A method, system and device for quantitatively evaluating visual function deterioration

Abstract

The application relates to the field of intelligent medical treatment, in particular to a visual function recession quantitative evaluation method, system and device. Early detection and accurate evaluation of visual function recession are key links for delaying disease progression and reducing the risk of irreversible visual function damage, however, existing clinical evaluation faces many limitations. In model training, the application corrects dimension sensitivity shrinkage by a psychological cognitive state, explicitly injects disc anatomical partition into the fundus feature extraction process, and introduces a physiological consistency ordering constraint between the visual function recession grade and the retinal age, so that the model accurately evaluates early visual function recession. The application provides a more accurate quantitative evaluation method, which has important clinical value and industrial value.

Description

Technical Field

[0001] This application relates to the field of intelligent healthcare, specifically to a method, system, and device for quantitatively assessing visual function decline. Background Technology

[0002] With the accelerating aging of the population and the increasing burden of chronic diseases, the incidence of eye diseases closely related to visual function decline, such as glaucoma, age-related macular degeneration, and diabetic retinopathy, continues to rise, posing a severe challenge to the healthcare system. Early detection and accurate assessment of visual function decline are crucial for slowing disease progression and reducing the risk of irreversible vision loss. However, existing clinical assessment methods face many limitations, necessitating more precise, robust, and interpretable quantitative assessment methods. Summary of the Invention

[0003] To address the above problems, this invention provides a method for quantitatively assessing visual function decline, specifically including: Acquire fundus images, visual function data, and psychological cognitive data of the tested samples; The fundus image is divided into regions and the gray-level features of each region are calculated; the global retinal structure features are obtained by feature encoding using the fundus image and the gray-level features of each region. Psychological confidence coefficients are generated from psychological cognitive data, and visual function data is corrected using these psychological confidence coefficients to obtain corrected visual function data. The corrected visual function data are used to calculate the comprehensive visual function impairment index vector through principal component transformation. Fusion features are obtained by modulating global retinal structural features using a comprehensive visual impairment index vector; The assessment results are obtained by using fusion features to evaluate the degree of visual function decline and / or retinal age.

[0004] The purpose of this invention is to provide a computer program product that includes a computer program or instructions, which are executed by a processor to implement the above-described method for quantitative assessment of visual function decline.

[0005] The purpose of this invention is to provide a computer device comprising a memory, a processor, and a computer program or instructions stored in the memory, wherein the computer program or instructions are executed by the processor to implement the above-described method for quantitative assessment of visual function decline.

[0006] The purpose of this invention is to provide a computer-readable storage medium having a computer program or instructions stored thereon, which are executed by a processor to implement the above-described method for quantitative assessment of visual function decline.

[0007] Advantages of this invention: 1. Existing technologies, when processing visual function test data, generally neglect the influence of psychological and cognitive states on the measurement results, or only employ simple sample elimination strategies. They fail to perform refined confidence adjustments for different visual function indicators based on dimensions such as depression, anxiety, and cognitive cooperation, leading to non-organic fluctuations being misjudged as true visual function decline. This invention uses a psychological and cognitive state-driven dimension-sensitive contraction correction to adjust the confidence of raw visual function test data. Unlike directly eliminating abnormal samples or uniformly scaling, this method generates psychological confidence coefficients based on depression, anxiety, and cognitive states, and performs personalized contraction based on the degree of psychological influence on different visual function dimensions, thus weakening the interference of non-organic fluctuations while preserving true organic damage information.

[0008] 2. Existing fundus image-based analysis methods mostly employ global encoding to extract features, failing to incorporate the anatomical patterns of the peripapillary nerve fiber layer as structural priors into the model. Since different anatomical regions exhibit differentiated damage patterns during visual function decline, global, indiscriminate encoding easily overlooks subtle texture changes in key local areas and lacks interpretability support for the areas of interest to the model. This invention explicitly injects anatomical partitions of the optic disc into the fundus feature extraction process, achieving partition-aware depth feature encoding. By generating spatial masks for six anatomical partitions—inferior, superior, nasal, temporal, foveal, and optic disc core regions—the model is guided to extract local depth features from each partition. These features are then weighted and aggregated by a partition attention module combined with untrained structural priors, ensuring that the areas of interest to the model align with clinical image interpretation logic and improving the interpretability of the assessment.

[0009] 3. Existing methods, when utilizing visual function testing data, typically use the test results of individual subjects in isolation, failing to leverage demographic information to uncover group decline patterns among similar populations. Single test results are susceptible to random factors and measurement noise, lacking group-level smoothing constraints, leading to insufficient assessment stability. This invention utilizes a nonparametric graph propagation enhancement mechanism based on demographic similarity to smooth individual visual function characteristics using the decline trends of similar populations. Using age and body mass index as anchors, neighborhood samples are searched in the training set. A comprehensive visual function impairment index of the neighborhood is aggregated using Gaussian decay similarity weighting and fused with individual characteristics, effectively reducing the impact of single test noise and random fluctuations on the assessment results.

[0010] 4. Existing multi-task learning methods related to visual function decline typically optimize decline level classification and age regression as independent tasks, without explicitly modeling the monotonic physiological relationship between the two. This can lead to contradictory results that do not conform to clinical common sense, such as classifying samples with low retinal physiological aging as having severe decline, thus weakening the clinical credibility of the assessment results. This invention introduces a physiological consistency ranking constraint between visual function decline level and retinal age in joint training. By constructing positive-order sample pairs and requiring that the predicted retinal age of samples with higher decline levels is no lower than that of samples with lower decline levels, physiological logic is directly encoded as the optimization objective. This avoids contradictory results such as the model outputting severe decline but with a younger retinal age, ensuring the inherent consistency of multi-task predictions. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the process for quantitative assessment of visual function decline provided in an embodiment of the present invention; Figure 2 The comparison results of psychological confidence coefficient correction and non-correction are provided for embodiments of the present invention; Figure 3 The comparison results before and after global retinal structural depth feature modulation are provided in the embodiments of the present invention; Figure 4 Demographic neighborhood search is provided for embodiments of the present invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0014] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as S101, S102, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0015] Figure 1 A schematic diagram of the method for quantitatively assessing visual function decline provided in this embodiment of the invention specifically includes: S1: Obtain fundus images, visual function data, and psychological cognitive data of the tested sample; In one embodiment, the fundus image of the tested sample can be represented as a fundus image vector after vectorization of the fundus image, the visual function data can be represented as the visual function vector data after vectorization of the collected visual function data, and the psychological cognitive data can be represented as the psychological cognitive data after vectorization of the psychological cognitive data.

[0016] Visual function data include any one or more of the following: visual acuity, contrast sensitivity, stereopsis, motion perception, average visual field defect, visual field pattern standard deviation, color vision, and binocular balance.

[0017] In one embodiment, the method further includes data preprocessing, which involves preprocessing the fundus image, visual function data, and psychological cognitive data to obtain preprocessed fundus image and preprocessed visual function data. The preprocessed data is then used to obtain psychological cognitive data. The preprocessed fundus image is partitioned and the grayscale features of each region are calculated. A psychological confidence coefficient is generated using the preprocessed psychological cognitive data. The preprocessed visual function data is then corrected using the psychological confidence coefficient to obtain corrected visual function data.

[0018] In one specific embodiment, visual function decline multi-source sample collection and task label construction: This invention first collects multi-source data from subjects in a unified manner, then establishes labels for visual function decline level and retinal age, and finally forms a multi-source sample set for model training, validation, and inference. The specific steps are as follows: 1>Collecting the first The original fundus images, original visual function test data, psychological cognitive scale data, and demographic data of each subject were used to form the first... A multi-source vision function sample ,in Indicates the sample index. , This indicates the total number of samples.

[0019] In one implementation, the fundus images, visual function test data, psychological cognitive scale data, and demographic data are derived from the image acquisition equipment, visual function test equipment, scale assessment process, and medical records, respectively. The original fundus images are denoted as... Original fundus image It refers to color fundus images acquired through fundus cameras or fundus imaging devices, which are used to provide structural information such as the optic disc, the nerve fiber layer around the optic disc, blood vessels, and the fovea of ​​the macula. The original visual function vector is denoted as Original visual function vector This refers to a vector composed of multiple visual function testing indicators, which may include visual acuity, contrast sensitivity, stereopsis, motion perception, average visual field defect, visual field pattern standard deviation, color vision, binocular balance, and other standard visual function testing indicators; the total number of visual function dimensions is denoted as . , Represents the original visual function vector The number of detection indicators can be taken as 8 to 15, depending on a method that is easy to implement. ; The data from the psychological cognition scale are recorded as follows: Psychological cognition scale data It refers to scale data used to assess the emotional state, attentional state and cognitive cooperation of the subjects, preferably including the total score of the depression scale, the total score of the anxiety scale and the total score of the mini cognitive scale; Demographic data are denoted as Demographic data This refers to basic information related to the patterns of visual function decline in the population, which may include age, body mass index, gender, medication history, and past eye disease history.

[0020] In a method that is easy to implement, age and body mass index are prioritized for subsequent group graph propagation, while other demographic data can be used as supplementary criteria for neighborhood screening.

[0021] 2> Based on clinical examination results or standard grading rules, it is classified as the first... A multi-source vision function sample Generate realistic visual function decline level labels True visual function decline level label It refers to an ordered category label of the degree of visual function decline of the subject, preferably divided into 4 levels, namely normal, mild decline, moderate decline and severe decline.

[0022] In practice, the true level of visual function decline can be determined based on visual acuity, contrast sensitivity, visual field impairment indicators, and the comprehensive judgment of clinicians. For example, if a subject's visual acuity and contrast sensitivity are within the normal range and there are no significant abnormalities in visual field impairment indicators, they are labeled as normal; if a subject has a slight decrease in contrast sensitivity or localized visual field defects, but the impact on daily visual function is minor, they are labeled as mild decline; if a subject has abnormalities in multiple visual function dimensions that affect daily activities, they are labeled as moderate decline; if a subject has significant visual field defects, low visual acuity, or severe decline in multiple visual functions, they are labeled as severe decline. Specific grading thresholds can be set according to clinical standards or research dataset labeling specifications, but the same grading rules must be used in the training, validation, testing, and inference phases.

[0023] 3> is the first A multi-source vision function sample Generate real retinal age labels True retinal age label It refers to a continuous numerical label used to monitor the regression of retinal physiological aging. It can be the subject's actual age or the retinal biological age output by an independent retinal age assessment system.

[0024] In one implementation, if independent retinal age assessment results are unavailable, the subject's actual age is used as the true retinal age label. This is used to constrain the model's learning of age-related changes in fundus structures; if a retinal biological age exists from an independent imaging model or clinical assessment, it is preferentially used as the true retinal age label. It is used to more directly characterize the physiological aging degree of retinal tissue.

[0025] 4. All multi-source visual function samples are divided into training, validation, and test sets. Standardized parameters, principal component transformation parameters, demographic neighborhood samples, and model training parameters are calculated only on the training set. The training set is the dataset used to calculate feature processing parameters and update model parameters; the validation set is the dataset used to select model hyperparameters and save the best model; and the test set is the dataset used to finally evaluate model performance.

[0026] It should be noted that the mean and standard deviation of visual function testing indicators, the principal component direction after psychological correction, the training sample neighborhood in demographic graph propagation, and the amplitude normalization range are all only allowed to be calculated from the training set. The validation set, test set, and online inference stage all directly use the processing parameters saved in the training stage, without recalculating global statistics based on the samples to be evaluated or the test samples, in order to avoid data leakage and distortion of evaluation results.

[0027] S2: Divide the fundus image into regions and calculate the grayscale features of each region; use the fundus image and the grayscale features of each region to perform feature encoding to obtain the global retinal structure features; In one embodiment, the partitioning process includes: extracting brightness from the fundus image according to the three channels (red, green, and blue) and then performing weighted fusion to obtain a single-channel brightness map; calculating the geometric features of each bright region in the single-channel brightness map; filtering candidate regions for optic disc features based on the geometric features; fitting the candidate regions to circles or ellipses and calculating the optic disc center and radius; using the optic disc center as the origin of polar coordinates, calculating the radial distance and angular direction of each pixel position relative to the optic disc center; and dividing the region based on the radial distance, angular direction, and optic disc radius; the geometric features include any one or more of area, circularity, and vascular convergence features.

[0028] In one embodiment, the feature encoding is: generating a mask of p optic disc partitions by partitioning the fundus image, where p is a natural number greater than 1; Multi-scale feature extraction is performed on fundus images to obtain multi-scale feature maps; the mask of each optic disc region is multiplied element-wise with the multi-scale feature map to obtain p weighted feature maps of the regions; Feature extraction is performed on the weighted feature maps of p partitions to obtain the depth features of p partitions; the depth features of p partitions are fused with the grayscale features of the corresponding partitions, and then attention is calculated to obtain the attention weight of each partition; The global retinal structure features are obtained by weighting and summing the depth features of p partitions with attention weights; gray-level features include any one or more of the following: average gray-level intensity, quantile contrast, gray-level co-occurrence matrix contrast, and gray-level co-occurrence matrix entropy.

[0029] In one specific embodiment, fundus structural damage, visual function detection abnormalities, psychological and cognitive states, and demographic decline patterns reflect different aspects of visual function decline. Directly inputting raw data into the model can easily lead to the model learning accidental correlations under small sample conditions, and it is difficult to explain the contribution of different data sources to the evaluation results. Therefore, this invention first extracts fundus structural feature vectors based on optic disc anatomical regions, then performs confidence correction on the original visual function vectors based on psychological and cognitive states, and finally performs nonparametric graph propagation enhancement based on demographic similarity to obtain fundus structural feature vectors, a comprehensive visual function impairment index vector, and group-enhanced visual function features. The processing of fundus structural feature vectors includes the construction of fundus structural feature vectors based on the course of nerve fiber layers in optic disc regions. Because the nerve fiber layers around the optic disc have different course directions and thickness distributions in the superior, inferior, nasal, and temporal regions, the reduced reflectivity, thinning of texture stripes, and local structural atrophy of the nerve fiber layers related to visual function decline often exhibit regional differences. If only global texture features of the entire fundus image are extracted, it is easy to mask the differences between local anatomical regions. To this end, this invention constructs a fundus structure feature vector with fixed spatial meaning through optic disc localization, scale normalization, anatomical partition mask generation, and partition texture statistics. Specifically: 1> Original fundus images Normalized fundus images are obtained by performing size uniformity, color channel uniformity, and pixel intensity normalization. Normalized fundus images This refers to a color fundus image that has been entered into a standardized format, with a preferred size of [size missing]. ,in Indicates spatial resolution. These represent the red, green, and blue channels.

[0030] In the specific implementation, the original fundus image is first... The images are adjusted to a fixed size by scaling proportionally along the long side or cropping to the center; then the pixel values ​​are scaled to the range of 0 to 1; if there is significant color deviation between different fundus cameras, color standardization can be further performed using the mean and standard deviation of the color channels in the training set. Based on this, fundus images acquired by different devices and under different exposure conditions have a consistent input range, reducing the impact of device differences on subsequent regional feature statistics.

[0031] 2> Based on normalized fundus images Determine the center and radius of the optic disc, with the center denoted as . ,in Indicates the center of the optic disc in the normalized fundus image The horizontal coordinate in the middle, Indicates the center of the optic disc in the normalized fundus image The vertical coordinate in the image; the radius of the viewing disk is denoted as... , Used to characterize the scale of the optic disc region.

[0032] In one implementation, normalized fundus images are first used... The brightness map is extracted, and then local contrast enhancement and threshold segmentation are performed on the brightness map to obtain candidate bright regions. Then, candidate regions of the optic disc are selected based on the area, circularity, and vascular convergence features of the candidate regions. Finally, circular or elliptical fitting is performed on the boundaries of the candidate regions of the optic disc to obtain the center of the optic disc. and the radius of the viewing disc .

[0033] In this implementation, the normalized fundus image Extracting the brightness map from the data, specifically... The red, green, and blue color channels are converted into a single-channel grayscale image using a weighted average method. This grayscale image is the luminance image. The weighting coefficients can be set according to the contribution of each color channel to the luminance. For example, a common luminance conversion formula can be used: Luminance value = ,in , , These represent the pixel values ​​for the red, green, and blue channels, respectively.

[0034] In this implementation, local contrast enhancement can be achieved using a contrast-limited adaptive histogram equalization method. First, the brightness map is divided into multiple small regions (e.g., (Pixel patches). Then, histogram equalization is performed in each region to enhance local contrast while limiting the magnitude of contrast amplification to avoid excessive noise amplification. Thresholding segmentation involves setting a brightness threshold on the enhanced image, marking pixels above the threshold as foreground (candidate regions) and pixels below the threshold as background. This threshold can be a fixed value or determined using adaptive methods such as Otsu's method.

[0035] In this implementation, when filtering candidate regions for the optic disc, the area, circularity, and vascular convergence features of each highlighted candidate region can be calculated separately. The area can be obtained by counting the number of pixels within the candidate region. Circularity can be calculated using the formula... The closer the value is to 1, the more circular the shape. Vessel convergence features can be obtained by analyzing the direction of blood vessels at the edge of the candidate region. Because retinal vessels converge radially at the optic disc, a typical feature is that the gradient direction of the pixels at the edge of the candidate region points towards the center of the region. By setting area range, lower limits for circularity, and lower limits for vessel convergence, candidate regions that best match the optic disc characteristics can be selected.

[0036] In this implementation, the set of boundary pixels of the candidate viewport region can be fitted with a circle or an ellipse using the least squares method. Circular fitting finds the optimal center coordinates by minimizing the error between the sum of the squares of the distances from the boundary pixels to the center and the square of the radius. and radius Ellipse fitting is more flexible, allowing for slight deformation of the optic disc shape. It solves for the general equation parameters of the ellipse by minimizing the algebraic distance, thereby obtaining the center coordinates, major axis, and minor axis.

[0037] It should be noted that all subsequent anatomical divisions are based on the same optic disc center. and the same viewing radius This process avoids inconsistencies in partition locations caused by using different spatial references in different steps.

[0038] 3> Center the view A circular statistical region around the viewing disk is generated as a reference, and a polar coordinate reference system is established within this region. Specifically, the radius of the circular statistical region around the viewing disk is denoted as... , This refers to the maximum radius of the structure surrounding the visual disc that is included in the statistics, with the preferred radius being... For normalized fundus images any pixel position in If the pixel position is near the center of the display disc The distance is no greater than If the pixel location is included in the circular statistical area around the viewing disk, then the pixel location is included in the statistical area of ​​the structure around the viewing disk; otherwise, the pixel location is not included in the statistical area of ​​the structure around the viewing disk.

[0039] Furthermore, with the center of the visual disc Using the origin of polar coordinates, calculate the position of each pixel. Relative to the center of the viewing disc The radial distance and angular direction are used to determine whether the pixel location belongs to the core region of the optic disc or the annular region surrounding the optic disc; the angular direction is used to determine whether the pixel location belongs to the upper region, lower region, nasal region, temporal region, or the region corresponding to the fovea of ​​the macula.

[0040] In one implementation, if left and right eye information can be obtained from the acquisition device or medical records, the nasal and temporal directions are directly determined based on this information. If left and right eye information is unavailable, a candidate region for the fovea that is darker, approximately elliptical in shape, and has low vascular density is searched on the outer side of the optic disc. The direction of this candidate region is taken as the temporal direction, and its opposite direction is taken as the nasal direction. Based on this, the anatomical direction definition remains consistent even when left and right eye information is missing.

[0041] In this implementation, for any pixel position within the circular statistical area surrounding the view disk... Its relative to the center of the optic disc radial distance Calculated using the Euclidean distance formula: ;in, Indicates pixel position The straight-line distance to the center of the viewing disc, expressed in pixels. The angular direction of this pixel position relative to the center of the viewing disc. Calculated using the arctangent function: ;in, It is the arctangent function in the fourth quadrant. The range of values ​​is Radius; preferably, from horizontal to the right (in the image coordinate system) (positive direction of axis) The radius is radian, with counterclockwise rotation being positive.

[0042] 4> Divide the circular statistical area around the viewing disc into Anatomical regions Represents the total number of anatomical regions, with the preferred selection. The six anatomical regions are, in order, the inferior region, the superior region, the nasal region, the temporal region, the region corresponding to the fovea of ​​the macula, and the optic disc core region. The set of pixels for each anatomical region is denoted as , Used to limit the The pixel locations within each anatomical region that participate in structural feature statistics; among them... This represents the anatomical partition index. .

[0043] In the specific implementation, first ensure that the radial distance is not greater than The pixel positions are then mapped into the core region of the optic disc; then, the region is mapped with the fovea as the center, an angle width of 30 to 60 degrees, and a radial range of... to The fan-shaped region is assigned to the corresponding region of the fovea; finally, the remaining annular region is divided into the lower region, upper region, nasal region, and temporal region according to the angular direction. If a pixel location satisfies multiple partitioning conditions, it is assigned according to the priority order of the optic disc core region, the corresponding region of the fovea, and other annular regions, so that each pixel location belongs to at most one anatomical partition, avoiding duplicate counting.

[0044] In one implementation, the determined temporal direction is defined with the optic disc center as the origin. direction (angle Following the positive angle direction in step 3 of S201 (counterclockwise is positive), the angle divisions of each annular region are as follows: Temporal region corresponding angle range The area above corresponds to The nasal side area corresponds to The area below corresponds to If information from the left and right eyes is missing, first determine the temporal direction (i.e., the direction of the fovea centralis) and then divide it according to the above rules.

[0045] 5> From normalized fundus images Extracting grayscale fundus images grayscale fundus images This refers to normalized fundus images A single-channel fundus image obtained by fusing the three color channels is used to characterize the reflectivity and texture changes of the tissues surrounding the optic disc.

[0046] In one implementation, the weight of the green channel is prioritized because it typically reveals the reflective texture of the retinal blood vessels and nerve fiber layer more clearly. Simultaneously, the small contributions of the red and blue channels are retained to reduce the impact of single-channel noise. If significant blood vessel occlusion exists in the retinal image, a linear structure enhancement operator can be used to extract candidate regions of blood vessels. These candidate regions are then assigned lower weights or excluded from the statistical analysis during partitioning. Based on this, the interference of coarse blood vessel shadows on the texture statistics of the nerve fiber layer can be reduced, allowing the retinal structural feature vector to focus more on the reflectivity and texture changes of the nerve fiber layer itself.

[0047] In this implementation, grayscale fundus images By analyzing normalized fundus images The three color channels (red, green, and blue) are weighted and fused together. To enhance the contrast of the nerve fiber layer and blood vessels around the optic disc, the weight of the green channel is preferentially increased. The calculation method is expressed as follows: ;in, , , Represent The red, green, and blue channels are used. This weighted combination fully utilizes the advantage of the green channel in displaying fundus details, while incorporating a small amount of red and blue channel information to reduce overall noise.

[0048] 6> Calculate the average gray intensity, quantile contrast, and gray-level co-occurrence matrix texture features within each anatomical region to obtain the first... The local structural descriptor of each anatomical region. The average gray intensity is used to characterize the first... The overall reflectance level of each anatomical region; quantile contrast is used to characterize the first... The difference in brightness within each anatomical region can reduce the impact of extreme noise pixels on contrast statistics; the gray-level co-occurrence matrix texture feature is used to characterize the gray-level combination relationship of adjacent pixels, which can reflect the continuity of nerve fiber layer stripes and local roughness.

[0049] In the specific implementation, first convert the grayscale fundus image Quantize to a fixed grayscale level; the number of grayscale levels can be 32 or 64; then at the... A collection of pixels from anatomical partitions Internal calculation of average gray intensity ,in Used to characterize the The overall reflectance level of each anatomical region; then calculate the... The difference between the 90% grayscale quantile and the 10% grayscale quantile within each anatomical region is used to obtain the quantile contrast. ,in Used to characterize the Stable local brightness and darkness differences within each anatomical region; finally, a gray-level co-occurrence matrix is ​​constructed according to a pixel distance of 1 pixel and orientations of 0 degrees, 45 degrees, 90 degrees, and 135 degrees. The gray-level co-occurrence matrix contrast and entropy are averaged across the four orientations to obtain the gray-level co-occurrence matrix contrast. and gray-level co-occurrence matrix entropy .in, Used to characterize the The range of local grayscale changes in each anatomical region Used to characterize the Texture complexity of each anatomical partition.

[0050] It should be noted that the local structure descriptor is used for the comprehensive description of the first... The vector of texture and reflectance features within each anatomical partition is derived from the average grayscale intensity calculated within that partition. Quantile contrast Gray-level co-occurrence matrix contrast and gray-level co-occurrence matrix entropy These four statistics are pieced together.

[0051] 7> By assembling the local structural descriptors of each anatomical region according to a fixed anatomical region order, the fundus structural feature vector is obtained. Fundus structure feature vector This refers to the non-trained structural feature vector obtained by concatenating the reflectance intensity features and texture features of the lower region, upper region, nasal region, temporal region, the corresponding region of the fovea, and the core region of the optic disc. This is used to characterize the differences in structural damage in different anatomical regions around the optic disc, with 24 dimensions. This represents the vector concatenation operator, used to join local structural descriptors from different anatomical regions in a fixed order. Indicates from index arrive , will the Local structural descriptors generated from each anatomical region Perform the concatenation (join) operation on the first and last parts in the order of the partition index; Indicates the first Average gray intensity of each anatomical region; Indicates the first Quantile contrast of each anatomical region; Indicates the first Gray-level co-occurrence matrix contrast of each anatomical region; Indicates the first The gray-level co-occurrence matrix entropy of each anatomical partition.

[0052] It should be noted that the fixed anatomical partitioning order remains consistent throughout the training, validation, testing, and inference phases. This is achieved by transforming the anatomical partitioning patterns of the peripapillary nerve fiber layer into a fixed-dimensional feature vector of fundus structures. It can provide a stable structural prior for the model without increasing the number of trainable parameters, and reduce the model's dependence on background brightness, random blood vessel textures and acquisition noise.

[0053] In one embodiment, for example, the input raw fundus image is standardized to Normalized fundus images The center of the optic disc was detected. and the radius of the viewing disc If there are 1 pixel, then the radius of the circular statistical area around the viewport is... Each pixel is then used to generate six anatomical regions, and within each region, the average grayscale intensity, quantile contrast, gray-level co-occurrence matrix contrast, and gray-level co-occurrence matrix entropy are statistically analyzed to obtain the final result. 3D fundus structure feature vector .

[0054] S3: Generate psychological confidence coefficients from psychological cognitive data, and use these psychological confidence coefficients to correct visual function data to obtain corrected visual function data; In one embodiment, the psychological cognitive data includes any two or more of the following: depression scale scores, anxiety scale scores, and Mini-Cognitive Scale scores. The composite score is obtained by normalizing any two or more of the following scores, then weighting and summing them, and finally converting the composite score using a smoothing decreasing function to obtain the psychological confidence coefficient.

[0055] In one specific embodiment, visual function test results are easily affected by the subject's attention, emotional state, and cognitive level. When the level of depression, anxiety, or cognitive cooperation decreases, non-organic fluctuations may occur in test indicators such as visual acuity, contrast sensitivity, stereopsis, and motion perception. If the raw visual function test data is used directly, the model may misjudge the measurement fluctuations caused by psychological and cognitive states as actual visual function decline. Therefore, this invention generates a psychological confidence coefficient based on psychological and cognitive scale data, performs dimensionality-sensitive contraction correction on standardized visual function values, and then obtains a comprehensive visual function impairment index vector through fixed principal component transformation, including: 1> For the original visual function vector The execution directions of each visual function dimension are unified and standardized to obtain a standardized visual function vector. Standardized visual function vectors This refers to the visual function vector obtained by converting visual function detection indicators with different dimensions and directions to a unified direction of impairment and a unified numerical scale. The standardized visual function value for each visual function dimension is denoted as... ,in This indicates the view function dimension index. .

[0056] In practical implementation, the first step is to unify the direction. Specifically, for test indicators whose original meaning is "the larger the value, the better the visual function" (such as visual acuity and contrast sensitivity), they are transformed into damage indicators that "the larger the value, the more severe the visual impairment" through a monotonically decreasing transformation. It is preferable to use the theoretical maximum value of the indicator minus the original value or its opposite. For test indicators whose original meaning is "the larger the value, the more severe the impairment," the original direction is kept unchanged. The direction is then recorded as follows: The first subject The value of the individual visual function dimension is ,in This is the index for the visual function dimension. Then, using only training set samples, the oriented unified index is calculated. Mean of each dimension and standard deviation z-score standardization was performed on all samples to obtain the standardized visual function values. ;in, Indicates the first training set The average value of each visual function dimension after unifying the direction. Indicates the first training set The standard deviation of each visual function dimension after unifying the directions. (This will include all...) Each dimension By assembling them in a fixed order, a standardized visual function vector is formed. The verification, testing, and inference phases all use data saved during the training phase. and Standardization should be implemented. Based on this, Each dimension has a uniform damage direction and unit variance, and its training set mean is 0.

[0057] 2> Data from psychological cognition scales This is converted into a unified range of psychological and cognitive state characteristics. (Psychological and cognitive scale data) It is a multidimensional vector, the dimension of which is equal to the number of scales collected, preferably psychological cognitive scale data. Including the total score of the depression scale Anxiety Scale Total Score and total score of the Mini-Cognitive Scale .in, Used to characterize the intensity of a depressive state, with values ​​ranging from 0 to 27; Used to characterize the intensity of anxiety, with values ​​ranging from 0 to 21; Used to characterize cognitive cooperation level, with values ​​ranging from 0 to 5.

[0058] In one implementation, the total score of the depression scale Anxiety Scale Total Score and total score of the Mini-Cognitive Scale The total scores of these scales are derived from data from psychological cognitive scales. The pre-defined structured fields were extracted. During the data collection phase, participants completed standard scales such as the PHQ-9 (Patient Health Questionnaire-9 Depression Scale), GAD-7 (Generalized Anxiety Scale), and Mini-Cog (Simplified Cognitive Assessment Scale). The total scores of these scales were used as data for the psychological cognitive scale. The components are recorded and stored. During the implementation of the method of this invention, directly from... These preset scores are read from the database, rather than being calculated in real time.

[0059] In this implementation, Divide by 27 to obtain the normalized depression scale score; Divide by 21 to obtain the normalized anxiety scale score; Divide by 5 to obtain the normalized simplified cognitive scale score. Since a higher total score on the simplified cognitive scale usually indicates more stable test compliance, subtract the normalized simplified cognitive scale score from 1 to obtain the degree of cognitive impairment. Based on this, increased depression, increased anxiety, and cognitive impairment all contribute to an increase in the intensity of psychological disturbance.

[0060] In one embodiment, for example, the data of a typical subject with a good mental and cognitive state might be as follows: However, the data from a subject in poor condition may be... First dimension This represents the total score of the depression scale, with a range of 0-27 (example); the second dimension... This represents the total score of the anxiety scale, with a range of 0-21 (example); the third dimension. This represents the total score of the Simplified Cognitive Scale, with a range of 0-5.

[0061] 3> Generate psychological confidence coefficients based on characteristics of psychological cognitive states. Psychological confidence coefficient It refers to the measure of the first The continuity coefficient for the reliability of visual function test results for each subject, with a value range of [value missing]. Up to 1, of which This represents the minimum retention rate, with a preferred value of 0.5. Psychological confidence coefficient. The closer the value is to 1, the more reliable the visual function test data; psychological confidence coefficient The closer This suggests that visual function test data is more likely to be affected by psychological and cognitive states.

[0062] In one implementation, the normalized depression scale score, the normalized anxiety scale score, and the degree of cognitive impairment are first summed according to their weights to obtain the comprehensive score of psychological disturbance. , It refers to the first The overall psychological disturbance score of each subject is used to characterize the overall impact of psychological and cognitive state on the stability of visual function testing; then, the overall psychological disturbance score is processed by a smoothing decreasing function. Converted to psychological confidence coefficient , is represented as: ; in, Indicates the minimum retention percentage; Represents the natural exponential function, used to generate a continuous and smooth confidence adjustment curve; psychological disturbance composite score. The influence strength of each scale can be controlled by the depression regulation coefficient, anxiety regulation coefficient, and cognitive insufficiency regulation coefficient. Each regulation coefficient can be taken from 0.5 to 2.0, and in a convenient implementation method, all of them are taken as 1.0.

[0063] 4> Based on psychological confidence coefficient For standardized visual function vectors Performing dimension-sensitive contraction correction yields the psychologically corrected visual function vector. Psychologically corrected visual function vector This refers to the standardized visual function vector based on psychological cognitive state. The confidence-adjusted visual function vector is used to reduce the impact of non-organic fluctuations on the expression of visual function impairment.

[0064] In the actual implementation, psychological sensitivity weights are set for each visual function dimension. , Indicates the first The degree to which visual function dimensions are influenced by psychological cognitive state, with values ​​ranging from 0 to 1. Indicators requiring high subjective coordination, such as contrast sensitivity, motion perception, and stereopsis, can be assigned higher psychological sensitivity weights; relatively stable objective examination indicators can be assigned lower psychological sensitivity weights. The calculation method for the psychologically corrected visual function value of each visual function dimension is expressed as follows: ; in, Indicates the first The first subject The psychologically corrected visual function value for each visual function dimension is the psychologically corrected visual function vector. The One element; Indicates the first training set The mean of each standardized visual function dimension is typically close to 0.

[0065] It should be noted that this contraction correction does not involve multiplying all visual function dimensions by a fixed percentage, but rather based on the psychological confidence coefficient. And psychological sensitivity weight This shrinks the visual function dimension with low confidence and high psychological sensitivity towards the mean of the training set, thereby reducing abnormal deviations caused by psychological cognitive state, while preserving as much as possible the true organic differences in visual function.

[0066] In one specific embodiment, the visual function vector after orientation unification and z-score normalization is compared with the psychologically corrected visual function vector after psychological confidence coefficient contraction correction. Figure 2 As shown, the blue bars represent the original standardized values, and the orange bars represent the corrected values. It can be seen that on some highly psychologically sensitive dimensions (such as dimensions 2 and 6), the corrected values ​​significantly shrink towards the training set mean (0 line). Psychological correction can effectively reduce non-organic fluctuations caused by depression, anxiety, or cognitive deficiencies, making the expression of visual function impairment closer to the true organic state. The horizontal axis represents the visual function dimension index (1-11, unitless), and the vertical axis represents the standardized impairment index value (the larger the value, the more severe the impairment; dimensionless).

[0067] S4: The comprehensive visual function impairment index vector is calculated from the corrected visual function data through principal component transformation; In one embodiment, the principal component transformation calculation involves centering the corrected visual function data, projecting it onto the principal component space, and then dividing each element by the whitening scaling vector to obtain the comprehensive visual function impairment index vector. The whitening scaling vector is constructed using the covariance matrix of the corrected visual function data in the sample set.

[0068] In one specific embodiment, fixed principal component transformation parameters are calculated based on the psychologically corrected visual function vectors from the training set, and the psychologically corrected visual function vectors are then... Converted into a comprehensive visual impairment index vector Comprehensive visual impairment index vector This refers to the visual function vector of psychological correction. The low-dimensional visual function impairment expression obtained after principal component projection and whitening is used to compactly characterize common abnormal patterns in multiple visual function dimensions, with a dimension of 6.

[0069] In practical implementation, fixed principal component transformation parameters are calculated based on the psychologically corrected visual function vectors in the training set, and the psychologically corrected visual function vectors are then... Converted into a comprehensive visual impairment index vector Specifically, firstly, only the mentally corrected visual function vectors of all samples in the training set are used. Calculate the mean vector And center each sample to obtain ;in, The training set mean vector represents the vector of the mentally corrected visual function, with dimension . Its first The element is the i-th element among all samples in the training set. The mean of the psychologically corrected visual function values ​​is used to assess... Centralized execution; It refers to the first The psychologically corrected visual function vector after centering each sample has a dimension of . Then, the covariance matrix of the centered vectors of the training set is calculated, and its eigenvalues ​​are decomposed. The eigenvalues ​​are then arranged in descending order, and the top eigenvalues ​​are retained. eigenvalues and their corresponding eigenvectors, which are then arranged column-wise to form a projection matrix. , By the retained former The eigenvectors corresponding to each eigenvalue are concatenated column-wise, with a dimension of . Used to project the centered vector to Principal component space; where, The first term represents the result of eigenvalue decomposition of the covariance matrix of the psychologically corrected visual function vectors centered on the training set. 1 eigenvalue, And satisfy This represents the variance along each principal component direction and is used to construct the whitening scaling vector. To eliminate scale differences among the principal components, the whitening scaling vector is calculated. Then, for any sample, its centered vector is... pass Project onto the principal component space, then divide element by element. The comprehensive visual function impairment index vector is obtained. ;in, Represents the whitening scaling vector. , dimension , used to divide element by element to eliminate scale differences among principal component components; This represents element-wise division. New samples in the validation, testing, and inference phases all use samples saved during the training phase. , and A fixed transformation is performed to avoid inconsistencies in the meaning of principal components between different data partitions and to prevent information leakage.

[0070] It should be noted that this step transforms psychological and cognitive states into confidence modulation signals for visual function test data, and performs personalized contraction correction in combination with the psychological sensitivity of different visual function dimensions. Compared with directly eliminating samples with abnormal psychological states, this step can preserve the effective visual function information of the subjects, while weakening non-organic fluctuations caused by depression, anxiety, and cognitive deficiencies.

[0071] In one embodiment, for example, if a subject's total score on a depression scale is 20, their total score on an anxiety scale is 16, and their total score on a mini-cognitive scale is 1, then their overall psychological disturbance score is high, and their psychological confidence coefficient is high. Reduce; weight of psychological sensitivity The contrast sensitivity dimension shows that the standardized visual function value of this dimension will significantly shrink towards the training set mean; for the psychological sensitivity weights... This is a relatively stable objective indicator, and this dimension only contracted slightly. If another subject's total score on the depression scale is 2, the total score on the anxiety scale is 1, and the total score on the mini-cognitive scale is 5, then the psychological confidence coefficient... The result is close to 1, and the original standardized results are basically retained for each visual function dimension.

[0072] S5: Fusion features are obtained by modulating global retinal structural features using a comprehensive visual function impairment index vector; In one embodiment, the modulation is achieved by generating an individual impairment scaling vector and an individual impairment bias vector by integrating the visual function impairment index vector. The individual impairment scaling vector is multiplied by the global retinal structural features and then added to the individual impairment bias vector to obtain the fused features. The generation is achieved by inputting the integrated visual function loss index vector into a mapping network with hidden layers and full connections.

[0073] In one embodiment, the method further includes group-enhanced visual function features, acquiring demographic data of the tested samples, screening a similar sample set using the demographic data of the tested samples, calculating the similarity between samples in the sample set and the tested samples, weighting the comprehensive visual function impairment index vector of the samples in the sample set using the similarity to obtain a domain visual function vector, fusing the domain visual function vector with the comprehensive visual function impairment index vector of the tested samples to obtain group-enhanced visual function features, modulating global retinal structural features using the comprehensive visual function impairment index vector and group-enhanced visual function features to obtain a first fusion feature, and using the first fusion feature to assess the level of visual function decline and / or retinal age to obtain an assessment result; the demographic data includes age and body mass index, the sample set is screened by age and body mass index, the age difference between the tested sample and the samples in the sample set is less than or equal to a first preset threshold, and the difference between the body mass index of the tested sample and the samples in the sample set is less than or equal to a second preset threshold, thus obtaining a similar sample set; the similarity is calculated by using age and body mass index to measure the similarity of the samples.

[0074] The basis for screening similar sample sets can be screening from training set samples or from collected and processed database data on visual functions.

[0075] Optionally, the global retinal structural features are first modulated by the comprehensive visual function impairment index vector, and then a second modulation is performed by the group-enhanced visual function features.

[0076] In one specific embodiment, demographic information such as age and body mass index exhibits a stable group pattern with visual function decline. Subjects of similar age and body mass index typically show similar decline trends. If the comprehensive visual function impairment index vector of a single subject is used in isolation, it is easily affected by single-detection noise and random fluctuations. Therefore, this invention constructs a nonparametric neighborhood graph based on demographic similarity and propagates the comprehensive visual function impairment index vector of similar training samples to the current sample to obtain group-enhanced visual function features, including: 1> From demographic data Extract age and body mass index to construct a demographic feature vector. Demographic feature vector It refers to a low-dimensional feature vector used to search for similar group samples, which includes at least age. and Body Mass Index ;in, Indicates the first The age of each subject is used to characterize life stage; Indicates the first The body mass index of each subject is used to characterize their body mass level.

[0077] In practical implementation, if the medication history or previous eye disease history is complete, it can be encoded as a binary or categorical variable, and training samples with consistent categories or similar clinical significance should be prioritized during neighborhood screening. To avoid sample sparsity due to too many neighborhood conditions, in an easy-to-implement approach, age and body mass index are used as basic neighborhood search conditions, while other demographic information is used as optional constraints.

[0078] 2> For the first For each subject, a demographic neighborhood set is obtained by searching for demographically similar samples in the training set. Demographic neighborhood set It refers to the first A training sample set of subjects with similar demographic characteristics was used to restrict population augmentation to samples from similar age groups and similar body mass levels.

[0079] In practice, training samples that meet the criteria of an age difference of no more than 5 years and a body mass index difference of no more than 3 are preferentially selected. and The There are training samples; among them, Indicates the training sample index. Indicates the first The age of each training sample. Indicates the first The body mass index of each training sample. If the number of training samples meeting the condition exceeds the number of samples in the maximum neighborhood. Then, the top candidates are selected according to their demographic similarity from highest to lowest. training samples, This represents the maximum number of neighborhood samples, which can be between 10 and 30. If no training samples that meet the conditions are found, graph propagation is not performed, and the comprehensive visual function impairment index vector is directly transferred. As a feature of group-enhanced visual function .

[0080] It should be noted that during the training phase, when performing graph propagation on the training samples, the current training sample itself is not included in its demographic neighborhood set. To avoid self-looping causing the neighborhood averaging to lose its smoothing effect, new samples in the validation, testing, and online inference phases can only search for neighborhood samples in the training set; cross-propagation between validation or testing set samples is not allowed.

[0081] 3> Calculate the first The subject and the first Demographic similarity between neighboring training samples Demographic similarity It is used to measure how close two samples are in the demographic space; the larger the value, the more similar the two samples are.

[0082] In one implementation, demographic similarity The Gaussian decay calculation, based on the age difference and body mass index difference, is expressed as follows: ; in, This represents the age bandwidth parameter, used to control the impact of age differences on demographic similarity. The optimal age for assessing the impact is 5 years. This represents the body mass index bandwidth parameter, used to control the impact of differences in body mass index on demographic similarity. The influence scale is preferably set to 3; This represents the natural exponential function, used to calculate demographic similarity. It decreases smoothly as the age difference and body mass index difference increase.

[0083] 4> Based on demographic similarity The visual function impairment index vector of the neighborhood training samples is weighted and averaged to obtain the visual function trend vector of the neighborhood group. Neighborhood group visual function trend vector It refers to the first The visual impairment trend formed by training samples of subjects with similar demographic characteristics is used to provide constraints on the population decline pattern.

[0084] In the specific implementation, the comprehensive visual function impairment index vector of each neighborhood training sample is used. Multiply by the corresponding demographic similarity Then, for the demographic neighborhood set The sum of all neighboring training samples is calculated, and then divided by the sum of similarities. If the sum of similarities is close to 0, a minimal constant is used. To ensure computational stability, Desirable .

[0085] In one embodiment, as an example, suppose the first Demographic neighborhood set of each subject Includes 2 training samples and The comprehensive visual impairment index vector (assumed to be 3-dimensional for simplicity) is as follows: and The calculated demographic similarities are as follows: and Therefore, the total similarity is... Neighborhood group visual function trend vector The calculation is as follows: .

[0086] 5> The first Vector of the comprehensive visual impairment index of each subject Visual function trend vector of neighboring groups Fusion yields group-enhanced visual function characteristics Group-enhanced visual function characteristics It refers to visual function characteristics that incorporate the decline trend of similar populations while preserving the individual's visual function impairment expression, with a dimension of 6.

[0087] In one implementation, a graph smoothing intensity coefficient is used. The calculation method for controlling the fusion ratio of individual and group information is as follows: ; in, The graph smoothing intensity coefficient represents the strength of the influence of the neighborhood group trend on the final feature. It can be taken from 0.3 to 0.7, with 0.5 being preferred.

[0088] It should be noted that this step utilizes similar population samples from the training set to provide a stable group prior. This approach reduces the impact of noise from single visual function tests on the overall visual function impairment index vector. This will have an impact, while also preventing data leakage from spreading within the validation or test sets.

[0089] In one embodiment, for example, the first If a participant is 68 years old and has a body mass index (BMI) of 24, then the training sample aged 63 to 73 years with a BMI between 21 and 27 will be preferentially selected from the training set as the demographic neighborhood set. If 18 neighborhood training samples are obtained, then the demographic similarity of each neighborhood training sample is calculated. The visual function impairment index vector of the 18 neighboring training samples was weighted and averaged to obtain the visual function trend vector of the neighborhood group. Finally, it was compared with the first... The individual subject's comprehensive visual function impairment index vector Fusion yields group-enhanced visual function characteristics .

[0090] In one specific embodiment, such as Figure 4 The demographic neighborhood search shown uses age (horizontal axis, unit: years) and body mass index (vertical axis, unit: kg / m²) as coordinates, displaying the training set samples (gray dots), the current sample (red asterisk, age 68, BMI 24), and neighborhood samples (blue dots, several in total) that meet the age difference ≤ 5 years and BMI difference ≤ 3. The red dashed rectangle marks the neighborhood search range. Based on age and BMI, demographically similar groups can be selected from the training set, forming the basis of the nonparametric neighborhood graph. This provides reasonable group reference samples for subsequent construction of enhanced visual function characteristics of the group, avoiding the use of samples with excessively different demographic characteristics from the current sample for trend estimation.

[0091] S6: The assessment results are obtained by evaluating the degree of visual function decline and / or retinal age through fusion features. In one embodiment, the grayscale features of each region are fused to obtain fundus structural features. The fundus structural features and the comprehensive visual function impairment index vector are then input into the evaluation model to assess the degree of visual function decline and / or retinal age. Alternatively, the group-enhanced visual function features, fundus structural features, and the comprehensive visual function impairment index vector are input into the evaluation model to assess the degree of visual function decline and / or retinal age.

[0092] Specifically, global retinal structural features are obtained by feature encoding using fundus images and fundus structural features. During the feature encoding process, grayscale features of each region are extracted using fundus structural features.

[0093] In one specific embodiment, visual function decline level and retinal age reflect the degree of disease progression and the degree of physiological aging of the retina, respectively. These two typically exhibit a monotonic correlation. Conventional models, if simply spliced ​​and independently trained on fundus images, visual function vectors, and age regression tasks, are prone to contradictory outputs such as high visual function decline level but low retinal age prediction. Therefore, this invention constructs a multi-task evaluation model, which includes an anatomical partition encoder, a dual-channel conditional gating module, a decline level prediction head, a retinal age regression prediction head, and physiological consistency ranking constraints, including: Construction of an anatomical partition encoder for retinal structural damage: Because the texture details of the nerve fiber layer around the optic disc are small, and different anatomical regions have different sensitivities to visual function decline, if a regular convolutional network is used to encode the entire normalized fundus image indiscriminately, the network is easily affected by blood vessels, exposure, and background texture. This invention introduces an anatomical partition mask of the optic disc into the feature map level, so that the model can extract the depth structural features of different anatomical partitions respectively, and then obtain the global retinal structural depth features through partition attention aggregation.

[0094] 1> Normalize fundus images Input image basic encoding module to obtain multi-scale basic feature map Multi-scale basic feature map This refers to the normalization of fundus images by the image fundamental coding module. After multi-stage convolution processing, the feature maps from different stages are fused through upsampling and concatenation operations to form a feature map with multi-scale information. The image base coding module refers to a trainable image coding network composed of multiple layers of convolution, normalization, activation, and downsampling operations, used to process normalized fundus images. Extract the structural response around the view disk.

[0095] In one implementation, the image basic coding module comprises three convolutional stages, each stage including two convolutional kernels with a size of [missing information]. The algorithm consists of convolutional layers, batch normalization layers, ReLU activation layers, and first-order spatial downsampling. The first convolutional stage preserves fine texture and small-scale reflectance changes in the nerve fiber layer; the second convolutional stage extracts the mid-scale relationship between blood vessels and nerve fiber layer texture; and the third convolutional stage extracts a larger range of peripapillary structural patterns. To avoid detail loss due to continuous downsampling, features from different convolutional stages can be processed using... Convolution with a uniform number of channels, then upsampling to the same spatial size and concatenating them, yields a multi-scale basic feature map. .

[0096] 2> Based on the above-generated set of 6 anatomical partition pixels Construct 6 viewport partition masks. Among them, the first... Each disk partition mask is denoted as View disk partition mask It refers to the first A binary space mask corresponding to each anatomical region is used to preserve the feature response of the corresponding anatomical region at the feature map level and suppress the response of other regions.

[0097] In one implementation, the normalized fundus image is first processed... Generate a viewport partition mask on the original spatial dimensions. Then, based on the multi-scale basic feature map Spatial dimensions for the viewport partition mask The nearest neighbor interpolation algorithm is used for scaling to create a scaled viewport partition mask. With multi-scale basic feature maps They have the same height and width. Preferably, using nearest neighbor interpolation can preserve the discrete partitioning meaning of the mask boundaries and avoid partition aliasing caused by interpolation.

[0098] In this implementation, the normalized fundus image The original space dimensions (e.g.) On the image, create a matrix with the same size as the image, with all pixel values ​​initialized to 0. Then, according to step 4 of S201... A collection of pixels from anatomical partitions , to include the elements in the matrix that belong to this set All pixel positions The value is set to 1, that is... At that time, the video disk partition mask The Line 1 Column elements Otherwise, it is 0; generate all in this way. Each video disc partition mask.

[0099] 3> The first Individual View Disk Partition Mask With multi-scale basic feature maps Perform element-wise multiplication to obtain the first... The partition weighted feature map, and the first ... The weighted feature map of each partition is input into the shared residual feature extraction module to obtain the first partition weighted feature map. Partition depth features , No. Partition depth features It refers to the first The structural damage representation of each anatomical region in the depth feature space has a dimension of 256.

[0100] In one implementation, the shared residual feature extraction module can consist of two convolutional kernels with a size of [missing information]. The system consists of convolutional layers, batch normalization layers, ReLU activation layers, residual connections, and global average pooling layers. The six anatomical regions share the same set of residual feature extraction parameters, which maps the local structural responses of different anatomical regions to the same feature space, facilitating subsequent comparison of the contributions of different regions to the judgment of visual function decline. Sharing parameters also reduces the number of trainable parameters and lowers the risk of overfitting under small sample conditions.

[0101] 4> The first Partition depth features With the The untrained local structural descriptors of each anatomical region are input into the region attention module to obtain the first... attention weights for each partition Partition attention weight Used to characterize the The contribution of each anatomical region to the assessment of visual function decline in the current sample is determined by the sum of the attention weights of the six regions, which is 1.

[0102] In the specific implementation, the first Partition depth features With the The untrained local structural descriptors of each anatomical region are input into the region attention module to obtain the region attention weights. The specific structure of the partitioned attention module is as follows: First, it starts from the fundus structure feature vector... Extract the first The local structure descriptors corresponding to each partition, namely the four statistics of average gray intensity, quantile contrast, gray-level co-occurrence matrix contrast, and gray-level co-occurrence matrix entropy, are mapped to a predefined fully connected layer. Vectors of the same dimension Then, and Element-wise addition yields the fused vector. Then The input is fed into an attention scoring layer, which consists of a single fully connected layer and a scalar output, to obtain the first... The original attention scores for each partition Finally, regarding The raw scores of each partition Normalization was performed using the Softmax function to obtain the attention weights for the six partitions. ,satisfy .in, It refers to the first The structural prior mapping vector of each partition is given by the local structural descriptor of that partition (i.e., (A total of 4 statistics) are obtained through a pre-defined fully connected layer, with a dimension of 256. It refers to the first The fusion vector of each partition, with a dimension of 256, is an element-wise fusion representation of deep features and structural priors. It refers to the first The original attention scores for each partition are derived from the fusion vector. The scalar value output by a single-layer fully connected network indicates that the partition is more important to the current evaluation task. This is achieved by using non-trained structural features. By introducing attention computation, the model can use the prior reflectivity and texture statistics of the partitions to help determine which anatomical regions are more important for assessing visual function decline in the current sample, thus improving the rationality and interpretability of attention.

[0103] 5> According to the attention weight of the partition Deep features of 6 partitions Weighted summation is performed to obtain the global retinal structural depth features. ,Right now Global retinal structural depth features This refers to the image feature vector obtained by fusing the depth structural responses of six optic disc anatomical regions, used to characterize normalized fundus images. The overall retinal structural damage pattern in the image is 256-dimensional.

[0104] It should be noted that this step injects anatomical priors into the image encoding process through optic disc partitioning masks and partitioned attention weights. This prevents the model from extracting features indiscriminately from the entire fundus image. Instead, it learns local damage patterns in the lower, upper, nasal, temporal, foveal, and optic disc core regions separately. This enhances the model's attention to sensitive areas of the nerve fiber layer and improves the interpretability of visual function decline assessment results.

[0105] In one embodiment, for example, if a subject shows significant decreases in reflectivity and increases in texture entropy in the lower and upper regions, the non-trained local structural descriptors corresponding to the lower and upper regions will indicate structural abnormalities; the partition attention module will increase the attention weight of the corresponding partition, thereby enhancing the global retinal structural depth features. More attention should be paid to anatomical areas associated with visual function decline.

[0106] In one specific embodiment, dual-channel conditionally gated fusion is used in parallel prediction of deterioration level and retinal age: Since retinal structural damage, individual visual function status, and population decline trends reflect the visual function decline process from different perspectives, simply concatenating global retinal structural depth features, comprehensive visual function impairment index vector, and population-enhanced visual function features makes it difficult for the model to determine which image channels should be enhanced by individual visual function impairment and which should be corrected by the decline trend of similar populations. Therefore, this invention employs a dual-channel conditional gating module to convert the comprehensive visual function impairment index vector and population-enhanced visual function features into channel-level modulation parameters, continuously modulating the global retinal structural depth features. This includes: 1> Input global retinal structure depth features Comprehensive visual impairment index vector and group-enhanced visual function characteristics Among them, global retinal structural depth features The output encoded from the above fundus images has a dimension of 256; the comprehensive visual function impairment index vector. The output, after processing the above visual function data, has 6 dimensions; group-enhanced visual function features. The output, processed from demographic data, has 6 dimensions.

[0107] In practical implementation, the comprehensive visual function impairment index vector is used. and group-enhanced visual function characteristics Before inputting into the dual-channel conditional gating module, the training set mean and standard deviation are used for standardization to avoid a certain principal component dimension or a certain population augmentation dimension having too high a weight in the gating mapping due to its large numerical range.

[0108] 2> Integrate the visual impairment index vector Input an individual conditional mapping network to generate an individual damage scaling vector. and individual damage bias vector Among them, the individual damage scaling vector It refers to the vector of the comprehensive visual impairment index. The generated channel-level scaling parameters are used to adjust the global retinal structure depth features. The response intensity of each channel to the individual's visual function impairment state; individual impairment bias vector. It refers to the vector of the comprehensive visual impairment index. The generated channel-level bias parameters are used to compensate for channel response offsets caused by individual visual function status.

[0109] In one implementation, the individual conditional mapping network adopts a single hidden layer fully connected structure with an input dimension of 6, a hidden layer dimension of 64 or 128, and an output dimension of 512. The first 256 outputs are smoothed and activated to form an individual damage scaling vector. The last 256 outputs form the individual damage bias vector. Individual damage scaling vector The value is preferably limited to the range of 0.5 to 1.5 to avoid excessive amplification or complete suppression of the image channel response; individual damage bias vector You can multiply it by a bias strength coefficient of 0.1 to ensure stability in the early stages of training.

[0110] 3> Enhance visual function characteristics of the group Input a population conditional mapping network to generate a population augmentation scaling vector. and group enhancement bias vector Among them, the population enhancement scaling vector This refers to the enhancement of visual function characteristics by the group. The generated channel-level scaling parameters are used to correct the global retinal structure depth features based on the decay trend of similar populations. Channel response; group enhancement bias vector This refers to the enhancement of visual function characteristics by the group. The generated channel-level bias parameters are used to compensate for the channel response offset caused by the population decay trend.

[0111] In one implementation, the group conditional mapping network and the individual conditional mapping network have the same structure but independent parameters. The purpose of parameter independence is to allow the individual's visual impairment state and the decline trend of similar groups to learn different modulation schemes, avoiding confusion between the two in the same mapping network.

[0112] 4> First use the individual damage scaling vector and individual damage bias vector Global retinal structural depth features Perform the first conditional modulation, then use the population enhancement scaling vector. and group enhancement bias vector Perform the second conditional modulation to obtain multi-source fusion features. Multi-source fusion characteristics It refers to a comprehensive discriminative feature that is simultaneously modulated by fundus structure, individual visual function impairment status, and decline trend in similar populations. It has 256 dimensions and is calculated as follows: ; in, This indicates element-wise multiplication, used to implement channel-level scaling modulation.

[0113] It should be noted that by employing a continuous modulation approach, first targeting individuals and then groups, the multi-task assessment model can first select relevant image channels based on the current visual function impairment status of the subject, and then correct the channel responses based on the decline trends of similar populations, thus enabling multi-source fusion features. It takes into account both individual differences and group patterns.

[0114] 5> Integrate multi-source features By inputting the visual function decline level prediction head and the retinal age regression prediction head respectively, a probability vector of visual function decline level is obtained. and retinal age prediction value Among them, the probability vector of visual function decline level. It is a 4-dimensional probability vector, with four elements corresponding to normal, mild, moderate, and severe retinal deterioration, respectively; retinal age prediction value. It refers to the continuous values ​​output by the retinal age regression prediction head, which are used to characterize the degree of physiological aging of the retina.

[0115] In practical implementation, the visual function decline level prediction head can consist of two fully connected layers and a normalized exponential function, outputting four level probabilities; the retinal age regression prediction head can consist of two fully connected layers, finally outputting a continuous value. Both prediction heads share multi-source fusion features. However, each has independent parameters, enabling classification and regression tasks to share the expression of fundus structural damage while learning their respective task boundaries.

[0116] It should be noted that the comprehensive visual impairment index vector Provides information on individual functional impairment and enhances visual function characteristics in groups. It provides information on the decline trend of similar populations, and the two directly adjust the global retinal structural depth features through channel-level scaling and channel-level bias. This can improve the interpretability and stability of multi-source fusion.

[0117] In one embodiment, as an example, if the comprehensive visual impairment index vector of a sample... The study indicated a significant decrease in contrast sensitivity and stereoscopic vision, while also showing enhanced visual function characteristics in the group. If the sample is in the range of aging and deterioration, the dual-channel conditional gating module will enhance the image channel response related to the weakening of the neural fiber layer reflectivity and the thinning of texture stripes, and suppress the brightness noise channel response that is inconsistent with the current functional state.

[0118] In one specific embodiment, the physiological consistency constraint between the degree of deterioration and retinal age is used in joint training: 1> Based on the probability vector of visual function decline level And true visual function decline level label Calculate the loss by the classification of the recession level Depreciation level classification loss The device is used to monitor the degree of visual function decline and accurately distinguish between four levels: normal, mild decline, moderate decline, and severe decline.

[0119] In practical implementation, cross-entropy loss can be used as the classification loss for descent levels. Optionally, if the number of samples at the four levels in the training set is imbalanced, a class-weighted cross-entropy decay classification loss can be used. This allows decay levels with fewer samples to receive higher loss weights during training. Class weights are calculated solely based on the number of samples for each level in the training set; they are not recalculated during the validation, testing, and inference phases.

[0120] 2> Based on retinal age prediction values and true retinal age label Calculating retinal age-regression loss retinal age regression loss It is used to supervise the estimation of continuous physiological aging by the retinal age regression prediction head.

[0121] In practical implementation, mean squared error loss or mean absolute error loss can be used as the retinal age regression loss. If the training uses the actual retinal age label... For standardization, the inference stage requires using the mean and standard deviation of the training set ages to predict the retinal age. Revert to age units; if age units are used directly during training, the predicted retinal age value will be... The output is the predicted age.

[0122] 3> Construct a set of positive order sample pairs within the same training batch The set of orthogonal sample pairs It refers to a set of sample pairs consisting of samples with higher actual visual function decline levels and samples with lower actual visual function decline levels, used to establish a monotonic constraint between visual function decline level and retinal age.

[0123] In practical implementation, for any two samples in the same training batch and If the first True visual function decline level label for each sample Higher than True visual function decline level label for each sample Then the sample pair Add to the set of ascending sample pairs If there are no sample pairs with different true visual function decline levels within the same training batch, then the physiological consistency ranking loss is not calculated for that training batch; only the decline level classification loss and retinal age regression loss are used to update the model parameters.

[0124] 4> Based on the set of ordinal samples Calculate the physiological consistency ranking loss Physiological consistency ranking loss To constrain the retinal age prediction value of samples with higher actual visual function decline levels to be higher than that of samples with lower actual visual function decline levels, and to at least meet a preset sorting interval, the calculation method is expressed as follows: ; in, Represents the set of ordinal sample pairs Number of sample pairs in the sample; This represents the sorting interval threshold, used to require that the predicted retinal age of samples with higher degradation levels be at least higher than a preset difference. If retinal age is trained in years... 1.0 year old is acceptable; Indicates the first Predicted retinal age for each sample; Indicates the first Predicted retinal age for each sample; This represents the function that takes the maximum of the two values.

[0125] 5> Classify losses by recession level retinal age regression loss And physiological consistency ranking loss Weighted summation yields the total losses from joint training. And based on the total loss of joint training Update the trainable parameters of the multi-task evaluation model and the joint training total loss. The calculation method is expressed as follows: ; in, This represents the weighting coefficient for the retinal age regression loss, used to balance the regression task's contribution to the total loss during joint training. The contribution can be between 0.2 and 1.0, with 0.5 being preferred; This represents the weighting coefficient of the physiological consistency ranking loss, used to balance the ranking constraint in the total loss during joint training. The contribution can be taken as 0.1 to 0.5, with 0.2 being preferred.

[0126] In one implementation, the Adam optimizer can be used to update the trainable parameters of the image base coding module, the shared residual feature extraction module, the partitioned attention module, the dual-channel conditional gating module, the visual function decline level prediction head, and the retinal age regression prediction head. The initial learning rate can be 0.001, the training batch size can be 16 to 64, and the number of training epochs can be 50 to 200. After each training epoch, the macro-average F1 score of visual function decline level, the mean absolute error of retinal age, and the proportion of physiological consistency violations are calculated on the validation set, and the model parameters with the best overall validation performance are saved.

[0127] It should be noted that this step establishes monotonic constraints on ordered deterioration levels and continuous retinal ages through physiological consistency ranking loss. This ensures that the multi-task assessment model not only optimizes the prediction error of individual tasks but also guarantees that the multi-task outputs conform to the physiological laws of visual function decline. This reduces inconsistent outputs such as "severe deterioration but younger retinal age" or "normal sample but abnormally high retinal age," thereby improving the clinical rationality of the assessment results.

[0128] In one embodiment, for example, if the first The true visual function decline level of the sample was severe decline. If the true visual function decline level of a sample is normal, then Forming a positive-order sample pair. If the model predicts the... The predicted retinal age for the first sample was 66.0 years. The predicted retinal age for each sample was 65.5 years, and the sorting interval threshold was [not specified]. If the age difference is only 0.5 years, it does not meet the requirement of being at least 1.0 year older. Therefore, the physiological consistency ranking loss will penalize this, pushing the model to improve its ranking in subsequent training. The sample relative to the first Differences in retinal age prediction among samples.

[0129] In one specific embodiment, histograms were used to compare the numerical distributions of global retinal structural depth features (dependent solely on fundus images) and multi-source fusion features modulated by dual-channel conditional gating. Figure 3 As shown, the mean and variance of the modulated feature distribution changed significantly, indicating that the combined visual impairment index and population enhancement features successfully adjusted the response intensity and pattern of the image features. The horizontal axis of both subplots represents the feature values ​​(dimensionless), and the vertical axis represents the frequency (i.e., the number of times the feature value occurs, a dimensionless integer).

[0130] In one specific embodiment, after completing the training of the multi-task assessment model, the trained processing parameters and model parameters can be deployed to the visual function decline auxiliary assessment system. During online operation, the system acquires the raw fundus images, raw visual function test data, psychological cognitive scale data, and demographic data of the subjects to be assessed according to the same data acquisition specifications as the training phase, and outputs the visual function decline level, retinal age prediction value, confidence cue, and anatomical region contribution information. The specific steps are as follows: 1> Obtain raw fundus images, raw visual function test data, psychological cognitive scale data, and demographic data of the subjects to be evaluated to form a multi-source visual function sample to be evaluated. Multi-source visual function samples to be evaluated This refers to the multi-source input samples that currently require auxiliary assessment of visual function decline, containing the same type of data as in the training phase.

[0131] In practical implementation, if the multi-source visual function samples to be evaluated If a score on a certain cognitive scale is missing, it can be imputed using the median of the corresponding scale in the training set, and the scale should be marked as missing in the system output. If a key visual function detection dimension is missing, the system should prompt for supplementary detection, or only output low-completeness assessment results related to fundus images. This can prevent missing data from being misinterpreted as normal data.

[0132] 2> Samples for evaluating multi-source visual function Perform the same non-training prior feature construction process as the training phase to obtain the feature vector of the fundus structure to be evaluated. Vector of the comprehensive visual impairment index to be evaluated Enhanced visual function characteristics of the group to be evaluated .

[0133] Specifically, the original fundus images to be evaluated are subjected to size unification, pixel normalization, optic disc localization, anatomical partition generation, and partition texture statistics to obtain the feature vector of the fundus structure to be evaluated. The process involves standardizing the direction of the original visual function test data to be evaluated, standardizing the training set statistics, applying psychological confidence correction, and performing fixed principal component transformation to obtain the comprehensive visual function impairment index vector to be evaluated. Based on the age and body mass index of the subjects to be evaluated, demographic neighborhood samples are searched in the training set, and similarity-weighted propagation is performed to obtain the enhanced visual function characteristics of the group to be evaluated. .

[0134] It should be noted that the online inference phase must use the standardized parameters, psychological sensitivity weights, principal component transformation parameters, demographic graph propagation parameters, and training set neighborhood samples saved during the training phase. These parameters must not be recalculated separately based on the samples to be evaluated, otherwise the meaning of the features will be changed and the stability of the evaluation will be affected.

[0135] 3> Combine the normalized fundus image to be evaluated with the feature vector of the fundus structure to be evaluated. Vector of the comprehensive visual impairment index to be evaluated Enhanced visual function characteristics of the group to be evaluated Input the trained multi-task evaluation model to obtain the probability vector of the visual function decline level to be evaluated. and the predicted retinal age to be evaluated .

[0136] Probability vector of visual function decline level to be assessed This refers to the predicted probability that the subject being evaluated belongs to one of four categories: normal, mild, moderate, or severe retinal age; the predicted retinal age value to be evaluated. This refers to the predicted physiological retinal age of the subject being evaluated. If the retinal age is standardized during the training phase, the predicted retinal age of the subject being evaluated will be calculated using the mean and standard deviation of the age saved during the training phase before output. Restore to the unit of age.

[0137] 4> Based on the probability vector of the visual function decline level to be assessed Output the visual function degradation level. The probability of the highest level is denoted as... , Represents the probability vector of the degree of visual function decline to be assessed. The highest probability value in the range; the level corresponding to the highest probability is used as the initial visual function decline level result. If Less than the confidence threshold If so, output a low confidence warning; if Not less than the confidence threshold Then, the visual function degradation level corresponding to the highest probability is output. Confidence threshold This represents the minimum probability threshold for accepting the model's grade judgment, which can be between 0.6 and 0.8, with 0.7 being the preferred value.

[0138] In one embodiment, as an example, if the probability vector of the visual function decline level to be evaluated is... And confidence threshold The maximum probability is 0.62, which is less than 0.7. The system outputs "Moderate tendency to decline but insufficient confidence; it is recommended to combine with retesting or clinical examination for confirmation"; if the probability vector of the visual function decline level to be evaluated is... And confidence threshold The maximum probability is 0.67, corresponding to severe degradation. The system outputs the degradation level result as severe degradation.

[0139] 5> Output anatomical region contribution information based on the region attention weight. Anatomical region contribution information refers to the degree of attention the multi-task assessment model pays to the lower region, upper region, nasal region, temporal region, foveal region, and optic disc core region when assessing visual function decline. It is used to help interpret the assessment results.

[0140] In practical implementation, the attention weights of the six partitions obtained above are output in a fixed partition order, and can be further combined with the fundus structure feature vector. The system generates textual prompts based on the regional reflectance and texture statistics. For example, when the regional attention weights of the lower and upper regions are high, and the average gray intensity of the corresponding regions decreases while the gray-level co-occurrence matrix entropy increases, the system can prompt that "the structural features related to the nerve fiber layers above and below the optic disc contribute significantly to the current assessment." This explanatory information does not replace clinical diagnosis but can help physicians understand the areas of focus of the model.

[0141] 6. Perform a time consistency check on the results of multiple follow-up visits for the same subject. If the system receives continuous follow-up data, perform a trend check on the visual function decline level and retinal age prediction value at different time points for the same subject. If the visual function decline level gradually increases and the retinal age prediction value increases synchronously during continuous follow-up, a decline progression indicator will be output; if a certain assessment result is significantly inconsistent with the results at previous and subsequent time points, and the psychological confidence coefficient or the confidence level of the grade prediction is low, it is suggested that the test may be affected by psychological cognitive state or test cooperation, and a retest is recommended.

[0142] The present invention also discloses a computer program product or system, including a computer program that, when executed by a processor, implements the above-described method steps.

[0143] The visual function decline quantitative assessment system provided in this embodiment of the invention specifically includes: Acquisition Unit: Acquires fundus images, visual function data, and psychological cognitive data of the tested sample; First feature unit: Divide the fundus image into regions and calculate the grayscale features of each region; use the fundus image and the grayscale features of each region to perform feature encoding to obtain the global retinal structure features; Correction Unit: Generates psychological confidence coefficients from psychological cognitive data, and uses these psychological confidence coefficients to correct visual function data to obtain corrected visual function data; The second feature unit: The comprehensive visual function impairment index vector is calculated from the corrected visual function data through principal component transformation; Fusion unit: Fusion features are obtained by modulating global retinal structural features using a comprehensive visual impairment index vector; Assessment Unit: The assessment results are obtained by evaluating the degree of visual function decline and / or retinal age through fusion features.

[0144] An embodiment of the present invention provides a computer device, specifically comprising: A memory and a processor; the memory is used to store program instructions; the processor is used to invoke the program instructions, when any of the above-described methods for quantitative assessment of visual function decline are executed.

[0145] The present invention also discloses a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, provides any of the above-described methods for quantitative assessment of visual function decline.

[0146] The verification results of this verification embodiment show that assigning inherent weights to indications can improve the performance of this method compared to the default settings. Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for example, the division of units is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separated; the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of this embodiment. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0147] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0148] The computer device provided by the present invention has been described in detail above. For those skilled in the art, there will be changes in the specific implementation and application scope based on the ideas of the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for quantitatively assessing visual function decline, characterized in that, include: Acquire fundus images, visual function data, and psychological cognitive data of the tested samples; The fundus image is divided into regions and the grayscale features of each region are calculated; Global retinal structural features are obtained by feature encoding using fundus images and grayscale features of each region; The feature encoding is: generating a mask for p optic disc partitions by partitioning the fundus image, where p is a natural number greater than 1; Multi-scale feature extraction is performed on fundus images to obtain multi-scale feature maps; the mask of each optic disc region is multiplied element-wise with the multi-scale feature map to obtain p weighted feature maps of the regions; Feature extraction is performed on the weighted feature maps of p partitions to obtain the depth features of p partitions; the depth features of p partitions are fused with the grayscale features of the corresponding partitions, and then attention is calculated to obtain the attention weight of each partition; The global retinal structure features are obtained by weighted summation of the depth features of p partitions and the attention weights. Gray-level features include any one or more of the following: average gray-level intensity, quantile contrast, gray-level co-occurrence matrix contrast, and gray-level co-occurrence matrix entropy. Psychological confidence coefficients are generated from psychological cognitive data, and visual function data is corrected using these psychological confidence coefficients to obtain corrected visual function data. The corrected visual function data are used to calculate the comprehensive visual function impairment index vector through principal component transformation. The fusion feature is obtained by modulating the global retinal structural features using the comprehensive visual function impairment index vector. The modulation is achieved by generating an individual impairment scaling vector and an individual impairment bias vector using the comprehensive visual function impairment index vector. The individual impairment scaling vector is multiplied by the global retinal structural features and then added to the individual impairment bias vector to obtain the fusion feature. The generation is achieved by inputting the comprehensive visual function loss index vector into a mapping network with hidden layers and full connections. The assessment results are obtained by evaluating the degree of visual function decline and / or retinal age through fusion features.

2. The method for quantitatively assessing visual function decline according to claim 1, characterized in that, The principal component transformation calculation involves centering the corrected visual function data, projecting it onto the principal component space, and then dividing each element by the whitening scaling vector to obtain the comprehensive visual function impairment index vector. The whitening scaling vector is constructed from the covariance matrix of the corrected visual function data in the sample set.

3. The method for quantitatively assessing visual function decline according to claim 1, characterized in that, The method further includes group-enhanced visual function features, obtaining demographic data of the tested samples, screening a similar sample set using the demographic data of the tested samples, calculating the similarity between the samples in the sample set and the tested samples, weighting the comprehensive visual function impairment index vector of the samples in the sample set with the similarity to obtain a domain visual function vector, fusing the domain visual function vector with the comprehensive visual function impairment index vector of the tested samples to obtain group-enhanced visual function features; modulating global retinal structural features with the comprehensive visual function impairment index vector and group-enhanced visual function features to obtain a first fusion feature, and using the first fusion feature to assess the degree of visual function decline and / or retinal age to obtain an assessment result; the similarity of the samples is calculated using age and body mass index.

4. The method for quantitatively assessing visual function decline according to claim 1, characterized in that, The psychological cognitive data includes any two or more of the following: depression scale scores, anxiety scale scores, and Mini-Cognitive Scale scores. The composite score is obtained by normalizing any two or more of the following scores, then weighting and summing them, and finally converting the composite score using a smoothing decreasing function to obtain the psychological confidence coefficient.

5. The method for quantitatively assessing visual function decline according to claim 1, characterized in that, The partitioning process includes: extracting brightness from the fundus image according to the three channels (red, green, and blue) and then performing weighted fusion to obtain a single-channel brightness map; calculating the geometric features of each bright region in the single-channel brightness map; filtering candidate regions for optic disc features based on the geometric features; fitting the candidate regions to circles or ellipses and calculating the optic disc center and radius; using the optic disc center as the origin of polar coordinates, calculating the radial distance and angular direction of each pixel position relative to the optic disc center; and dividing the region based on the radial distance, angular direction, and optic disc radius; the geometric features include any one or more of the following: area, circularity, and vascular convergence features.

6. A computer program product comprising a computer program or instructions, characterized in that, The computer program or instructions are executed by a processor to implement the visual function decline quantitative assessment method according to any one of claims 1-5.

7. A computer device comprising a memory, a processor, and a computer program or instructions stored in the memory, characterized in that, The computer program or instructions are executed by a processor to implement the visual function decline quantitative assessment method according to any one of claims 1-5.

8. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, The computer program or instructions are executed by a processor to implement the visual function decline quantitative assessment method according to any one of claims 1-5.

Images