Optic nerve multi-modal image prognostic marker extraction method

Abstract

The present application relates to a method for extracting optic nerve multimodal image prognostic markers, obtaining optical coherence tomography, magnetic resonance imaging and retinal electrophysiological data; based on the signal-to-noise ratio of the region, the signal-to-noise ratio of the mode is calculated, the signal cooperative change region is identified according to the multimodal noise non-uniform fitting error function, and the signal stable coupling region is constructed; the mode structure consistency deviation rate is calculated, the local region of the abnormal split of the cross-modal structure is identified, and is mapped to the original data of the corresponding mode; the time point difference spectrum deflection degree of the follow-up image sequence of the same subject is calculated, the nonlinear mutation time point is identified, the candidate region of the prognosis event is determined; the cross entropy approximation value of the response of each mode is calculated, the mode redundancy penalty index and the marker exclusive information density are determined, and the marker feature point with low redundancy and high contribution degree is screened; the posterior condition transition probability is calculated, the potential correlation path of the marker is generated, and the structure degeneration direction and the prognosis risk grade are determined.

Description

Technical Field

[0001] This invention relates to the field of optic nerve imaging technology, specifically to a method for extracting prognostic biomarkers from multimodal optic nerve imaging. Background Technology

[0002] As disclosed in Chinese patent document CN117426748A, "A Method for Detecting MCI Based on Multimodal Retinal Imaging," current methods for extracting prognostic biomarkers from multimodal optic nerve imaging still face numerous technical bottlenecks and shortcomings in application. These shortcomings are mainly reflected in the following aspects: First, existing methods generally focus on improving classification performance. For example, this method classifies MCI, AD, and normal cognitive individuals by constructing a dual-stream attention neural network model. While this has improved prediction accuracy to some extent, its core objective remains static pattern recognition, lacking the ability to model and predict the long-term evolution of optic nerve structures. This makes it difficult to support early intervention and dynamic tracking management of individualized optic nerve degenerative changes. In contrast, optic nerve diseases such as glaucoma and optic atrophy exhibit significant temporal progression characteristics. If their subtle evolutionary trajectories and mutation nodes cannot be captured, it is difficult to provide early warnings of prognostic risks, limiting the depth of application of multimodal imaging in clinical prognostic analysis.

[0003] Secondly, although this method proposes a multimodal fusion mechanism and introduces an attention mechanism to enhance the feature interaction between OCT and fundus images, its fusion strategy is mainly based on the explicit feature stacking and channel weighting of deep networks. It does not analyze the cooperative stability and structural heterogeneity between different modalities at the signal ontology level, and lacks systematic modeling of underlying parameters such as signal quality drift, energy attenuation consistency, and structural coupling regions. As a result, the obtained fused features lack interpretability and localization ability in a biophysical sense, making it difficult to accurately anchor key prognostic regions. Thirdly, the existing method has not established a measurement system for the distribution of redundancy-exclusive information between modalities. When identifying key regions, this method uses class activation mapping for visual localization and relies on gradient information from intermediate layers of the neural network to generate weighted heatmaps. Although this method can improve image interpretability, it does not conduct causal analysis and information separation evaluation of the contributions of different modalities. It cannot quantify whether there is a truly unique important structural feature in a certain modality that affects classification judgment, thus failing to achieve specific enhancement of marker screening and effective removal of noise redundancy. Fourth, the annotation and training methods in this approach are highly dependent on cognitive test scores, falling under the category of label-supervised methods. This makes it extremely sensitive to the size of the training sample and the quality of the labels, and it suffers from poor generalization ability across devices and institutions. Especially in optic neuropathy scenarios, actual prognostic labels are often difficult to obtain in a timely manner, leading to difficulties in large-scale data modeling. Therefore, there is a lack of a mechanism that can automatically extract unsupervised or weakly supervised potential prognostic features from the internal variation patterns of image data.

[0004] Fifth, regarding temporal modeling, this method does not involve the processing of multi-period image data, relying solely on static features at a single time point for judgment, thus ignoring the nonlinear progression of optic nerve structures. Sixth, the existing scheme also lacks sufficient accuracy in spatial localization. The class activation maps extracted by this method are limited by the backprojection capability of deep semantic features of the network, making it difficult to accurately locate the structural details of the original image. Summary of the Invention

[0005] The purpose of this invention is to provide a method for extracting prognostic biomarkers from multimodal optic nerve imaging, thereby addressing some of the drawbacks and shortcomings mentioned in the background art.

[0006] The present invention adopts the following technical solution to solve the above-mentioned technical problems: a method for extracting prognostic biomarkers of optic nerve multimodal imaging, comprising: acquiring optical coherence tomography, magnetic resonance imaging and retinal electrophysiological data; calculating the intermodal signal-to-noise ratio change rate based on regional signal-to-noise ratio covariance, identifying signal covariance regions according to the multimodal noise non-mean fitting error function, and constructing a signal stable coupling region; For the stable coupling region of the signal, the modal structure consistency deviation rate is calculated, local regions of cross-modal structural abnormal splitting are identified, and the regions are mapped to the corresponding modal raw data and labeled as prognostic sensitive regions; for the follow-up image sequence of the same subject, the time point difference spectrum skewness is calculated, nonlinear mutation time points are identified, and candidate regions of prognostic events are determined; For each candidate region of the prognostic event, the cross-entropy approximation value of each modal response is calculated, the modal redundancy penalty index and the exclusive information density of the marker are determined, and the marker feature points with low redundancy and high contribution are screened; the posterior conditional transition probability is calculated based on the spatial distribution and variation characteristics of the marker feature points, the potential association path of the marker is generated, and the direction of structural degradation and the prognostic risk level are determined.

[0007] Furthermore, the calculation of regional signal-to-noise ratio covariance includes establishing a local signal stability weight field and separating the noise source from the structural signal through covariance directionality analysis of the change trend of the neighboring signal; the determination of the signal stable coupling region includes performing differential filtering analysis on the energy distribution of the multimodal signal to identify regions with the same energy attenuation trajectory.

[0008] Furthermore, before calculating the modal structure consistency deviation rate, structural domain boundary tensor decomposition is performed to extract the main components of cross-modal structural deformation; the identification of cross-modal abnormal splitting regions includes the introduction of hierarchical texture coupling analysis to detect the depth distribution pattern of structural fractures between modes.

[0009] Furthermore, when mapping the abnormal split region, reverse resampling is performed based on the spatiotemporal registration residual sequence to reduce the impact of local geometric distortion on the region mapping; the calibration of the prognostic sensitive region is performed by evaluating the region self-stabilization index, which is based on the joint determination of temporal domain intensity fluctuation and spatial domain texture perturbation.

[0010] Furthermore, the calculation of the time difference spectrum skew includes constructing a sliding window cluster of changing trends to distinguish between instantaneous noise changes and real structural abrupt changes; when identifying nonlinear abrupt change time points, a cross-modal time alignment mechanism is introduced to correct the impact of different modal acquisition delays on the time difference calculation.

[0011] Furthermore, a modal confidence constraint matrix is ​​introduced during the calculation of the cross-entropy approximation to achieve unbalanced weighting in the multimodal prediction distribution; the modal redundancy penalty index is calculated by performing multi-scale mutual information decomposition to extract the components of independent information gain across modalities.

[0012] Furthermore, the calculation of the exclusive information density of the marker includes establishing a feature contribution ranking sequence among modes and determining effective feature points through information stability criteria; when generating the potential association path, the sequential similarity mapping method is used to establish the coupling relationship between the temporal evolution sequence and the spatial variation trajectory; and local convolution smoothing of the path vector field is performed during the determination of the structural degradation direction.

[0013] Furthermore, the establishment of the feature contribution ranking sequence includes performing multi-level feature resolution segmentation processing and determining the ranking weight based on the feature stability distribution at different spatial scales; the determination of the information stability criterion adopts a time-staggered stability comparison mechanism, which identifies feature points with long-term stable expression by calculating the response duration rate of feature points in different time windows.

[0014] Furthermore, when establishing the coupling relationship between the temporal evolution sequence and the spatial variation trajectory, the distance matrix of the corresponding node in the spatial domain is updated in reverse after the temporal similarity converges, so as to enhance the path continuity.

[0015] Furthermore, the multimodal noise non-mean fitting error function is a residual evaluation function obtained by weighted summation of the difference between the non-mean deviation of the signal intensity of each mode in the same region and its covariance weight, which is used to quantify the degree of signal co-change between modes; The multimodal noise non-mean fitting error function is defined in a weighted integral form and is used to quantitatively measure the signal stability and co-variation degree of different modal images in the same region. Its expression is: in: This is the overall energy function of the non-mean fitting error for multimodal noise, used to measure the degree of cooperative deviation between signals; For the local spatial region involved in the calculation; The number of image modalities involved in the calculation; For the first The spatial location of each modality The signal strength; A non-mean reference function for all modal signals within the same region, used to reflect the overall signal distribution trend; For the first The signal confidence weights for each mode are adaptively determined based on the mode noise variance. For the first Each mode in the region The covariance response function within; This is a nonlinear amplification factor used to enhance the contribution of high-deviation regions to the error; It is a covariance balancing factor used to control the strength of the suppression of the overall error by the gradient changes between modes.

[0016] The beneficial effects of this invention are as follows: The method for extracting prognostic biomarkers from optic nerve multimodal images proposed in this invention, by introducing regional signal-to-noise ratio covariance analysis and a multimodal noise mean fitting error function, achieves quantitative calculation of the signal coordination relationship between different image modalities. It can automatically identify stable signal coupling regions under high noise conditions, effectively distinguishing structural information from artifact noise, thereby significantly improving the accuracy of multimodal image fusion. Compared with traditional fusion methods based on registration or pixel averaging, this method, through covariance directionality analysis and energy attenuation differential filtering mechanism, ensures the consistency of signal expression within the same structural domain in cross-modal data, enabling reliable extraction of early, subtle pathological changes.

[0017] Furthermore, this invention proposes a comprehensive judgment mechanism for modal structural consistency deviation rate and time point difference spectrum skewness in the structural anomaly identification and prognostic inference stages. This mechanism can accurately identify nonlinear mutation time points during follow-up and establish a coupling relationship between temporal evolution and spatial variation, thereby achieving dynamic tracking of the optic nerve degeneration process. Attached Figure Description

[0018] Figure 1 This is a flowchart of the multimodal image prognostic biomarker extraction process of the present invention.

[0019] Figure 2 This is a diagram showing the relationship between the stability of local signals in multimodal images and the structural anomaly detection function of this invention.

[0020] Figure 3 This is a functional diagram illustrating the screening and pathway relationships of multimodal optic nerve imaging biomarkers in this invention.

[0021] Figure 4 This is a flowchart of the multimodal optic nerve case analysis in Embodiment 1 of the present invention.

[0022] Figure 5 This is a flowchart of the multimodal optic nerve imaging biomarker path and trend modeling process in Embodiment 2 of the present invention. Detailed Implementation

[0023] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] Combined with appendix Figure 1 This invention discloses a method for extracting prognostic biomarkers from multimodal optic nerve imaging. The method acquires multimodal optic nerve imaging data from subjects, including optical coherence tomography (OCT) images primarily focused on the eye, magnetic resonance imaging (MRI) images primarily focusing on the optic nerve pathway, and retinal electrophysiological response data reflecting neural function transmission characteristics. These three types of data provide multi-source information on the optic nerve state from structural, pathway, and functional dimensions, respectively. After acquiring the original images, the optic nerve regions in each modality are registered and aligned, and local signal blocks are established within the same spatial region. Statistical analysis is used to obtain the signal-to-noise ratio variation of each modality within this region and to calculate its covariance matrix. The signal stability of the region is judged based on the consistency of the covariance direction. Using the established covariance information, combined with a non-mean fitting error function, residual calculations are performed to assess the degree of deviation of the multimodal signals. The signal intensity of each modality within the local region is compared with the overall non-mean reference signal, and integral weighting is performed using modal weights and covariance deviation terms to finally form a residual energy distribution function, which is used to measure the degree of co-variance of different modal signals within the region. When the residual value is lower than the set threshold, the region is determined to be a region of multimodal signal co-change, which has high data consistency and structural expression stability. Therefore, the region is defined as a signal stable coupling region and used as a candidate region input for subsequent feature extraction and marker screening.

[0025] After constructing the signal-stable coupling region, multimodal structural consistency analysis is performed on this region. By calculating the differences in structural parameters of each mode at the same spatial location, a modal structural consistency deviation rate is formed, which measures the degree of consistency in structural representation among multiple modes. To improve the accuracy of the deviation rate calculation, tensor decomposition is performed on the structural domain boundary information in each modal image before performing consistency deviation analysis. This extracts the structural deformation components in the dominant direction to eliminate interference from non-critical regions on the overall morphological differences. After obtaining the deviation rate data, threshold comparison is performed on each region to identify local regions where the deviation of modal structural parameters significantly exceeds the average range. These regions are defined as cross-modal structural anomalous split regions. The identified split regions are back-mapped to the original data layer of each modality using spatial indexing, and reverse resampling is performed using registration residuals to improve the localization accuracy of anomalous regions on the original image. Subsequently, based on the time dimension, structural parameter sequences were extracted from the follow-up images of the same subject at multiple time points to construct a structural change trend map. The slope and offset direction of each structural region on the time axis were calculated. The time point difference spectrum skewness index was used to identify the location of abrupt trend changes. Time points with drastic slope changes and obvious disruption of continuity were identified as nonlinear abrupt change points. The spatial region corresponding to this time point was then used as a potential prognostic event candidate region for subsequent biomarker screening and risk prediction analysis.

[0026] Information content analysis was performed on the response signals of each imaging modality within the candidate prognostic event regions. Cross-entropy approximations were calculated to measure the degree of deviation in prediction results for the same lesion region among different modalities. The cross-entropy approximation constructs an error term by comparing the dispersion of the modal response distribution with the multimodal joint response benchmark, revealing potential information redundancy between modalities. Combined with the cross-entropy results, a modal redundancy penalty index was introduced to quantify the non-independence of each modality in feature representation; a higher index value indicates stronger repetition among modalities and weaker expressive power. To eliminate modal overlap interference and improve feature recognition accuracy, a biomarker-specific information density index was constructed based on contribution ranking results and the stability of modal information distribution. All feature points were ranked and screened, retaining a set of feature points with high information independence and strong contribution to form high-confidence candidate biomarkers. Subsequently, based on the spatial distribution patterns and structural variation attributes of the selected biomarker feature points, a conditional transition probability assessment method was used to calculate the correlation strength between each biomarker during the pathological evolution process. Potential correlation paths were generated using a maximum a posteriori (MAP) transition path identification algorithm, and the direction of optic nerve degeneration trend was constructed using the principal vector of structural directional changes within these paths. Finally, based on the cumulative transition probability, spatial spread range, and number of activated biomarkers within the paths, the prognostic risk level of the corresponding region was comprehensively assessed, achieving visualization of structural degradation trends and individualized risk quantification output.

[0027] Combined with appendix Figure 2 For each modal image, a local signal stability weight field is constructed. This weight field uses pixel neighborhoods as basic units. By calculating the mean shift of signal intensity within this neighborhood and the difference in response across consecutive frames, a stability index reflecting the temporal consistency and spatial smoothness of the signal is generated. Based on this index, stable signal regions in each modal image are assigned higher weights, while noisy or highly volatile regions are assigned lower weights, forming a differentiated weighted distribution. Subsequently, based on this weight field, a covariance directionality analysis method is used to construct tensors for the change directions of each modal signal within the neighborhood and extract the principal direction components. The consistency of the principal directions across different modalities is compared, thereby achieving decoupled identification of structural changes and random noise. When a region exhibits consistent principal covariance directions across multiple modalities and the corresponding region has high stability weights, this region is identified as a region dominated by structural signals. To further optimize the signal coupling region identification results, differential filtering was applied to the energy distribution characteristics of the multimodal images. The energy attenuation rate of image blocks was calculated within a fixed window scale, and the consistency of the attenuation trend direction was used as a joint screening criterion. Finally, regions exhibiting similar signal energy decrease trajectories across different modalities were extracted. These regions were identified as stable signal coupling regions and can serve as reliable image region inputs for subsequent structural analysis and marker extraction.

[0028] Before performing modal structural consistency analysis, high-dimensional feature modeling is performed on the structural boundary information of the target region in the multimodal image. By constructing a structural domain boundary tensor and performing decomposition, the principal direction components with the most significant deformation between modes are extracted to accurately characterize the local structural deformation trend across modes. The tensor decomposition process is based on the gradient change matrix of each mode in the edge region. Stable and spatially descriptive structural principal axes are extracted through singular direction tensor reconstruction, thus providing a high-precision basic representation for subsequent consistency calculations. After obtaining the principal component representations, the structural consistency deviation rate is calculated by comparing the curvature of the structural distribution and the degree of boundary diffusion in the multimodal image in this region. Regions exhibiting discontinuities or misalignments in multiple modes are selected as potential structural anomaly splitting regions. To further enhance the deep feature representation of split recognition, hierarchical texture coupling analysis was performed on the above-mentioned regions in combination with local multi-level texture features. The coupling degree of different modal images in terms of gray-level directionality, edge complexity, and frequency domain detail distribution was extracted, and a texture alignment tensor map was constructed. Through this map, the deep distribution pattern of abnormal regions was modeled, realizing the accurate localization and multi-scale verification of potential structural rupture regions between modalities. This ensures that the extracted split regions have a real pathological basis and provides a structural basis for subsequent prognostic sensitive area labeling and marker extraction.

[0029] Because the imaging mechanisms of multimodal images differ during acquisition, nonlinear distortions in local scale, deformation, or tissue boundaries are easily caused. Therefore, a spatiotemporal registration residual sequence is introduced as a relocation reference during region mapping. By acquiring the temporal transformation parameters and spatial alignment error terms of the registered images between each modality, a residual sequence containing positional deviation and deformation residuals is constructed. Based on this sequence, reverse resampling is performed within the target region to accurately reconstruct the true boundaries and structural positions of the abnormal region from the original modal image, thereby reducing the mapping offset caused by registration errors and improving the stability and accuracy of abnormal structure mapping. After mapping, the prognostic sensitivity of the located abnormal regions is further evaluated. For this purpose, a region self-stabilization index is introduced as a sensitivity criterion, which considers both the signal intensity fluctuation amplitude in the temporal dimension and the texture perturbation changes in the spatial dimension. By analyzing the signal change curves and edge texture complexity changes of the region in continuous follow-up images, regions with large intensity fluctuations and frequent texture perturbations are assigned higher instability scores. Finally, regions with a self-stabilization index below a set threshold and exhibiting a tendency for pathological variation were identified as prognostic sensitive regions and used as the core input regions for subsequent feature extraction and risk prediction.

[0030] To accurately identify abrupt changes in optic nerve structures during follow-up, structural parameters were extracted and temporally sequenced from multimodal image sequences acquired at multiple time points from the same subject. In the temporal dimension, a time-point difference spectrum was constructed by calculating the trend curves of structural indices over time. Based on this spectrum, the spectral skewness of the time-point difference was calculated to quantify the changing trend of structural evolution rate across different time periods. To avoid interference from transient abnormal fluctuations or imaging artifacts in the overall assessment, a sliding window clustering mechanism was introduced during the calculation process. The entire follow-up data was divided into multiple time window sets with overlapping ranges, and trend fitting and slope analysis were performed separately within each window. By comparing the differences in slope changes between adjacent windows, it was determined whether the change belonged to a continuous trend evolution or an abrupt change process. Through significance analysis of the slope differences between windows, it was possible to effectively distinguish between false positive fluctuations caused by transient noise and temporal abrupt changes caused by genuine structural variations. To improve the accuracy of cross-modal temporal comparison, a cross-modal temporal alignment mechanism is introduced. By comparing the structural consistency time points of adjacent images between different modalities, a temporal alignment mapping table is constructed. Temporal remapping is then performed based on the differences in modal acquisition timestamps and the similarity of their corresponding structural evolution states. This eliminates the temporal offset error caused by asynchronous modal acquisition and ensures that the calculation of the difference spectrum skewness at all time points is completed under a unified time reference.

[0031] When performing feature analysis on candidate regions for prognostic events, the predicted response distribution of each modality in that region is calculated separately, and a multimodal prediction distribution set is constructed. Subsequently, a cross-entropy approximation is calculated based on this set to evaluate the degree of difference between the prediction results of each modality and the strength of deviation relative to the joint prediction distribution. To overcome the differences in prediction reliability among different modalities caused by variations in signal-to-noise characteristics, data integrity, or diagnostic sensitivity, a modality confidence constraint matrix is ​​introduced during the cross-entropy calculation. This matrix generates a set of modality weight parameters based on the stability performance of each modality during training, signal intensity distribution, and consistency in fitting prior labels. These parameters are then applied to the cross-entropy loss term to achieve unbalanced weighting of the multimodal prediction results. Modality predictions with high confidence have a greater influence on the overall difference calculation, while the deviation contribution of low-confidence modalities is suppressed, thus effectively avoiding prediction misleading caused by low-quality modalities.

[0032] After weighted cross-entropy evaluation, a modal redundancy penalty index is calculated to further eliminate redundant expressions between modalities and extract feature components with independent value. In this calculation, a multi-scale mutual information decomposition method is employed to jointly model the information flow features of different modalities in the candidate region. By constructing multi-scale image patches and their corresponding statistical feature spaces, frequency domain decomposition and directional component extraction are performed on the mutual information intensity between modalities to identify the proportion of shared information and the distribution density of independent gain information. Based on mutual information decomposition, the redundancy degree of each modality is quantified, a redundancy penalty factor is constructed, and combined with the original predicted information intensity. This factor is used for weight suppression and exclusive contribution calculation in the biomarker screening stage, thereby prioritizing the retention of low-redundancy, high-independent feature points and ensuring that subsequently extracted biomarkers have higher specificity and diagnostic value.

[0033] Combined with appendix Figure 3To extract high-predictive-value prognostic biomarkers for the optic nerve from multimodal image data, the performance intensity of all candidate feature points in each modality was analyzed. The contribution of each modality to the predicted target within the candidate region was calculated, and a feature contribution ranking sequence was established across modalities. This ranking sequence is arranged from high to low contribution, and the ranking priority is dynamically updated based on intermodal expression differences. Based on the ranking, an information stability criterion is introduced to screen the effectiveness of feature points. This criterion consists of two dimensions: the response stability of feature points in multi-time-point images and the coherence of signal changes in their neighborhood regions within the spatial domain. Feature points that maintain stable expression in both time and space dimensions are determined through a comprehensive stability score and selected as the final set of biomarkers with high exclusive information density. After obtaining the biomarker set, to further identify their evolutionary relationships in the process of optic nerve structural degeneration, potential association paths are generated using a sequence similarity mapping method. This method constructs a similarity matrix based on the activation order of feature points in a time series, combining spatial distance, response similarity, and structural coherence. Through dynamic matching and the principle of shortest path, it forms a path structure that couples the temporal evolution sequence with the spatial variation trajectory. To enhance the continuity and directional consistency of the path structure, a local convolution smoothing operation is performed on the path vector field after path generation. By constructing a Gaussian kernel based on spatial neighborhood, the direction of local path vectors is adjusted and mutations are suppressed, making the overall path more continuous, smooth, and consistent with the physiological structural orientation. Ultimately, this method is used to determine the dominant direction of optic nerve degeneration and provide quantifiable structural evidence for prognostic risk assessment.

[0034] By segmenting the original image data into multiple levels according to spatial resolution, the response stability of feature points at each spatial scale is calculated, and a ranking weight system based on the feature stability distribution is constructed. This system reflects the contribution of different modalities to candidate biomarker features at different spatial resolutions, thereby establishing a ranking sequence of feature contributions among modalities as the basis for subsequent information screening. After ranking, an information stability criterion screening process is performed. This process uses a time-staggered stability comparison mechanism, dividing the multi-time point image sequence into multiple staggered time windows. The long-term expression consistency of each candidate feature point is evaluated by statistically analyzing the response persistence rate within different time windows. If a feature point consistently exhibits a high response or significant structural change signal within multiple staggered time windows, the point is considered to have long-term stable expression ability and is thus marked as a valid biomarker feature point.

[0035] To accurately establish the coupling relationship between the temporal evolution and spatial variation trajectories of biomarkers in multimodal optic nerve imaging, an initial temporal evolution sequence is constructed based on the activation order and structural response intensity of each biomarker feature point in multi-time point images. A matching operation is then performed in the temporal domain using a sequence similarity mapping method to identify feature point paths with continuity and trends in the temporal structure. With iterative optimization of the similarity metric, once the similarity of the temporal evolution sequence reaches a convergence threshold, a set of nodes corresponding to that evolution sequence is extracted in the spatial domain. Based on their deformation trajectories and spatial distribution relationships in each modality, an initial distance matrix between nodes is constructed. To enhance the continuity and stability of the paths in the spatial structure, a reverse update operation is performed on the aforementioned spatial node distance matrix after the temporal similarity converges. This operation dynamically adjusts the spatial distance weights between nodes by comparing the sequential consistency and structural morphological change trends between path points in the time series, thereby redistributing path connectivity relationships and making the paths exhibit higher coherence and physiological structural consistency in space. Ultimately, a continuous path structure with a high degree of coupling between temporal evolution sequence and spatial variation trajectory is formed, which is used to guide the determination of subsequent optic nerve degeneration direction and the assessment of prognostic risk level.

[0036] To effectively distinguish structural signals from noise interference during multimodal optic nerve image analysis and to quantitatively characterize the degree of co-variance between modalities, a method for constructing a non-mean-fitting error function for multimodal noise is proposed. This method establishes a non-mean-fitting error energy model by analyzing the signal intensity differences and covariance weighting relationships among modalities within the same region, thereby measuring the degree of signal deviation and co-variance stability between modalities. The error function is defined using a weighted integral form to comprehensively quantify the signal stability and co-variance trends of different modal images within a unified spatial region. Its mathematical expression is: in: The overall energy function representing the non-mean fitting error of multimodal noise is used to measure the degree of cooperative deviation between multimodal signals; Indicates the local spatial region involved in the calculation; The total number of image modalities involved in the analysis; Indicates the first The signal strength of each mode at spatial location x; It represents the non-mean reference function for all modal signals within the same region, used to describe the overall signal stability trend; For the first The signal confidence weights for each mode are adaptively determined based on the noise variance and signal consistency of the mode. For the first Each mode in the region The covariance response function within the range is used to reflect the local correlation of signal changes; It is a nonlinear amplification factor used to enhance the contribution of high-deviation regions to the overall error; It is a covariance balancing factor used to adjust the effect of gradient changes on the suppression of error energy.

[0037] By calculating the changes in the non-mean deviation and covariance gradient of each mode within the same region, it is possible to effectively distinguish the regions of coordinated change between noise disturbances and structural signals. The residual energy distribution optimized by this function is used to identify regions of consistent signal changes, thus providing a high-confidence input basis for subsequent signal coupling region extraction and structural consistency analysis.

[0038] Example 1: Combined with appendix Figure 4 This embodiment uses a male subject as the research subject, whose chief complaint is blurred vision and visual field defects. During the initial consultation, and at follow-up stages at 3 months and 6 months, three types of multimodal imaging data of the optic nerve were collected, including optical coherence tomography (OCT) images, T2-weighted magnetic resonance imaging (MRI) images of the optic nerve along the axial direction, and retinal electrophysiological ERG response curve data. The sampling resolutions were 5 micrometers horizontally for OCT, 0.5 millimeters horizontally for MRI, and 1000 points per second for ERG. After image registration, the images were unified to a single spatial resolution of 10 micrometers per pixel.

[0039] First, the system constructs a local signal stability weight field. For each... Based on the pixel neighborhood, the coefficient of variation (COP) of the signal intensity sequence in this region is calculated in both OCT and MRI images. If the COP is less than 0.25, the weight is set to 1; if the COP is greater than 0.5, the weight is reduced to below 0.6. Simultaneously, covariance directionality analysis is performed in each neighborhood, calculating the covariance matrix of the local gradient vector of each pixel relative to the dominant direction. This reveals a strip-like structure with strong covariance directionality consistency in the nasal optic nerve head region. This method effectively eliminates high-frequency noise regions caused by vascular artifacts in MRI images while preserving anatomically continuous fibrous tissue signals in OCT images.

[0040] Further differential filtering energy analysis was performed on the above signals. The signal energy was calculated for each electrophysiological curve region in the ERG image. The system executes a sliding difference operator and analyzes its similarity to the energy change trajectory of the interlayer structure signal in OCT. The system found that around the optic disc in the 3 o'clock direction, the energy attenuation of ERG signal is highly consistent with the decreasing trend of the thickness of the nerve fiber layer in OCT. By setting a Pearson similarity coefficient greater than 0.85 as a threshold, multiple regions with coordinated energy decrease were screened out and defined as signal stable coupling regions.

[0041] The system then calculates the modal structural consistency deviation rate for these regions. To reduce interference from differences in boundary perception between different modalities, a structural domain boundary tensor decomposition method is used. The boundary intensity map of each modality is represented as a multidimensional tensor, and the first two principal components are retained after Tucker decomposition to extract the deformation principal axis. In several regions, especially in the 15 to 20 pixel area below the optic disc nose, the principal component directions of MRI and OCT form a 47-degree angle, significantly deviating from the low-response band direction of ERG. Based on this, the system identifies structural inconsistencies in this region and constructs a consistency deviation rate matrix with an average deviation rate of 0.36, which is much higher than the average of 0.12 for other regions.

[0042] Finally, to identify specific structural aberration patterns, a multi-level texture coupling analysis method was introduced. The system extracted texture features at different directions and scales based on Gabor filter banks and statistically analyzed the texture alignment rate between different modalities. In the nasal structural offset region, the OCT-MRI texture coupling strength was found to be less than 0.4, significantly lower than the average of 0.78 in other regions. Depth distribution pattern analysis showed that this type of structural breakage extended from the superficial nerve fiber layer to the deep reticular layer, consistent with the evolution trend of early glaucoma damage. The system identified this as a cross-modal structural aberration region and mapped it to the original images of each modality as a prognostic sensitive area.

[0043] After identifying cross-modal structural anomalous splitting regions, precise mapping operations are further performed on these regions to reduce the interference of geometric distortion caused by registration errors in each modality on the analysis results. The system introduces a reverse resampling method based on spatiotemporal registration residual sequences. In this example, MRI images are used as a global structural reference, and OCT and ERG data are registered to it respectively, recording the residual trajectory of each pixel during the registration process. Taking the nasal optic disc region as an example, this region still has an average residual of 4.2 pixels after initial registration, with a maximum residual of 8.7 pixels, mainly concentrated in the torsional transition area of ​​optic nerve fibers.

[0044] The system generates a spatiotemporal perturbation model based on these registration residuals, reverse-maps the original coordinates of the anomalous split regions to the image space before registration, and resamples to obtain the true physical location. Before resampling, the residual sequence is smoothed with a one-dimensional filter to remove abrupt noise, and a modality weighting factor is introduced to prioritize modes with more stable local signal intensity as the mapping reference. Using this method, the maximum mapping deviation of the nasal side structure offset region is reduced to 2.1 pixels, effectively correcting the stretching and compression deformation in the structural extension direction.

[0045] After spatial mapping, to accurately identify regions with long-term degradation risks, a regional self-stabilization index was calculated for each anomalous structural region. This index comprehensively considers two dimensions: temporal intensity fluctuations and spatial texture perturbations. The temporal feature is derived from the standard deviation of the intensity change in the region during three follow-up periods. In this example, the standard deviation of the region corresponding to the ERG curve is 0.22, significantly higher than the average of 0.07 for the entire visual disk. Spatial texture perturbations are calculated by measuring the coefficient of variation of the local second-order statistical matrices of adjacent pixels. The texture perturbation index of the OCT data in this region reaches 0.35, much higher than the 0.12 of the background region.

[0046] The system normalizes the two dimensions mentioned above and uses a weighted combination method to obtain the final regional self-stabilization index. The index value of the nasal split region was evaluated to be 0.79, which exceeds the set threshold of 0.65. It was finally identified by the system as a region with high prognostic sensitivity.

[0047] After evaluating the mapping accuracy and stability, to analyze the potential degenerative trend of the patient's optic nerve structure during follow-up, a time-point difference spectrum skewness index was introduced to identify time points with sudden evolutionary behavior. Based on multimodal imaging data from three time points, the system divided the structural parameter sequences extracted from each prognostic sensitive region into sliding window clusters. Taking the optic nerve fiber layer thickness sequence in OCT as an example, its values ​​at initial diagnosis, 3-month follow-up, and 6-month follow-up were 126 μm, 121 μm, and 108 μm, respectively. After constructing a difference spectrum sliding sequence with a window width of 2, the first window difference value was -5, and the second window difference value was -13. Calculating the skewness of these two window differences revealed a skewness value of 0.72, which is much higher than the system's set mutation threshold of 0.4, indicating significant nonlinear degeneration.

[0048] To avoid biases in the analysis results caused by time differences between different modalities during the acquisition process, a cross-modal time alignment mechanism is introduced. Considering that MRI acquisition is delayed by about 2 hours compared to OCT and ERG by about 30 minutes, the system performs time interpolation on the data of each modality and aligns them to a unified reference time when calculating the difference spectrum skewness. This ensures the consistency of parameter change trends and improves the accuracy of identifying time abrupt changes.

[0049] After identifying time points and corresponding regions with nonlinear mutation tendencies, the system further performs intermodal feature contribution analysis. First, it introduces the calculation of cross-entropy approximation to assess the distributional differences in the predicted responses of each modality. Taking the visual function score of the target region predicted by the three modalities as an example, the OCT predicted value is 0.78, the MRI predicted value is 0.69, the ERG is 0.81, and the actual clinical score is 0.74. The system constructs a modal confidence constraint matrix, assigning a higher weight of 0.45 to the low-noise OCT, while setting the highly volatile MRI to 0.25. After introducing this unbalanced weighting mechanism, the corrected cross-entropy approximation loss is 0.063, which is better than the 0.088 under the unweighted condition, indicating that the confidence constraint has a significant effect on improving the consistency of multimodal predictions.

[0050] To further identify the modal information that contributes most to prognostic prediction, the system performs multi-scale mutual information decomposition on the same region. Analysis of the mutual information values ​​between different modal images at different spatial scales revealed that the mutual information between OCT and ERG in the local high-resolution region (scale 1) is 0.46, while the mutual information between MRI and OCT in the mid-scale region (scale 3) is only 0.22, indicating that MRI contains a significant amount of redundant information in this region. Based on this, the system constructs a modal redundancy penalty index to weaken modal information with low mutual information and low confidence, making the finally selected biomarker feature points more stable, independent, and prognostically valuable.

[0051] Example 2: Combined with appendix Figure 5 Based on Example 1, for the identified nonlinear mutation time points and their corresponding prognostic event candidate regions, further prognostic biomarker feature points were extracted and trend modeling of structural degradation direction was carried out. Based on the joint analysis of OCT, MRI, and ERG trimodal data, a cross-modal feature contribution ranking sequence was constructed. Using structural morphology, texture complexity, and signal consistency as indicators, after feature variance normalization, the contribution of GCL thickness in OCT was 0.83, optic nerve sheath tension in MRI was 0.62, and photoreaction potential in ERG was 0.77. After weighted comparative analysis within the same region, the system ranked the contribution values ​​sequentially and performed screening in conjunction with subsequent stability assessment.

[0052] An information stability criterion was introduced to exclude feature points with drastic short-term fluctuations or occasional anomalies. In actual data, the response persistence rate of the GCL thickness sequence was calculated over a three-month follow-up period, resulting in 87.5%, while the response persistence rate of the potential value in the ERG was 52.3%. The system set a stability threshold of 70%, therefore only GCL thickness was retained as a valid feature point, and its exclusive information density index was calculated, with a density value of 0.93, indicating a highly non-redundant independent contribution in this region.

[0053] Next, the sequential similarity mapping method is used to construct paths for feature points to characterize their structural evolution sequence and spatial deformation trajectory at different time points. Based on the spatial projection positions of the GCL thickness at three time points, an initial set of path nodes is established. Then, the comprehensive similarity between each node in terms of temporal trend and spatial geometric change is calculated to form a temporal evolution path. The similarity values ​​are set to increase by 0.91, 0.94, and 0.89 over time. The system determines that its sequential evolution is good, so the path is retained for subsequent structural trend modeling.

[0054] Furthermore, a local convolution smoothing mechanism was introduced into the path vector field. Using the constructed path vector as input, first-order derivative convolution was performed on three local regions to eliminate directional jumps caused by pixel perturbations or registration errors. The standard deviation of path angle change before processing was 17.6°, which decreased to 6.3° after processing, improving the continuity of the structural degeneration trend. Finally, the system combined the path direction vector with the trend of feature value changes to identify the degeneration trend spreading from the nasal region of the fovea to the superior temporal quadrant. Based on the information density and mutation index superposition value in the path, the prognostic risk level was determined to be medium to high risk. It was recommended that the patient increase the follow-up frequency in the next six months and focus on observing the fluctuations in neurological function in the superior temporal region.

[0055] For the multimodal imaging data acquired from this 52-year-old subject in the early stages of optic neuropathy, the system executed a process to establish a feature contribution ranking sequence. In the OCT data, the thickness of the retinal nerve fiber layer (RNFL) fluctuated significantly across multiple spatial layers, with a mean standard deviation of 7.2 μm, while the fluctuation was lowest in the periphery of the temporal fovea, at 2.3 μm. In the MRI data, the cross-sectional area of ​​the optic nerve sheath structure showed a consistent deformation trend across multiple resolution layers, with the variation range concentrated within 5%. Simultaneously, in the ERG, the b-wave amplitude exhibited significant temporal correlation in the central region, with a root mean square deviation of less than 0.8 μV across multiple measurements.

[0056] To accurately assess the stability contribution of these features at different spatial scales, the image data was segmented into multi-level feature resolution segments within spatial windows of 32×32, 64×64, and 128×128, respectively. Taking RNFL thickness as an example, its spatial stability weight at the 64×64 level is 0.86, higher than its performance at other levels; therefore, the system assigns this feature a ranking weight of 0.91 at this level. Meanwhile, the stability score of the nerve sheath tension feature in MRI at the 128×128 level is 0.79, with a corresponding ranking weight of 0.84.

[0057] Subsequently, a time-staggered stability comparison mechanism was introduced for candidate feature points in the sorted sequence. Based on the patient's data from four follow-up visits over six months, the system extracted whether each feature point consistently expressed its trend characteristics in adjacent time windows. For example, GCL thickness showed a continuous linear decrease across all time windows, with a response persistence rate of 100%; while b-wave amplitude showed a reverse jump in the latter two follow-up visits, with a persistence rate of only 50%. Based on this, the system set a discrimination threshold of 70% to select feature points with long-term stable expression for subsequent evolutionary path construction.

[0058] In the path coupling construction phase, the initial temporal evolution sequence is constructed using the time series of sorted high-confidence feature points. Sliding window clustering is used to identify the spatial changes at each time point, forming an initial path node chain. For the GCL thickness feature, it appears in coordinate region A1 at the first time point, subsequently forming continuous paths in regions B3, C5, and D6. After evaluating the temporal similarity between nodes, the system found a path convergence rate of 94%, meeting the path stability requirements.

[0059] To enhance the spatial coherence of the path, the spatial distance matrix between nodes is updated in reverse. Taking nodes B3 and C5 in the temporal evolution path as an example, the initial Euclidean distance is 2.1 mm, which is adjusted to 1.6 mm after weighting for temporal trend similarity. The final total path length is compressed from the original 7.4 mm to 5.8 mm, indicating that the path continuity is effectively improved. The area covered by the path ultimately falls in the optic nerve axonal junction region of the patient's inferotemporal quadrant, which is consistent with the subsequent pathological expansion direction, confirming that the proposed method can predict the trend of structural degeneration in advance.

[0060] Three modalities of imaging data were collected from the subjects: optical coherence tomography (OCT), magnetic resonance imaging (MRI), and retinal electrophysiological (ERG) data. The spatial extent of the macular region in the right eye was systematically selected. Perform error function evaluation and analysis, with a region size of [missing information]. Pixel.

[0061] 1. Setting up the original multimodal signal Set a point in this area The modal signal strength at the location is as follows: The value is 82.5, representing the OCT image signal strength. The signal intensity of the MRI image is 79.3. The signal strength after normalizing the amplitude of the ERG response potential is set to 85.7. The number of modes is .

[0062] 2. Calculation of Non-Mean Reference Signal Use the non-mean reference function of the three-mode signal at the same point in this region: 3. Gradient setting of modal covariance response The derivative of the covariance response function at this point is calculated as follows, simulating known local statistical results: 4. Signal confidence weight The confidence weight is set by inversely proportional to the noise variance of each mode within this region: This indicates that the OCT noise is relatively small; ; ; 5. Setting the parameters of the fitting error function Set the nonlinear amplification factor Covariance balance factor .

[0063] 6. Error function calculation expression Substituting the above values ​​into the weighted integral error function: Because the calculation is a single point The integral can be viewed as a single-point discrete sum, and each term can be calculated by substituting the modal values: for : for : For i=3: 7. Error Function Results Add the three items together: This value indicates a strong cooperative deviation behavior among multiple modalities at this point, suggesting that the signal is in a weakly stable state or an early precursor region of variation. By accurately defining and calculating the multimodal noise mean fitting error function, not only can the stable quantification of cross-modal signal differences be achieved, but it can also serve as an important basis for subsequent identification of cooperative change regions and coupled structure boundaries. In the above example, the effective calculation of this error function on the real data point reflects a consistent deviation of ERG and MRI from the overall reference trend at this point, verifying the possibility of multimodal abnormal signals at this point and providing a precise basis for subsequent lesion localization and prognostic risk modeling.

[0064] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for extracting prognostic biomarkers from multimodal optic nerve imaging, characterized in that, include: Acquire optical coherence tomography, magnetic resonance imaging, and retinal electrophysiological data; The visual nerve regions in each modality of image are registered and aligned, and local signal blocks are established within the same spatial region. The calculation of the inter-modal signal-to-noise ratio change rate based on the regional signal-to-noise ratio covariance includes: obtaining the signal-to-noise ratio change of each mode in the local region through statistical analysis and calculating its covariance matrix; and judging the signal stability of the local region based on the consistency of the covariance direction. A local signal stability weight field is constructed for each modal image. The local signal stability weight field uses the pixel neighborhood as the basic unit. By calculating the mean shift of the signal intensity within the pixel neighborhood and the difference in response between consecutive frames, a stability index reflecting the temporal consistency and spatial smoothness of the signal is generated. Based on the stability index, stable signal regions in each modal image are assigned higher weights, while noisy or high-fluctuation regions are assigned lower weights. Based on the local signal stability weight field, a covariance directionality analysis method is used to construct tensors for the change direction of each modal signal within the neighborhood and extract the principal direction components. The consistency of the principal direction components among different modalities is then compared. The residual calculation is performed on the deviation of the multimodal signal according to the multimodal noise non-mean fitting error function to identify the signal cooperative change region; the multimodal noise non-mean fitting error function is a residual evaluation function obtained by weighted summation of the difference between the non-mean deviation of the signal intensity of each mode in the same region and its covariance weight, which is used to quantify the degree of signal cooperative change between modes. The multimodal noise non-mean fitting error function is defined in weighted integral form, and its expression is: in, The overall energy function representing the non-mean fitting error of multimodal noise is used to measure the degree of cooperative deviation between multimodal signals; Indicates the local spatial region involved in the calculation; The total number of image modalities involved in the analysis; Indicates the first The spatial location of each modality Signal strength at the location; It represents the non-mean reference function for all modal signals within the same region, used to describe the overall signal stability trend; For the first The signal confidence weights for each mode are adaptively determined based on the noise variance and signal consistency of the mode. For the first Each mode in the region The covariance response function within the range is used to reflect the local correlation of signal changes; It is a nonlinear amplification factor used to enhance the contribution of high-deviation regions to the overall error; It is a covariance balancing factor used to adjust the effect of gradient changes on the suppression of error energy; When the residual value obtained from the residual calculation is lower than a set threshold, the corresponding region is determined to be a multimodal signal co-variation region; when a region has the same principal covariance direction in multiple modes and the corresponding region has high stability weights, the region is determined to be a structure signal dominant region; differential filtering is performed on the energy distribution characteristics of the multimodal image, the energy attenuation rate of the image block is calculated under a fixed window scale, and the consistency of the attenuation trend direction is used as the joint screening criterion to extract regions that simultaneously show similar signal energy decline trajectories under different modes, and the regions are identified as signal stable coupling regions; For the signal stable coupling region, the structural parameter differences of each modality at the same spatial location are calculated, and the curvature of the structural distribution and the degree of boundary diffusion of the multimodal image in the signal stable coupling region are compared to form a modal structure consistency deviation rate, which is used to measure the consistency of structural expression among multimodalities. The modal structure consistency deviation rate of each region is compared with a threshold to identify local regions where the deviation of modal structural parameters significantly exceeds the average value range. The local regions are identified as cross-modal structural abnormal split regions, and the cross-modal structural abnormal split regions are mapped to the corresponding modal original data and labeled as prognostic sensitive regions. Structural parameters were extracted and time-series processed from multimodal follow-up image sequences collected at multiple time points for the same subject. Structural change trend maps and time-point difference maps were constructed. Based on the time-point difference maps, time-point difference spectral skewness was calculated to quantify the changing trend of structural evolution rate across different time periods. The calculation of time-point difference spectral skewness involved constructing a sliding window cluster to divide the multimodal follow-up image sequence into multiple time window sets with overlapping ranges. Trend fitting and slope analysis were performed within each time window. By comparing the slope changes of adjacent time windows, transient noise changes and true structural abrupt changes were distinguished. Time points with drastic slope changes and significant disruption of continuity were identified as nonlinear abrupt change time points, and the spatial regions corresponding to these nonlinear abrupt change time points were used as candidate regions for prognostic events. Information content analysis is performed on the response signals of each image modality in the candidate regions of the prognostic events, and the cross-entropy approximation value of each modality response is calculated. The cross-entropy approximation value is used to construct an error term by comparing the dispersion of the modal response distribution with the multimodal joint response benchmark, which is used to measure the degree of deviation of the prediction results of each modality for the same lesion area. The modal redundancy penalty index is determined by combining the cross-entropy approximation. The modal redundancy penalty index is used to quantify the non-independence of each modality in feature representation. The higher the modal redundancy penalty index, the stronger the repetition between modalities. The calculation of the modal redundancy penalty index includes performing multi-scale mutual information decomposition, jointly modeling the information flow features of different modalities in the prognostic event candidate region, constructing multi-scale image patches and their corresponding statistical feature spaces, performing frequency domain decomposition and directional component extraction on the mutual information intensity between each modality, identifying the proportion of shared information between modalities and the distribution density of independent information gain, quantifying the redundancy of each modality based on the mutual information decomposition, constructing a redundancy penalty factor and combining it with the original prediction information intensity. Based on the contribution ranking results and the stability of modal information distribution, a marker-specific information density is constructed. The calculation of the marker-specific information density includes calculating the contribution of each modality to the predicted target within the candidate region of the prognostic event, establishing a feature contribution ranking sequence among modalities, and determining effective feature points through information stability criteria. The feature contribution ranking sequence is arranged from high to low contribution. The information stability criteria include the response stability of feature points in multi-time point images and the signal change coherence of feature points in their neighborhood regions within the spatial domain. Based on the modal redundancy penalty index and the marker-specific information density, feature points are ranked and filtered, retaining marker feature points with low redundancy, high information independence, and high contribution. Based on the spatial distribution and variation characteristics of the marker feature points, the posterior conditional transition probability is calculated, potential association paths of the markers are generated, and the direction of structural degradation and prognostic risk level are determined.

2. The method for extracting prognostic biomarkers from optic nerve multimodal imaging according to claim 1, characterized in that, Before calculating the modal structure consistency deviation rate, structural domain boundary tensor decomposition is performed to extract the main components of cross-modal structural deformation; the identification of cross-modal structural abnormal splitting regions includes the introduction of hierarchical texture coupling analysis to detect the depth distribution pattern of structural fractures between modes.

3. The method for extracting prognostic biomarkers from optic nerve multimodal imaging according to claim 2, characterized in that, When mapping the cross-modal structural anomaly splitting region, reverse resampling is performed based on the spatiotemporal registration residual sequence to reduce the impact of local geometric distortion on the region mapping; the calibration of the prognostic sensitive region is performed by evaluating the region self-stabilization index, which is based on the joint determination of temporal domain intensity fluctuation and spatial domain texture perturbation.

4. The method for extracting prognostic biomarkers from optic nerve multimodal imaging according to claim 1, characterized in that, A cross-modal time alignment mechanism is introduced when identifying the nonlinear abrupt change time point to correct the impact of different modal acquisition delays on time difference calculation.

5. The method for extracting prognostic biomarkers from optic nerve multimodal imaging according to claim 1, characterized in that, The calculation of the cross-entropy approximation introduces a modal confidence constraint matrix to achieve unbalanced weighting in the multimodal prediction distribution.

6. The method for extracting prognostic biomarkers from optic nerve multimodal imaging according to claim 1, characterized in that, When generating the potential association path of the marker, the sequential similarity mapping method is used to establish the coupling relationship between the temporal evolution sequence and the spatial variation trajectory; the local convolution smoothing of the path vector field is performed during the determination of the structural degradation direction.

7. The method for extracting prognostic biomarkers from optic nerve multimodal imaging according to claim 6, characterized in that, The establishment of the feature contribution ranking sequence includes performing multi-level feature resolution segmentation processing and determining the ranking weight based on the feature stability distribution at different spatial scales; the determination of the information stability criterion adopts a time-staggered stability comparison mechanism, which identifies feature points with long-term stable expression by calculating the response duration rate of feature points in different time windows.

8. The method for extracting prognostic biomarkers from optic nerve multimodal imaging according to claim 6, characterized in that, When establishing the coupling relationship between the temporal evolution sequence and the spatial variation trajectory, the distance matrix of the corresponding node in the spatial domain is updated in reverse after the temporal similarity converges, so as to enhance the path continuity.

Images