Exosome and optic nerve injury prognosis evaluation method and system

Abstract

The application discloses an exosome and optic nerve injury prognosis evaluation method and system, relates to the field of data processing and artificial intelligence, and the method comprises the following steps: collecting and marking the prognosis grade of the optic nerve MRI data; adopting anisotropic Gaussian interpolation to perform adaptive processing on three-dimensional MRI voxel data, so as to retain the micron-level structure of the optic nerve fiber layer; constructing a double-branch gated classification network, extracting axon injury and fusion prognosis related features through independent branches, and training by adopting a dynamic focus loss function combined with a confidence loss; and finally outputting the probability distribution of the prognosis of the patient through the model. The application can effectively fuse multi-modal information, enhance the fine damage feature representation, and improve the accuracy and interpretability of the prognosis evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing and artificial intelligence, in particular to an exosome and optic nerve injury prognosis evaluation method and system. BACKGROUND

[0002] Optic nerve injury is an important complication of various ophthalmic and nervous system diseases, its pathological process is complex, involving axonal injury, glial cell reaction and microstructure destruction, and different degrees of injury directly affect the visual function recovery and long-term prognosis of patients. The clinical routine prognosis evaluation mainly depends on imaging examination and physician experience judgment, but the traditional magnetic resonance imaging has significant limitations in the aspects of insufficient interlayer resolution, weak signal of subtle lesions and insufficient multi-modal information integration, resulting in lack of objectivity and accuracy in early identification and prognosis grading of optic nerve injury.

[0003] The prior art including the invention patent with publication number CN119400393B proposes an infarction change prediction method and system based on brain parenchymal MRI image, which has the following problems: the conventional bicubic or linear interpolation method is prone to blur effect when processing interlayer low resolution data, resulting in loss of optic nerve fiber layer microstructure and affecting accurate extraction of injury features; the existing multi-modal image fusion method mostly uses direct splicing or weighted average, which is easy to cause feature conflict and information redundancy between different modalities, and difficult to highlight the weak signal features of the optic nerve subtle injury area; the existing three-dimensional convolutional neural network based on fixed convolution kernel is difficult to adapt to the high variability of the injury area morphology, the receptive field is limited, and the local details and global context information cannot be considered at the same time, resulting in insufficient feature extraction; the existing method often mixes the features of different pathological and physiological processes in the prognosis evaluation, lacks independence constraint, and the commonly used loss function does not fully consider the class imbalance and artificial labeling confidence difference, resulting in insufficient prediction accuracy and stability of the model. SUMMARY

[0004] In order to solve the technical problem of insufficient prediction accuracy and stability of the model in the prior art, the present application provides an exosome and optic nerve injury prognosis evaluation method and system.

[0005] The present application is realized by the following technical solutions:

[0006] An exosome and optic nerve injury prognosis evaluation method, comprising:

[0007] S1, optic nerve MRI data acquisition and labeling; including acquiring optic nerve MRI data and performing artificial labeling, the labeling categories including prognosis grade;

[0008] S2, anisotropy Gaussian interpolation processing of optic nerve MRI data; specifically, based on an anisotropy Gaussian interpolation kernel function, the original three-dimensional MRI voxel data is adaptively interpolated;

[0009] S3, prognosis evaluation model construction and training; the model adopts a double-branch gated classification network, extracts axon damage features and discriminates information of fused prognosis-related features through independent branches, adopts a dynamic focal loss function combined with a prognosis confidence loss for training, and dynamically adjusts the difficult sample weight through a time-related adjustment factor;

[0010] S4, exosome and optic nerve injury prognosis evaluation; the processed optic nerve MRI data of the patient is input into the trained prognosis evaluation model to obtain a prediction probability distribution representing the probability that the patient belongs to a prognosis grade.

[0011] Further, the adaptive interpolation processing of the original three-dimensional MRI voxel data based on the anisotropy Gaussian interpolation kernel function includes: using a small standard deviation to reserve high-resolution details in the XY plane, using a large standard deviation to achieve weak interpolation in the Z axis, and obtaining interpolated three-dimensional MRI voxel data through local neighborhood weighted summation.

[0012] Further, the prognosis evaluation model construction further includes constructing a multi-modal feature projection and subtle damage response enhancement module, using a cross-modal projection matrix to transform and fuse the T1, T2, and DWI modal data processed by the anisotropy Gaussian interpolation, and using a subtle damage response factor based on gradient amplitude to enhance the subtle damage features.

[0013] Further, the subtle damage feature enhancement further includes adaptively weighting the fused multi-modal feature tensor using the subtle damage response factor, preserving the original features in the damage area, and enhancing the signal in the non-damage area through power operation to obtain damage-enhanced features.

[0014] Further, the prognosis evaluation model construction further includes constructing a deformable subtle feature extraction module, introducing a deformable convolution mechanism based on standard 3D convolution, and learning an offset and a modulation factor for each convolution sampling point to enable the convolution kernel to adaptively adjust the sampling position according to the input features.

[0015] Further, the prognosis evaluation model construction further includes constructing a multi-scale context aggregation module, which fuses feature responses under different receptive fields by constructing a multi-scale hollow convolution group and an adaptive receptive field weighting mechanism.

[0016] Further, the prognosis evaluation model construction further includes constructing a channel-space dual-path attention module, which guides channel and spatial attention calculation through exosome distribution prior knowledge, and realizes adaptive weighting of dual-path attention through a gating fusion mechanism.

[0017] Further, the prognosis evaluation model construction further comprises constructing a feature decoupling and enhancement module, and the coupled features are decomposed into axonal injury related features and glial cell reaction related features by adopting a prognosis factor decoupling layer and an orthogonal constraint loss.

[0018] The application also provides an exosome and optic nerve injury prognosis evaluation system based on the exosome and optic nerve injury prognosis evaluation method.

[0019] An optic nerve MRI data acquisition and labeling module is used to acquire optic nerve MRI data and perform artificial labeling, and the labeling categories include prognosis grades.

[0020] An anisotropic Gaussian interpolation processing module is used to perform adaptive interpolation processing on the original three-dimensional MRI voxel data by using an anisotropic Gaussian interpolation kernel function, to retain high-resolution details in the layer in the XY plane using a small standard deviation, to achieve weak interpolation in the Z-axis using a large standard deviation, and to obtain interpolated three-dimensional MRI voxel data by locally weighted summation.

[0021] A prognosis evaluation model construction and training module comprises a multi-modal feature projection and subtle damage response enhancement module, a deformable subtle feature extraction module, a multi-scale context aggregation module, a channel-space dual-path attention module, a feature decoupling and enhancement module, a double-branch gated classification module, a loss function calculation module, a prognosis evaluation model iterative training and parameter updating module.

[0022] An exosome and optic nerve injury prognosis evaluation module is used to evaluate the prognosis of exosome and optic nerve injury based on the trained model, and output a prediction probability distribution representing the probability that the patient belongs to a prognosis grade.

[0023] In addition, to achieve the above-mentioned purposes, the application further provides a computer readable storage medium, and the computer readable storage medium stores a program instruction of an exosome and optic nerve injury prognosis evaluation method.

[0024] Compared with the prior art, the application has the following beneficial effects:

[0025] 1. The anisotropic Gaussian interpolation method proposed in the application maintains high-resolution details in the layer direction while enhancing tissue continuity through weak interpolation in the interlayer direction, and can effectively retain the micron-level structural features of the optic nerve fiber layer.

[0026] 2、The application fully fuses the complementary information in the multi-modal image through the cross-modal projection matrix and the subtle damage response factor mechanism, and enhances the subtle damage area of the weak signal, avoiding that the key lesion characteristics are covered by the background information.

[0027] 3、The application adopts deformable convolution and multi-scale context aggregation mechanism, so that the convolution kernel can be adaptively fitted to the morphologically complex damage area, and the representation ability of the model to different spatial scale damages is improved through multi-scale feature weighting.

[0028] 4、The application separates the axon damage feature and the glial cell reaction feature by using feature decoupling and orthogonal constraint, and combines the double-branch gated classification structure and the dynamic loss function design, which improves the accuracy of prognosis evaluation and enhances the interpretability of the result. BRIEF DESCRIPTION OF DRAWINGS

[0029] The drawings described herein are intended to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their description serve to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0030] Figure 1 is a flowchart of an exosome and optic nerve damage prognosis evaluation method according to an embodiment of the present application;

[0031] Figure 2 is a PSNR comparison of different interpolation methods according to an embodiment of the present application;

[0032] Figure 3 is a structure similarity comparison of different interpolation methods according to an embodiment of the present application;

[0033] Figure 4 is a classification accuracy comparison of different feature fusion methods according to an embodiment of the present application;

[0034] Figure 5 is a feature heat map of different scales and a multi-scale context feature heat map, wherein Figure 5(a) is a small-scale feature heat map of d=1, Figure 5(b) is a medium-scale feature heat map of d=3, Figure 5(c) is a large-scale feature heat map of d=5, and Figure 5(d) is a multi-scale context feature heat map;

[0035] Figure 6 From left to right are heat maps of axon damage features, glial cell reaction features, and fused prognosis-related features;

[0036] Figure 7 is an accuracy comparison of different methods on each prognosis level according to an embodiment of the present application. DETAILED DESCRIPTION

[0037] Embodiments of the present application will be described in detail below with reference to the drawings.

[0038] The above embodiments of the present application are only used to illustrate the technical solutions of the present application and their advantages and are not used to limit the protective scope of the present application. Apparent variations or modifications can be made to the embodiments according to specific working conditions and embodiments of the present application. Therefore, the protective scope of the present application should be subject to the scope of the claims of the present application.

[0039] It should be further noted that the drawings provided in the following embodiments are only schematic and are not drawn to scale. Only the components directly related to the present application are shown in the drawings, and the number, shape and size of the components are not drawn according to the actual implementation. The shape, number and ratio of the components in the actual implementation can be changed at will, and the layout of the components can be more complex.

[0040] Reference is made to Figure 1 A method for evaluating the prognosis of optic nerve injury by using exosomes, comprising the following steps:

[0041] S1, optic nerve MRI data acquisition and labeling

[0042] The optic nerve MRI data acquisition process uses a high-resolution magnetic resonance imaging device, such as a 3T MRI scanner, to perform three-dimensional multi-modal scanning on the optic nerve region through a standard clinical protocol to obtain original three-dimensional MRI voxel data. Specifically, the patient is placed in a supine position, the head is fixed to reduce motion artifacts, a special optic nerve coil is used, and thin-layer scanning is performed along the optic nerve axis. The layer thickness is set to 0.6-1.0 millimeters, and the interlayer spacing is ≤0.1 millimeters, ensuring that the in-layer resolution (XY plane) reaches a sub-millimeter level, and the interlayer resolution (Z axis) is consistent with the layer thickness.

[0043] The scanning sequence covers three modalities of T1-weighted imaging (T1WI), T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI), each of which generates independent three-dimensional voxel data.

[0044] After the original data acquisition is completed, two or more experienced neuro-radiologists independently label. The labeling method is based on DICOM format images, and a medical image labeling software is used to label categories including prognosis grade, including:

[0045] Grade I: slight injury, good prognosis;

[0046] Grade II: moderate injury, general prognosis;

[0047] Grade III: severe injury, poor prognosis.

[0048] At the same time, the degree of injury is evaluated by artificial assessment, and the corresponding confidence vector is marked for each data, which adopts a three-dimensional vector format, and the elements represent the confidence probability of grade I, grade II and grade III respectively, such as [0.9, 0.1, 0.0] representing 90% confidence in grade I.

[0049] The labeling results are cross-validated, and the inconsistent places are arbitrated by experienced doctors, and finally the prognosis label and confidence vector matched with each sample are generated as the basic truth value for model training.

[0050] S2, anisotropic Gaussian interpolation processing of optic nerve MRI data

[0051] The interlayer resolution of the optic nerve MRI data is much larger than the intra-layer resolution, and the conventional bicubic interpolation method will produce a significant blurring effect when processing such anisotropic data, especially in the cross-layer interpolation direction, which is easy to cause the loss of micron-level structural details of the optic nerve fiber layer, seriously hindering the accurate extraction of subsequent subtle damage features, and further affecting the accuracy of prognosis evaluation.

[0052] The application uses an anisotropic Gaussian interpolation kernel function to perform adaptive interpolation processing on the original three-dimensional MRI voxel data, uses a smaller standard deviation in the XY plane to retain the high-resolution details within the layer, and uses a larger standard deviation in the Z axis to achieve weak interpolation, and obtains interpolated three-dimensional MRI voxel data by local neighborhood weighted summation, thereby enhancing the interlayer continuity while retaining the micron-level structural features of the optic nerve fiber layer, which is represented as:

[0053]

[0054] In the formula, represents the gray value of the interpolated three-dimensional MRI voxel data at position ;

[0055] x is the X-axis coordinate index;

[0056] y is the Y-axis coordinate index;

[0057] z is the Z-axis coordinate index;

[0058] represents the gray value of the original three-dimensional MRI voxel data at position ; specifically, it is the MRI voxel data of any modality in T1, T2 and DWI multi-modal data;

[0059] represents an anisotropic Gaussian interpolation kernel function, used to control the interpolation weight in different directions, and the calculation method is represented as ;

[0060] is a natural exponential function;

[0061] represents the integer offset in the X direction, and the value range is ;

[0062] represents the integer offset in the Y direction, and the value range is ;

[0063] represents the integer offset in the Z direction, and the value range is ;

[0064] represents the standard deviation of the Gaussian kernel in the XY plane direction, which controls the layer interpolation strength, and the value is 0.7;

[0065] represents the standard deviation of the Gaussian kernel in the Z axis direction, which controls the interlayer interpolation strength, and the value is 1.8.

[0066] Further, each voxel position in the original three-dimensional MRI voxel data is traversed, and the current voxel is taken as the center, and the weighted sum is obtained in the local neighborhood according to the anisotropic Gaussian interpolation kernel function to obtain the interpolated three-dimensional MRI voxel data This process effectively suppresses the blurring effect caused by interlayer interpolation while maintaining the high-resolution details within the layer, and preserves the micron-level structural characteristics of the optic nerve fiber layer.

[0067] It should be noted that considering the physical characteristics of low interlayer resolution and high intra-layer resolution of MRI voxel data, on the basis of anisotropic interpolation, the anisotropic Gaussian kernel function is used, which uses a smaller standard deviation in the XY plane to preserve high-resolution details within the layer, and uses a larger standard deviation in the Z axis to achieve weak interpolation and avoid over-smoothing the interlayer structure, while enhancing the interlayer coherence while preserving the micron-level structure of the optic nerve fiber layer.

[0068] It should be noted that the interpolated three-dimensional MRI voxel data is a single modality interpolation result, and for T1, T2, DWI multi-modal data, anisotropic Gaussian interpolation processing needs to be performed independently for each modality, that is, the interpolated three-dimensional MRI voxel data of the th modality is defined as , m is the modality index, corresponding to the T1 modality data, is T2 modality data, is DWI modality data.

[0069] In one embodiment, the influence of different interpolation methods on image quality is analyzed, and the influence of different interpolation methods on the interlayer resolution reconstruction quality of the optic nerve magnetic resonance imaging data is evaluated. The experiment compares four interpolation techniques: bicubic interpolation (a conventional technique), linear interpolation (a conventional technique), nearest neighbor interpolation (a conventional technique), and anisotropic Gaussian interpolation proposed by the present technique. The advantages of the present technique in preserving micron-level structures of the optic nerve fiber layer are verified by quantifying two key indicators: peak signal-to-noise ratio (in decibels) and structural similarity index (dimensionless, value range 0-1) of image reconstruction quality. As shown in FIG. 1, in the PSNR comparison chart of different interpolation methods, the experimental data is intuitively displayed through a violin plot, and the violin main body position of the present technique is significantly higher than that of other methods, and the distribution is more concentrated, indicating that the reconstructed image has lower noise and more complete details; as shown in FIG. 2, in the structural similarity comparison chart of different interpolation methods, the violin plot of the method of the present application not only has the highest position, but also has a significant upward shift in the upper quartile region, proving that the reconstructed image has a significant advantage in structural information preservation. Under the two indicators, the "tail" (outlier region) of the violin of the present technique is significantly shortened, indicating that its performance on different samples is more stable, and the calculation method of the anisotropic Gaussian kernel function through XY plane weak interpolation and Z axis strong interpolation effectively suppresses the layer blur effect caused by conventional methods. Figure 2 Figure 3

[0070] S3, prognosis evaluation model construction and training

[0071] S301, constructing a multi-modal feature projection and subtle damage response enhancement module

[0072] Using conventional multi-parameter MRI feature fusion methods such as direct channel splicing or simple weighted averaging on T1, T2, and DWI modality data can easily cause feature conflicts and information redundancy between different modalities, making it difficult to fully utilize the complementary information between modalities. At the same time, due to the very weak signal of the optic nerve subtle damage area in MRI, these conventional fusion methods cannot effectively highlight the features of these key areas, resulting in important subtle features being easily overwhelmed by background noise or other information during prognosis evaluation;

[0073] The present application transforms and fuses the T1, T2, and DWI modality data processed by the anisotropic Gaussian interpolation through a cross-modality projection matrix, and enhances the subtle damage features by using a gradient amplitude-based subtle damage response factor. The specific steps are as follows:

[0074] 1) Multi-modal feature projection fusion

[0075] ​​The interpolated three-dimensional MRI voxel data of the anisotropic Gaussian interpolation processed T1, T2 and DWI modalities are cross-modality feature transformed through a projection matrix, and then processed through batch normalization and ReLU activation function to obtain a fused multi-modality feature tensor, so as to integrate the complementary information of the multi-modality, represented as:

[0076]

[0077] In the formula, represents the interpolated three-dimensional MRI voxel data of the i-th modality, with a dimension of , , corresponding to T1, T2 and DWI modalities respectively;

[0078] m is a modality index, ;

[0079] H is the height of the MRI voxel data, that is, the number of rows within the layer;

[0080] W is the width of the MRI voxel data, that is, the number of columns within the layer;

[0081] D is the depth of the MRI voxel data, that is, the number of layers;

[0082] represents the projection matrix of the i-th modality, which is a trainable parameter, used to realize cross-modality feature transformation, with a dimension of ;

[0083] is the output channel dimension of the projection matrix ;

[0084] is the transpose of ;

[0085] represents a batch normalization operation, used to stabilize the training process and accelerate convergence;

[0086] represents a rectified linear unit activation function, which is a nonlinear transformation function;

[0087] represents a fused multi-modality feature tensor, representing the feature tensor after multi-modality fusion, which integrates the complementary information of T1, T2 and DWI.

[0088] It should be noted that the dimensions of the T1, T2 and DWI three modalities are the same, all being HxWxD, but the gray scale distribution and contrast are different, reflecting different biophysical properties.

[0089] It should also be noted that the projection operation is equivalent to​ convolution, which extends each voxel gray value to a feature vector.

[0090] 2) Calculate the subtle damage response factor

[0091] Based on the spatial gradient amplitude of the fused multi-modal feature tensor, the subtle damage response factor is calculated by the Sigmoid function, which maps the gradient amplitude to the range of 0 to 1, and the value of the region with large gradient is close to 1, so as to highlight the weak signal of the damage region, which is expressed as:

[0092]

[0093] In the formula, The spatial gradient tensor of the fused multi-modal feature tensor is calculated by the three-dimensional difference operator, which represents the change intensity of the feature between voxels, and is specifically calculated by in the X direction in the Y direction in the Z direction , and then merged into the gradient tensor, expressed as

[0094] is the partial derivative symbol;

[0095] The spatial gradient tensor of the fused multi-modal feature tensor is calculated by the three-dimensional difference operator, which represents the change intensity of the feature between voxels, and is specifically calculated by

[0096] The subtle damage response factor is expressed as, the value range is (0, 1), the larger the gradient amplitude of the region, the closer the value to 1, indicating the potential damage region.

[0097] It should be noted that the calculation of the subtle damage response factor uses the calculation method of the Sigmoid function to map the gradient amplitude to the interval (0, 1), and the output of the damage boundary region with large gradient is close to 1, and the output of the region with small gradient is close to 0. It is particularly noted that the conventional Sigmoid function is usually used as a neuron activation function for nonlinear mapping, but in the calculation process of the subtle damage response factor , the role of the Sigmoid function is to construct an adaptive weight mapping, and the scaling factor of the denominator is set to -5, which is used to adjust the steepness of the curve, so that the subtle damage response factor can sensitively capture the small gradient change and highlight the weak signal of the damage region.

[0098] 3) Subtle damage feature enhancement

[0099] An adaptive weighting of the fused multimodal feature tensor is performed using a subtle damage response factor. This preserves the original features in the damaged region and enhances the signal in the non-damaged region through exponentiation, resulting in damage-enhanced features. This enhances the signal in the non-damaged region while preserving the subtle structure of the damaged region, and is expressed as:

[0100]

[0101] In the formula, This represents element-wise multiplication.

[0102] Represents the fusion of multimodal feature tensors Element-wise exponentiation is performed to the power of 1.2 to moderately enhance the signal strength in the non-damaged region;

[0103] This indicates that signal enhancement is performed in the non-damaged area;

[0104] It represents the damage enhancement features, characterizing the feature tensor after damage response weighting enhancement, preserving the original features in the damaged area, and moderately enhancing the signal in the non-damaged area.

[0105] It should be noted that, The term uses a power of 1.2 to calculate the response factor for subtle damage. Non-damaged regions close to 0 perform non-linear enhancement on feature values, increasing the signal intensity of the background region and preventing damage features from being submerged. Too high a power will lead to over-enhancement and noise, while too low a power will result in insufficient enhancement. 1.2 is an empirical value preset by humans.

[0106] It should also be noted that, The characterization preserves the original features in the damaged area and the response factors in subtle damage. Damage areas close to 1 retain their original features, avoiding over-enhancement that leads to distortion and ensuring that the fine structure of the damaged area is preserved.

[0107] In one embodiment, the impact of different feature fusion methods on classification performance is analyzed, focusing on the influence of different feature fusion strategies on the classification accuracy of prognostic assessment. Experiments compare three methods: direct splicing fusion, weighted average fusion, and the cross-modal projection fusion of this invention. Classification accuracy is used as an evaluation index; a higher value indicates a more accurate judgment of the prognostic level. This index ranges from 0 to 1, with values ​​closer to 1 representing better performance. Figure 4As shown, the experimental results show that the cross-modal projection fusion method proposed in the application is obviously better than the conventional fusion strategy in classification accuracy. From the data distribution, the precision value of the technology not only has the highest average value, but also has the smallest data fluctuation range, indicating that the method has better stability and reliability. Therefore, the cross-modal projection matrix can effectively integrate the complementary information of different modalities, reduce the feature conflict and information redundancy, and combine the subtle damage response factor mechanism to highlight the weak signal characteristics of the damage area, thereby improving the recognition ability of the subtle damage characteristics, and finally improving the accuracy of the prognosis evaluation.

[0108] S302, a deformable subtle feature extraction module is constructed

[0109] The optic nerve injury morphology has high variability. The conventional 3D convolution operation adopts a fixed geometric structure sampling grid, and its receptive field is limited and cannot adaptively adjust the sampling position according to the input content. The rigid structure is difficult to effectively fit the complex damage area, and it is easy to cause insufficient extraction of subtle lesion characteristics of variable damage, and finally affect the accuracy of prognosis evaluation.

[0110] The deformable convolution mechanism is introduced on the basis of the standard 3D convolution, and the offset and modulation factor are learned for each convolution sampling point, so that the convolution kernel can adaptively adjust the sampling position according to the input feature, thereby enhancing the representation and extraction ability of the morphologically variable damage characteristics. The specific steps are as follows:

[0111] 1) Constructing a deformable convolution operation

[0112] Based on the damage-enhanced feature, a learnable offset and modulation factor are introduced for each sampling point on the basis of the fixed sampling position of the standard 3D convolution, so that the convolution sampling position can be adaptively adjusted according to the input feature, thereby enhancing the representation ability of the morphologically variable damage characteristics, which is expressed as:

[0113]

[0114] In the formula, represents the output feature value of the deformable convolution at position , that is, the feature value corresponding to the position coordinate of the deformable convolution feature at position .

[0115] represents the position coordinate index, which is a three-dimensional vector in form;

[0116] represents the number of convolution kernels, which determines the number of channels of the output feature;

[0117] is the index of the sampling point in the convolution kernel.

[0118] denotes the weight parameter of the th convolution kernel;

[0119] denotes the sampling at the deformed position of the damage-enhanced feature ;

[0120] denotes the fixed offset of the th sampling point in the standard 3D convolution, which is artificially preset and used to determine the regular sampling grid. For example, for a 3x3x3 convolution kernel, , , the value range is , covering all integer offset combinations of the neighborhood around the center voxel;

[0121] denotes the offset of the th sampling point, which is a trainable parameter and is learned from the deformable convolution feature by the preset convolution layer;

[0122] denotes the modulation factor of the th sampling point, which takes a value range of , and is a trainable parameter that is learned from the deformable convolution feature by the preset convolution layer through the Sigmoid function activation function to adjust the importance of the sampling position.

[0123] It should be noted that in the construction process of deformable convolution, on the basis of standard 3D convolution, a learnable offset and a modulation factor are used for each sampling point, so that the convolution kernel can adaptively adjust the sampling position according to the input feature, better fit the morphological changes of the optic nerve damage, and enhance the ability to capture irregular and subtle structures, and improve the feature extraction accuracy of morphologically variable damage.

[0124] 2) Restricting the offset range to prevent excessive deformation

[0125] An L2 norm constraint is imposed on the learnable offset to limit the offset of each sampling point to no more than 2 voxel distances, preventing the sampling point from shifting too much and ensuring that the deformation is within a physically credible range, which is represented as:

[0126]

[0127] In the formula, denotes the L2 norm.​​

[0128] It should be noted that the constraints Limit the offset of each sampling point to no more than 2 voxel distances to prevent excessive sampling point offset from causing unstable feature extraction or exceeding the reasonable range, ensuring that the deformation is within the physically reliable range, and avoiding overfitting noise.

[0129] 3) Achieve sub-voxel precision sampling

[0130] For sampling positions with non-integer coordinates, feature values ​​are calculated using triple linear interpolation, and bilinear interpolation weights are used for distance-based calculation to achieve sub-voxel precision sampling, ensuring the differentiability of deformable convolution, expressed as:

[0131]

[0132] In the formula, The bilinear interpolation weights in the x-direction are expressed as follows: The closer the distance, the greater the weight;

[0133] The bilinear interpolation weights in the y-direction are expressed as follows: The closer the distance, the greater the weight;

[0134] The bilinear interpolation weights in the y-direction are expressed as follows: The closer the distance, the greater the weight;

[0135] This indicates the operation of taking the maximum value of the two options;

[0136] This represents a positional coordinate index distinct from p, and is a three-dimensional vector. form;

[0137] For position of coordinate;

[0138] For position of coordinate;

[0139] For position of coordinate.

[0140] 4) Generate deformable convolutional features

[0141] Deformable convolution features are generated based on deformable convolution operations, and defined. The deformable convolution feature represents a feature processed by a deformable convolution operation, and has stronger morphological adaptability.

[0142] In an implementation manner, the number of channels of the deformable convolution output feature is set as The dimension of the deformable convolution feature is .

[0143] The offset constraint and the sub-voxel interpolation of the term can ensure the stability and differentiability of the operation, are suitable for subtle structural changes of the optic nerve damage, and can enhance the adaptability of the model to complex morphologies and improve the robustness of feature expression.

[0144] S303, constructing a multi-scale context aggregation module

[0145] The optic nerve damage region has various morphologies and uneven spatial distribution, and a conventional single-scale convolution is difficult to simultaneously capture local subtle structure and global context correlation information, and a method of directly splicing different scale features is prone to introduce redundant information and weaken the discrimination ability of the model to the micro lesions.

[0146] The application fuses feature responses under different receptive fields by constructing a multi-scale hollow convolution group and an adaptive receptive field weighting mechanism, enhances the representation ability of the multi-scale damage structure, and the specific steps are as follows:

[0147] 1) Multi-scale hollow feature splicing

[0148] The deformable convolution feature is subjected to convolution operation with different hollow rates to obtain response features, the features are spliced along the channel dimension, a multi-scale fusion feature tensor is obtained, context information under different receptive fields is integrated, and the multi-scale fusion feature tensor is represented as:

[0149]

[0150] In the formula, The multi-scale fusion feature tensor represents a feature tensor fused by multi-scale hollow convolution, and integrates context information under different receptive fields.

[0151] The splicing operation along the channel dimension is represented as

[0152] The response feature of the deformable convolution feature with a hollow rate of is a feature obtained by subjecting the deformable convolution feature to convolution operation with a hollow rate of , such as The convolution kernel is sampled at intervals of 2 points to expand the receptive field, and the response feature refers to the output feature after the convolution operation.

[0153] This represents the porosity, controls the receptive field size of the convolution kernel, and sets... These correspond to convolution operations with dilation rates of 1, 3, and 5, respectively, providing progressively larger receptive fields.

[0154] 2) Generate multi-scale contextual features

[0155] Global average pooling is performed on the response features at each scale to obtain channel vectors. Adaptive weight coefficients are calculated using the weight vector and the Softmax function. Multi-scale features are then weighted and fused to obtain multi-scale context features, achieving adaptive selection of important scales, as shown below:

[0156]

[0157] In the formula, Representing multi-scale contextual features, dimensions, and response features. same;

[0158] Indicates the first The adaptive weighting coefficients for the scale are calculated as follows: ;

[0159] This represents a global average pooling operation that compresses the feature tensor into a channel vector along the spatial dimension.

[0160] This indicates that deformable convolutional features have a dilation rate of [missing information]. Response characteristics;

[0161] This indicates the void ratio, which is distinct from d.

[0162] Indicates the first The weight vector of the scale is a trainable parameter;

[0163] Indicates the first The weight vector of the scale is a trainable parameter;

[0164] Representation and Scale The corresponding learnable weight vector.

[0165] It should be noted that the adaptive weighting coefficients The calculation adopts the Softmax function. By calculating the weights of each scale, the model can adaptively select important scales according to the input features, which enhances the pertinence of multi-scale fusion and avoids interference from redundant information.

[0166] In one embodiment, the heat maps of features of different scales and the heat map of multi-scale context features are visualized as shown in Figure 5: Figure 5(a) is a small-scale feature heat map of d = 1, Figure 5(b) is a medium-scale feature heat map of d = 3, Figure 5(c) is a large-scale feature heat map of d = 5, and Figure 5(d) is a heat map of multi-scale context features.

[0167] S304, constructing a channel-space dual-path attention module

[0168] The spatial distribution of exosomes in the optic nerve MRI has specificity, and the conventional attention mechanism lacks the guidance of such biological prior knowledge, making it difficult to effectively focus on these key areas. At the same time, simply superimposing channel attention and spatial attention can easily lead to information conflict, and cannot effectively fuse the relevance of exosome biomarker information and image spatial features.

[0169] The present application guides channel and spatial attention calculation through exosome distribution prior knowledge, and realizes adaptive weighting of dual-path attention through a gating fusion mechanism. The specific steps are as follows:

[0170] 1) Channel attention calculation

[0171] The multi-scale context features are globally averaged and pooled, and the channel attention weight matrix is calculated through a multi-layer perceptron and a Sigmoid activation function to enhance the feature channels related to the distribution of exosomes, represented as:

[0172]

[0173] In the formula, represents the channel attention weight matrix, which is used to enhance the feature channels with high correlation to the distribution of exosomes;

[0174] represents a multi-layer perceptron composed of two fully connected layers, and the activation function uses a ReLU activation function;

[0175] represents a Sigmoid activation function that maps the weight to the interval.

[0176] 2) Spatial attention calculation

[0177] The multi-scale context features are maximally pooled along the channel dimension, and the spatial attention weight matrix is calculated through a three-dimensional convolution and a Sigmoid activation function to enhance the spatial regions rich in exosomes, represented as:

[0178]

[0179] In the formula, denotes a spatial attention weight matrix, used to enhance the spatial region of exosome enrichment;

[0180] denotes a max-pooling operation along the channel dimension, The output dimension of the term is ;

[0181] denotes a three-dimensional convolution operation with a convolution kernel size of .

[0182] 3) Gated fusion

[0183] The channel attention weight matrix and the spatial attention weight matrix are concatenated along the channel dimension, and a fused attention weight matrix is generated through a weight matrix and a hyperbolic tangent activation function, realizing adaptive fusion of channel and spatial attention, denoted as:

[0184]

[0185] In the formula, denotes a fused attention weight matrix, representing the weight matrix after the fusion of channel and spatial attention, used to weight and enhance exosome-related features;

[0186] denotes concatenating the channel attention weight matrix and the spatial attention weight matrix along the channel dimension;

[0187] denotes a gated fusion weight matrix, used to adjust the contribution ratio of the dual-channel attention, and is a trainable parameter;

[0188] denotes a hyperbolic tangent activation function, which constrains the output range to the interval .

[0189] 4) Attention fusion enhancement

[0190] The fused attention weight matrix is multiplied element-wise with the multi-scale context feature to obtain an attention fusion feature, enhancing the feature response of the exosome-related region, denoted as:

[0191]

[0192] In the formula, denotes an attention fusion feature, representing a feature tensor enhanced by attention, and the dimension is the same as the multi-scale context feature .

[0193] In one embodiment,

[0194] S305, construct a feature decoupling and enhancement module

[0195] In optic nerve prognosis evaluation, the features reflecting axonal injury and the features reflecting glial cell reaction are often coupled in the same feature space, which leads to poor explainability of the prognostic factors, the conventional feature fusion methods cannot effectively separate the features from different pathophysiological processes, and there is a lack of constraint mechanism for feature independence, resulting in reduced accuracy of prognosis prediction.

[0196] The application decouples the coupled features into axonal injury related features and glial cell reaction related features by adopting a prognostic factor decoupling layer and an orthogonal constraint loss, and the specific steps are as follows:

[0197] 1) Feature decoupling

[0198] The attention fusion features are subjected to channel dimension transformation through 1x1x1 convolution, and then split into axonal injury features and glial cell reaction features along the channel dimension, so as to realize feature decoupling, which is expressed as:

[0199]

[0200] In the formula, represents a convolution operation with a kernel size of , which is used for feature channel dimension transformation and decoupling;

[0201] represents an operation of splitting a tensor into two sub-tensors along the channel dimension, for example, if the output 64-channel features are split into two 32-channel tensors by the splitting operation;

[0202] is an axonal injury feature, representing the axonal injury related feature obtained after decoupling;

[0203] is a glial cell reaction feature, representing the glial cell reaction related feature obtained after decoupling;

[0204] represents the combination of splitting a tensor into two sub-tensors of axonal injury features and glial cell reaction features along the channel dimension.

[0205] 2) Orthogonal constraint enhancement

[0206] The Frobenius norm square of the transpose of the axonal injury features and the glial cell reaction features is calculated as an orthogonal constraint loss, which forces the two types of features to be orthogonal and enhances the feature independence, which is expressed as:

[0207]

[0208] In the formula, is the transpose of .

[0209] is the Frobenius norm;

[0210] represents the orthogonal constraint loss, and the smaller the value is, the stronger the feature independence is.

[0211] It should be noted that the orthogonal constraint loss adopts the calculation mode of the Frobenius norm, and at the same time forces the two types of features to be orthogonal in the vector space, thereby reducing semantic confusion, enhancing feature independence, and improving the explainability of the prognostic factor.

[0212] 3) Prognostic feature weighted fusion

[0213] The axonal injury features and glial cell reaction features are weighted and fused, the axonal injury features retain the original value, the glial cell reaction features are multiplied by a weight coefficient, and the fused prognostic related features are obtained, which are represented as:

[0214]

[0215] In the formula, represents the fused prognostic related features, and characterizes the fused prognostic related features;

[0216] represents the weight coefficient of the glial cell reaction features, and the preferred value is 0.3.

[0217] In one embodiment, the heat maps of the axonal injury features, the glial cell reaction features and the fused prognostic related features are visualized as shown in Figure 6 from left to right are the heat maps of the axonal injury features, the glial cell reaction features and the fused prognostic related features.

[0218] S306, constructing a double-branch gated classification module

[0219] Optic nerve prognosis evaluation needs to comprehensively consider axonal injury features and glial reaction features, but the contribution modes of the two types of features to the prognosis result are different, and a conventional single-branch classifier is difficult to distinguish such different contributions, and ignores the dynamic interaction between the features, increasing the risk of misjudgment of the prognosis grade.

[0220] The application adopts a double-branch gated classification network, extracts discriminant information of the axonal injury features and the fused prognostic related features through independent branches, and the specific steps are as follows:

[0221] 1) Double-branch prognosis probability prediction

[0222] The global average pooling is performed on the axon injury features and the features related to fusion prognosis respectively, and then the prediction probability of the axon injury branch and the fusion branch is generated by using a Softmax function through a weight matrix and a bias vector, which is represented as:

[0223]

[0224] In the formula, represents the prediction probability of the axon injury branch for the first category, with a dimension of ;

[0225] represents the prediction probability of the fusion prognosis branch for the first category, with a dimension of ;

[0226] represents a global average pooling operation, which compresses the feature tensor into a channel vector (input dimension , output );

[0227] is the axon branch weight matrix, which is a trainable parameter, with a dimension of ;

[0228] is the number of channels of the axon injury feature ;

[0229] is the fusion branch weight matrix, which is a trainable parameter, with a dimension of ;

[0230] is the number of channels of the fusion prognosis related feature ;

[0231] is the axon branch bias vector, which is a trainable parameter;

[0232] is the fusion branch bias vector, which is a trainable parameter;

[0233] is a normalized exponential function.

[0234] 2) Feature saliency gating fusion

[0235] The global average pooling results of the axon injury features and the fusion prognosis related features are spliced, the gating weight is calculated by using a weight vector and a Sigmoid function, and then the prediction probability of the two branches is weighted and fused to obtain the final prediction probability, which is represented as:

[0236]

[0237]

[0238] wherein, is the gating weight of axon branch, is a scalar;

[0239] is the feature saliency gating weight vector, dimension is a trainable parameter;

[0240] is the transpose of

[0241] denotes the vector concatenation operation, the concatenated features are 2 groups;

[0242] denotes the prediction probability of the model for the

[0243] S307, loss function calculation

[0244] There is a significant class imbalance problem in the optic nerve exosome prognosis evaluation, and the conventional cross-entropy loss function will dominate the majority class gradient, so that the model is not sensitive to the minority class prognosis level, and doctors have subjective confidence differences in the labeling process, and the conventional loss function cannot effectively fuse such prior information.

[0245] The application adopts a dynamic focal point loss function, dynamically adjusts the difficult sample weight through a time-dependent adjustment factor, and adopts a prognosis confidence loss to fuse artificial labeling uncertainty, and the specific steps are as follows:

[0246] 1) Calculate the classification loss

[0247] The dynamic focal point adjustment factor is used to calculate the classification loss based on the prediction probability and the true label, the weight of the difficult sample is dynamically adjusted, and it is represented as:

[0248]

[0249] wherein, denotes the classification loss, which is used to optimize the prognosis level prediction;

[0250] denotes the prediction probability of the model for the

[0251] denotes the class index, which is valued in , corresponding to the prognosis level I, II, III;

[0252] ​​​ denotes the dynamic focal point adjustment factor of the i-th iteration, which increases with the training step number, and is calculated as

[0253] is the initial focal point factor, which is preferably set as

[0254] is the adjustment coefficient, which is preferably set as

[0255] is the training step number index;

[0256] is the total training step number;

[0257] denotes the indicator function, which takes the value of 1 when the true label is equal to the i-th class; is the true label;

[0258] is the logarithmic function, and the default base is the natural constant.

[0259] 2) Calculate the prognosis confidence loss

[0260] Calculate the L2 norm square error between the model prediction probability distribution and the doctor's labeled confidence vector to obtain the prognosis confidence loss, which aligns the model prediction with the preset confidence, and is represented as:

[0261]

[0262] In the formula, denotes the prognosis confidence loss, which is used to align the model prediction with the doctor's confidence;

[0263] denotes the number of samples;

[0264] denotes the preset confidence vector of the i-th sample, with a dimension of is the confidence vector provided during manual labeling, such as [0.9, 0.1, 0.0], which indicates that the confidence of the prognosis level evaluation for this sample is 90% for level I, 10% for level II, and 0% for level III;

[0265] denotes the model prediction probability distribution of the i-th sample, with a dimension of

[0266] denotes the model prediction probability distribution of the i-th sample, with a dimension of ​​​​​The prediction probability distribution of a sample output, such as [0.8, 0.15, 0.05], indicates that the model predicts that the sample belongs to Class I with a probability of 80%, belongs to Class II with a probability of 15%, and belongs to Class III with a probability of 5%.

[0267] 3) Calculate the total loss function

[0268] The classification loss, confidence loss, and orthogonal constraint loss are weighted and summed to obtain the total loss function, which is used for model optimization, and is represented as:

[0269]

[0270] In the formula, The total loss function is used for overall model optimization.

[0271] It should be noted that the coefficients 0.5 and 0.2 are empirical weights used to balance the influence of different loss terms.

[0272] S308, iterative training and parameter updating of prognosis evaluation model

[0273] The process of iterative training and parameter updating of the prognosis evaluation model is as follows:

[0274] 1) After randomly initializing the trainable parameters of the model, iterative optimization is performed using the mini-batch gradient descent method, and the Adam optimizer is used as the optimizer;

[0275] 2) Each iteration includes forward propagation and backward propagation:

[0276] In the forward propagation stage, the interpolated three-dimensional MRI voxel data of the input batch of samples is sequentially passed through the multi-modal feature projection and subtle damage response enhancement module, the deformable subtle feature extraction module, the multi-scale context aggregation module, the channel-space dual-path attention module, the feature decoupling and enhancement module, and the dual-branch gated classification module.

[0277] In the backward propagation stage, the gradient is calculated based on the total loss function Loss, and all trainable parameters are updated.

[0278] During training, the learning rate is set to 0.001, and the stopping condition is based on the performance monitoring of the validation set. If the validation loss does not decrease for 10 consecutive iterations, or if the maximum number of iterations, such as 10,000 iterations, is reached, the training is terminated and the optimal model parameters are saved.

[0279] In one embodiment, the accuracy of different methods in each prognosis level is compared and analyzed, and the accuracy of different technical methods in three prognosis levels is compared, and the results are as follows Figure 7The experiment compares a conventional convolutional neural network, a residual network, a U-Net network, and the method proposed in the application. The accuracy index measures the ability of the model to correctly predict the prognosis, and the higher the value, the better the performance. The experimental results show that the method proposed in the application achieves the highest accuracy in all prognosis levels, especially in the prediction of severe injury levels, which reflects the overall superiority of the technical solution. The conventional convolutional neural network performs well in the mild injury level, but as the injury degree increases, its performance decreases significantly. The deformable fine feature extraction module and the multi-scale context aggregation module in the application can better capture the morphologically variable injury features, thereby performing well in handling complex cases.

[0280] In the embodiment, based on the anisotropic Gaussian interpolation method, high-resolution details are maintained in the intra-layer direction, while the tissue continuity is enhanced through weak interpolation in the inter-layer direction, which can effectively preserve the micron-level structural features of the retinal nerve fiber layer; and through the cross-modal projection matrix and the fine injury response factor mechanism, the complementary information in the multi-modal images is fully fused, and the fine injury area of the weak signal is enhanced, avoiding the key lesion features being covered by the background information; at the same time, the deformable convolution and the multi-scale context aggregation mechanism are adopted, so that the convolution kernel can adaptively fit the morphologically complex injury area, and the representation ability of the model to different spatial scale injuries is improved through multi-scale feature weighting; the feature decoupling and orthogonal constraint are used to separate the axon injury features and the glial cell response features, and combined with the double-branch gated classification structure and the dynamic loss function design, the accuracy of the prognosis evaluation is improved, and the result interpretability is enhanced.

[0281] The embodiment of the application also proposes an exosome and optic nerve injury prognosis evaluation system based on the exosome and optic nerve injury prognosis evaluation method as described above, comprising:

[0282] An optic nerve MRI data acquisition and annotation module is used to acquire optic nerve MRI data and perform manual annotation, and the annotation categories include prognosis levels;

[0283] An anisotropic Gaussian interpolation processing module is used to perform adaptive interpolation processing on the original three-dimensional MRI voxel data through an anisotropic Gaussian interpolation kernel function, to retain high-resolution details in the XY plane using a smaller standard deviation, to realize weak interpolation in the Z-axis using a larger standard deviation, and to obtain interpolated three-dimensional MRI voxel data through local neighborhood weighted summation;

[0284] a prognosis evaluation model construction and training module, which comprises a multi-modal feature projection and subtle damage response enhancement module, a deformable subtle feature extraction module, a multi-scale context aggregation module, a channel-space dual-path attention module, a feature decoupling and enhancement module, a double-branch gated classification module, a loss function calculation module, a prognosis evaluation model iterative training and parameter updating module;

[0285] an exosome and optic nerve injury prognosis evaluation module, which is used for evaluating the prognosis of exosome and optic nerve injury based on the trained model, and outputting a prediction probability distribution representing the probability that the patient belongs to a prognosis grade.

[0286] In addition, an embodiment of the present application further provides a computer readable storage medium, wherein the computer readable storage medium stores program instructions of an exosome and optic nerve injury prognosis evaluation method, and the program instructions of the exosome and optic nerve injury prognosis evaluation method can be executed by one or more processors to implement the steps of the exosome and optic nerve injury prognosis evaluation method.

[0287] The above-described embodiments are only used to describe the preferred embodiments of the present application, and are not used to limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope of the present application defined by the claims.

Claims

1. An exosome and optic nerve injury prognosis evaluation method, characterized by, The method comprises the following steps: S1, optic nerve MRI data acquisition and labeling; It includes collecting optic nerve MRI data and manually labeling, and the labeling categories include prognosis grade; S2, anisotropy Gaussian interpolation processing of optic nerve MRI data; Specifically, based on an anisotropy Gaussian interpolation kernel function, the original three-dimensional MRI voxel data is adaptively interpolated; The adaptive interpolation of the original three-dimensional MRI voxel data based on the anisotropy Gaussian interpolation kernel function includes: using a small standard deviation to retain high-resolution details in the XY plane, using a large standard deviation to achieve weak interpolation in the Z axis, and obtaining interpolated three-dimensional MRI voxel data by locally weighted summation; S3, prognosis evaluation model construction and training; the model adopts a double-branch gated classification network, extracts axon damage features and discriminates information of fused prognosis related features through independent branches; a dynamic focal loss function is used in combination with a prognosis confidence loss for training, and a time-related adjustment factor is used to dynamically adjust the weight of difficult samples; the prognosis evaluation model construction further includes constructing a feature decoupling and enhancement module, which decouples the coupled features into axon damage related features and glial cell response related features by using a prognosis factor decoupling layer and an orthogonal constraint loss; S4, exosome and optic nerve injury prognosis evaluation; the processed optic nerve MRI data of the patient is input into the trained prognosis evaluation model to obtain a prediction probability distribution, which represents the probability that the patient belongs to a prognosis grade.

2. The exosome of claim 1, wherein the method for evaluating the prognosis of the optic nerve injury is characterized by, The prognosis evaluation model construction further includes constructing a multi-modal feature projection and subtle damage response enhancement module, using a cross-modal projection matrix to perform feature transformation and fusion on the T1, T2 and DWI modal data processed by anisotropy Gaussian interpolation, and using a gradient amplitude based subtle damage response factor to realize subtle damage feature enhancement.

3. The exosome of claim 2, wherein the method for evaluating the prognosis of the optic nerve injury is characterized by, The subtle damage feature enhancement further includes adaptively weighting the fused multi-modal feature tensor using a subtle damage response factor, retaining the original features in the damage area, and enhancing the signal in the non-damage area through power operation to obtain damage enhanced features.

4. The exosome of claim 1, wherein the method for evaluating the prognosis of the optic nerve injury is characterized by, The prognosis evaluation model construction further includes constructing a deformable subtle feature extraction module, introducing a deformable convolution mechanism based on standard 3D convolution, learning an offset and a modulation factor for each convolution sampling point, so that the convolution kernel can adaptively adjust the sampling position according to the input features and combine the constraint offset range to prevent excessive deformation.

5. The exosome of claim 1, wherein the method for assessing prognosis of optic nerve injury is characterized by, The prognosis evaluation model construction further includes constructing a multi-scale context aggregation module, which fuses feature responses under different receptive fields and performs global average pooling on the response features of each scale to obtain a channel vector, calculates adaptive weight coefficients through a weight vector and a Softmax function, and weightedly fuses multi-scale features to obtain multi-scale context features.

6. The exosome of claim 1, wherein the method for assessing prognosis of optic nerve injury is characterized by, The prognosis evaluation model construction further includes constructing a channel-spatial dual-path attention module, which guides channel and spatial attention calculation through exosome distribution prior knowledge, and realizes adaptive weighting of dual-path attention through a gating fusion mechanism.

7. An exosome and optic nerve injury prognosis evaluation system based on the exosome and optic nerve injury prognosis evaluation method according to any one of claims 1 to 6, characterized by Also comprising: An optic nerve MRI data acquisition and labeling module, which is used to acquire optic nerve MRI data and manually label, and the labeling categories include prognosis grades; An anisotropic Gaussian interpolation processing module, which is used to perform adaptive interpolation processing on the original three-dimensional MRI voxel data through an anisotropic Gaussian interpolation kernel function, to retain high-resolution details within the layer using a smaller standard deviation in the XY plane, to achieve weak interpolation using a larger standard deviation in the Z axis, and to obtain interpolated three-dimensional MRI voxel data through local neighborhood weighted summation; A prognosis evaluation model construction and training module, which includes a multi-modal feature projection and subtle damage response enhancement module, a deformable subtle feature extraction module, a multi-scale context aggregation module, a channel-space dual-path attention module, a feature decoupling and enhancement module, a double-branch gated classification module, a loss function calculation module, a prognosis evaluation model iterative training and parameter updating module; an exosome and optic nerve injury prognosis evaluation module, which is used to evaluate the prognosis of exosomes and optic nerve injury based on the trained model, and output the prediction probability distribution, indicating the probability that the patient belongs to the prognosis grade.

8. A computer readable storage medium, characterized by, The computer readable storage medium stores program instructions of an exosome and optic nerve injury prognosis evaluation method, the program instructions of the exosome and optic nerve injury prognosis evaluation method can be executed by one or more processors to implement the steps of the exosome and optic nerve injury prognosis evaluation method as claimed in any one of claims 1 to 6.

Images