which the system collects eye movement data and calculates the baseline of fixation stability. The four-corner saccade task requires the user to fixate on target points in the four corners of the screen in sequence, and collects the baseline of the main saccade sequence parameters through multiple directional saccades. The resting tracking task requires the user to fixate on a light spot that moves slowly along a smooth trajectory, and collects the baseline of natural microsaccade activity during smooth tracking. The light reflection task uses the brightness switching stimulus of the screen from white to black to collect the baseline of the dynamic response and variation characteristics of the pupil diameter. The total time of the above calibration tasks is approximately 2 to 3 minutes, which is completed in one go when the user registers for the first time. The calibration data is securely stored in the system database after authorization by the user.

[0059] The individual baseline parameter solving unit, based on the multidimensional feature data collected during the calibration task, constructs an individual baseline vector B, the mathematical expression of which is:

[0060] in, Mean of fixation stability For the standard deviation of fixation stability, The main sequence ratio coefficient for saccades, Main sequence index, For the baseline of microsaccade frequency, This represents the baseline for the pupil diameter variation coefficient. This six-dimensional vector comprehensively characterizes the individual normal range of a user's eye movement function.

[0061] The baseline parameter solution process incorporates physiological constraint optimization to ensure that the estimation results conform to human physiological laws: Constraint 1 requires the main sequence index It falls within the range of [0.3, 0.5], which is consistent with the main sequence index of healthy individuals reported in the eye movement physiology literature; Constraint 2 requires microsaccade frequency It falls within the [0.5, 2.0] Hz range, which corresponds to the normal microsaccade activity level of a healthy adult. Constraint 3 introduces an age correction factor to adjust the fixation stability parameter a priori based on the user's age. Specifically, the fixation stability baseline for older users is appropriately relaxed to reflect normal age-related decline.

[0062] The baseline parameters are solved using a constrained least squares optimization algorithm to balance goodness of fit and physiological rationality, avoiding baseline distortion caused by overfitting. The objective function of the optimization problem is a weighted sum of the likelihood term of the calibration data and the constraint violation penalty term, which is solved using the Lagrange multiplier method or the projected gradient method.

[0063] The deviation calculation unit is configured to map the measured eye movement features to a standardized deviation relative to the individual baseline in subsequent periodic tests, and calculate the gaze stability deviation. Eye saccade main sequence deviation Microsaccade frequency deviation and pupil diameter variation deviation , forming the feature deviation vector .

[0064] The temporal consistency joint inference and risk assessment module is used to perform risk classification based on the standardized deviation, historical detection records and feature quality adaptive weighting, and output risk probability, risk level and structured detection report.

[0065] As attached Figure 4 As shown, the temporal consistency joint inference and risk assessment module includes a feature quality adaptive weighting unit, a temporal consistency constraint unit, a classification and risk assessment unit, and a personalized feedback generation unit.

[0066] The feature quality adaptive weighting unit is configured to calculate feature-level reliability weights based on the acquisition process quality of each feature, including gaze stability weights. Eye sac weight and micro-saccade weight .

[0067] Specifically, attention stability weights Defined as the mean pupil visibility in the effective fixation frame, its mathematical expression is:

[0068] in, A set of time indices for the gaze task segment. This represents the cardinality of the set. The larger this weight, the more ideal the measurement conditions for the gaze stability feature, and the more reliable the results.

[0069] eye sac weight Defined as the proportion of validly detected saccade events out of all detected saccade events, i.e. ,in The number of saccade events required to meet the quality threshold, This represents the total number of twitching events detected.

[0070] Micro-eye sac weight Defined using an exponential decay function:

[0071] in, Let r(t) be the coefficient of variation of the residual signal r(t) during the gaze segment. This is the attenuation coefficient. This weighting design ensures that micro-saccade features with higher noise levels receive lower weights, thus suppressing the influence of low-quality data on the inference results.

[0072] Based on the above three quality weights and pupil deviation, a weighted feature vector is constructed. Its mathematical expression is:

[0073] This weighted feature vector scales the deviation of each dimension according to its measurement quality, giving higher quality features greater influence and appropriately suppressing low-quality features, thereby improving the robustness of risk assessment.

[0074] The temporal consistency constraint unit is configured to incorporate user historical detection records H and establish a temporal consistency penalty term to prevent misjudgments caused by outliers in a single detection. Specifically, if the deviation of a certain dimension of the current detection changes drastically compared to the most recent historical detection, and the magnitude of the change exceeds the physiologically reasonable range, a temporal smoothing correction is triggered. The corrected deviation is calculated using an exponentially weighted average. ,in This represents the time-series confidence coefficient.

[0075] This smoothing strategy assumes that the actual changes in eye movement function are gradual processes, without extreme jumps in the short term, thus effectively filtering out noise and occasional artifacts in a single detection. Temporal consistency constraints are only enabled when a user has a historical record; for first-time users, the current observation is used directly.

[0076] The classification and risk assessment unit is configured to use a weighted support vector machine as the risk classifier, and the optimization objective function is: ; Where w is the normal vector of the classification hyperplane, b is the bias term, C is the regularization parameter, and N is the number of training samples. Let i be the slack variable for the i-th sample. is the comprehensive weight of the i-th training sample.

[0077] Overall weight It consists of two parts: first, the calculated feature quality weights, which reflect the measurement reliability of the samples; and second, the class balance factor, which adjusts the contributions of positive and negative samples through oversampling or undersampling strategies to solve the class imbalance problem in the training data where there are far more healthy samples than diseased samples.

[0078] The classifier outputs the risk probability of the sample. This indicates the probability that the current user has Parkinson's disease or is in the prodromal stage. Based on the risk probability, the system performs three levels of risk quantification: Low risk level corresponds to The determination is based on the absolute values ​​of all deviation components. And the timing consistency is good; Medium risk level corresponds to The judgment criteria are that one or two feature deviation components satisfy the condition. ; High-risk level corresponds to The criteria for judgment are that multiple features deviate significantly or have components. .

[0079] The personalized feedback generation unit is configured to automatically generate structured test reports based on risk assessment levels and specific deviation indicators, providing differentiated health advice and medical guidance for users with different risk levels.

[0080] Specifically, for high-risk users, the system provides a recommendation to "visit the neurology department within 2 weeks for UPDRS scale assessment and further examination with dopamine transporter imaging (DAT-SPECT)," and details the specific abnormal indicators and their deviations in the report (e.g., "Your saccade master sequence slope deviates from the individual baseline by 37%"). This helps users understand the source of risk and provides reference information for clinicians.

[0081] For users at medium risk, the system provides intervention suggestions such as "increase aerobic exercise (no less than 150 minutes of moderate-intensity exercise per week) and retest eye movement indicators after 3 months" to reduce the probability of risk progression through lifestyle adjustments.

[0082] For low-risk users, the system provides maintenance advice such as "Eye movement indicators are within the normal fluctuation range of the individual baseline; it is recommended to maintain a healthy lifestyle and undergo regular screening annually," avoiding unnecessary anxiety and over-medicalization. The report is presented in a visual format combining text and radar charts. The radar charts show the relative position of each characteristic's deviation from the normal range, allowing users to intuitively understand their own eye movement function status.

[0083] Non-contact Parkinson's disease prevention detection methods based on retinal imaging ocular motion characteristics include: S1: Multimodal eye image acquisition Raw eye video sequences are captured at a fixed frame rate using a high-resolution visible light camera deployed on a mobile device or workstation. Optionally, a near-infrared light source with a wavelength of 850 nm or 940 nm is synchronously triggered to acquire a near-infrared auxiliary frame sequence. The system automatically selects either single-mode or dual-mode operating mode based on the hardware configuration. S2: Shared Feature Encoding The original eye video sequence is input into the shared encoder E, and a depthwise separable convolution is used to replace the standard convolution operation. A shared feature map F is extracted by stacking multiple downsampling layers. S3: Multi-task parallel decoding Three decoding tasks are performed in parallel based on the shared feature map F: Using eye semantic segmentation head Output pixel-by-pixel semantic mask Discrete category labels for background, pupil, iris, and corneal reflective bright spots are obtained after Softmax normalization and argmax operation. Image quality assessment head Output frame-level quality vector Including image sharpness indicators Pupil visibility index Eyelid occlusion index With light uniformity index ; Pupil coarse positioning head Output single-channel heatmap Initialize the coarse positioning coordinates of the pupil center with the peak position; S4: Multi-task collaborative training Through weighted multi-task loss function End-to-end collaborative optimization is performed on the shared encoder and the three decoding heads, wherein... We employ cross-entropy loss with class weights. Mean squared error loss is used. Weighted mean square error loss is used; S5: Precise localization of pupil and corneal reflexes For each frame t in the video sequence, from the semantic mask Extracting the pupil mask channel With corneal reflection mask channel By filtering the largest connected component through connected component analysis and calculating the geometric center, the coordinates of the pupil center were obtained. Coordinates of the center of the corneal reflective spot ; Calculate the pupil-corneal reflection vector based on the two center coordinates. Simultaneously calculate the equivalent pupil diameter based on the pixel area of ​​the pupil mask. ; S6: Temporal reconstruction of eye-tracking trajectory and artifact removal Eye movement trajectory established based on pupil-corneal reflex vector calculated frame by frame. For each frame, a judgment is made based on the quality vector Q: if the eyelid occlusion degree... Exceeding the preset threshold or pupil visibility Below the preset threshold If a frame is marked as invalid, then the frame is marked as invalid. For a continuous interval marked as invalid, cubic spline interpolation is used, with the PCCR vectors of the valid frames at both ends of the interval and their temporal differences as boundary conditions for smooth reconstruction, and a valid frame mask is generated simultaneously. ; S7: Calculation of 3D Eye-Motion Features The pupil-corneal reflection vector on the two-dimensional image plane is mapped to the viewing angle coordinate system using the device calibration matrix to obtain the viewing angle coordinate sequence. And calculate multidimensional features for different eye-tracking subtasks: Fixation stability analysis: Calculate the two-dimensional covariance matrix of the effective frame viewpoint coordinate sequence within a continuous fixation task segment. Extracting feature values Calculate the discrete standard deviation of the fixation point. and the area of ​​the ellipse enclosed by the point of fixation ; Eye saccade velocity analysis: Calculate the visual angular velocity between adjacent valid frames. Based on speed threshold Detect saccade events and extract saccade parameter sets. And fit the main sequence relationship of eye saccades. ; Microsaccade frequency analysis: Linear detrending processing is performed on the visual signal after excluding saccade events within the fixation segment. Energy peaks are detected in the 1 to 3 Hz frequency band using continuous wavelet transform, and the number of microsaccades per unit time is counted. and amplitude distribution skewness ; S8: Standard Calibration Task Acquisition When users use the system for the first time, they are guided to complete a standardized four-degree-of-freedom eye-tracking calibration task, including a central fixation task, a four-corner saccade task, a resting tracking task, and a light reflection task, to collect multidimensional eye-tracking feature data. S9: Solving Individual Baseline Parameters Based on the multidimensional feature data collected during the calibration task, an individual baseline vector is constructed. ; Introducing physiological constraints for optimization, including: main sequence index The frequency of micro-saccades falls within the range of [0.3, 0.5]. It falls within the [0.5, 2.0] Hz range, and the gaze stability parameter is adjusted a priori based on the user's age; S10: Deviation Calculation In subsequent periodic tests, the measured eye movement features are mapped to a standardized deviation relative to the individual baseline, and the gaze stability deviation is calculated. Eye saccade main sequence deviation Microsaccade frequency deviation and pupil diameter variation deviation , forming the feature deviation vector ; S11: Adaptive weighting of feature quality The feature-level reliability weights, including gaze stability weights, are calculated based on the quality of the acquisition process for each feature. Eye sac weight and micro-saccade weight Based on the above quality weights, a weighted feature vector is constructed. ; S12: Timing Consistency Constraints Introducing user historical detection records H, when the deviation of a certain dimension of the current detection shows a jump compared to the most recent historical detection, and the jump exceeds the physiologically reasonable range, a time-series smoothing correction is triggered. The corrected deviation is calculated using an exponentially weighted average. ,in For time-series confidence coefficient; S13: Classification and Risk Assessment We use a weighted support vector machine as the risk classifier and map the decision values ​​to risk probabilities using the sigmoid function. And implement a three-level risk quantification: low risk corresponds to Medium-risk response High-risk response ; S14: Personalized Feedback Generation Based on the risk assessment level and specific deviation indicators, a structured detection report is automatically generated to provide differentiated health advice and medical guidance for users with different risk levels.

[0084] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A non-contact Parkinson's disease prevention detection system based on retinal imaging eye movement characteristics, characterized in that, include: The multi-task eye image perception module is used to receive the raw eye video sequence, extract high-dimensional spatiotemporal features through a shared encoder, and output the eye semantic mask sequence, frame-level quality vector sequence and pupil coarse localization heatmap in parallel. The eye-tracking feature parsing and quality constraint module is used to perform precise localization of pupil and corneal reflection based on the eye semantic mask sequence and pupil coarse localization heatmap, and to perform artifact removal and trajectory interpolation in combination with the frame-level quality vector sequence to calculate a multi-dimensional eye-tracking feature set; The individualized eye movement baseline construction and deviation measurement module is used to construct an individualized eye movement baseline based on the eye movement features collected by the standard calibration task during the first detection, and to map the measured eye movement features into a standardized deviation relative to the individualized eye movement baseline in subsequent detections; The temporal consistency joint inference and risk assessment module is used to perform risk classification based on the standardized deviation, historical detection records and feature quality adaptive weighting, and output risk probability, risk level and structured detection report.

2. The non-contact Parkinson's disease prevention detection system based on retinal imaging eye movement characteristics according to claim 1, characterized in that, The multi-task eye image perception module includes: The shared feature encoding unit is configured to take the original eye video sequence as input, use depthwise separable convolution to replace the standard convolution operation, and extract the shared feature map by stacking multiple downsampling layers. The multi-task parallel decoding unit is configured to connect in parallel to the eye semantic segmentation head, the image quality assessment head, and the pupil coarse localization head based on the shared feature map, and output the eye semantic mask sequence, the frame-level quality vector sequence, and the pupil coarse localization heatmap, respectively. The multi-task collaborative training unit is configured to perform end-to-end collaborative optimization of the shared feature encoding unit and the multi-task parallel decoding unit through a weighted multi-task loss function.

3. The non-contact Parkinson's disease prevention detection system based on retinal imaging eye movement characteristics according to claim 2, characterized in that, The eye-tracking feature analysis and quality constraint module includes: The pupil and corneal reflection fine localization unit is configured to extract the pupil mask channel and the corneal reflection mask channel from the semantic mask sequence, filter the largest connected component through connected component analysis and calculate the geometric center to obtain the pupil center coordinates and the corneal reflection bright spot center coordinates, and then calculate the pupil-corneal reflection vector. The eye-track temporal reconstruction and artifact removal unit is configured to establish eye-track based on the pupil-corneal reflection vector calculated frame by frame, mark invalid frames according to the eyelid occlusion and pupil visibility in the frame-level quality vector, and perform smooth reconstruction of the continuous interval marked as invalid frames using cubic spline interpolation. The three-dimensional eye movement feature calculation unit is configured to map the pupil-corneal reflection vector to the visual coordinate system through the device calibration matrix, and calculate the gaze stability feature, saccade velocity feature and microsaccade frequency feature.

4. The non-contact Parkinson's disease prevention detection system based on retinal imaging eye movement characteristics according to claim 3, characterized in that, The time-series consistency joint inference and risk assessment module includes: The feature quality adaptive weighting unit is configured to calculate feature-level reliability weights based on the acquisition process quality of each feature, and construct a weighted feature vector based on the reliability weights and the standardized deviation. The temporal consistency constraint unit is configured to incorporate user historical detection records. When the deviation of a certain dimension of the current detection changes abruptly compared with the most recent historical detection and the magnitude of the change exceeds the physiologically reasonable range, temporal smoothing correction is triggered. The classification and risk assessment unit is configured to use a weighted support vector machine as a risk classifier, map the weighted feature vectors to risk probabilities, and perform three-level risk quantification. The personalized feedback generation unit is configured to automatically generate a structured detection report based on the risk assessment level and specific deviation indicators.

5. A non-contact Parkinson's disease preventive detection method based on retinal image eye movement characteristics, used to implement the non-contact Parkinson's disease preventive detection system based on retinal image eye movement characteristics as described in any one of claims 1-4, characterized in that, include: S1. Multimodal eye image acquisition: The system acquires raw eye video sequences at a fixed frame rate using a visible light camera, and optionally triggers a near-infrared light source to acquire auxiliary frame sequences. The system automatically selects single-modal or dual-modal working mode according to the hardware configuration. S2. Shared Feature Encoding and Multi-Task Parallel Decoding: The original eye video sequence is input into a shared encoder, and a shared feature map is extracted by replacing the standard convolution operation with a depthwise separable convolution. In parallel, a pixel-wise semantic mask is output through an eye semantic segmentation head, a frame-level quality vector is output through an image quality assessment head, and a pupil coarse localization heatmap is output through a pupil coarse localization head. The shared encoder and the three decoding heads are then optimized end-to-end through a weighted multi-task loss function. S3. Precise localization of pupil and corneal reflection and reconstruction of eye movement trajectory: Extract the pupil mask channel and corneal reflection mask channel from the semantic mask, calculate the geometric center through connected component analysis to obtain the coordinates of the pupil center and the center of the corneal reflection spot, and then calculate the pupil-corneal reflection vector. Eye movement trajectories are established based on the pupil-corneal reflection vector frame by frame. Invalid frames are marked according to the eyelid occlusion and pupil visibility in the frame-level quality vector. Cubic spline interpolation is used to smoothly reconstruct the continuous intervals of invalid frames. S4. 3D eye movement feature calculation: The pupil-corneal reflection vector is mapped to the visual coordinate system through the device calibration matrix, and the gaze stability feature, saccade velocity feature and microsaccade frequency feature are calculated; wherein, the saccade velocity feature includes detecting saccade events and fitting the saccade master sequence relationship, and the microsaccade frequency feature is obtained by detecting energy peaks through continuous wavelet transform in the 1 to 3 Hz frequency band. S5. Personalized Eye-Tracking Baseline Construction: When users use the service for the first time, they are guided to complete a standardized four-degree-of-freedom eye-tracking calibration task, collect multi-dimensional eye-tracking feature data, and construct an individual baseline vector; physiological constraints are introduced to optimize the solution, including the main sequence index falling within the range of [0.3, 0.5], the micro-saccade frequency falling within the range of [0.5, 2.0] Hz, and prior adjustments to the fixation stability parameters based on the user's age; S6. Deviation Measurement and Risk Assessment: In subsequent periodic tests, the measured eye movement features are mapped to a standardized deviation relative to the individual baseline. A weighted feature vector is constructed by combining feature quality adaptive weights and temporal consistency constraints. A weighted support vector machine is used for risk classification and three-level risk quantification is performed. A structured test report is automatically generated based on the risk level and specific deviation indicators.

Images