Anti-NMDAR encephalitis clinical prognosis evaluation method based on artificial intelligence

Abstract

The invention relates to an anti-NMDAR encephalitis clinical prognosis evaluation method based on artificial intelligence, and belongs to the technical field of artificial intelligence and data processing. The method comprises the following steps: acquiring multi-modal nerve image data of a patient; performing fusion preprocessing on the multi-modal neural image data by adopting a tensor decomposition fusion strategy, and reserving cross-modal spatial correlation through low-rank constraint to obtain a fused output tensor; carrying out focus perception anisotropic diffusion filtering on the fused output tensor to obtain an output image after diffusion filtering; an anti-NMDAR encephalitis clinical prognosis evaluation model is constructed, the output image after diffusion filtering is input into the model for training, an Adam adaptive optimizer is adopted to optimize the training process, and finally a trained model is obtained; and inputting a to-be-evaluated output image after diffusion filtering into the trained model to obtain an evaluation classification result. The identification and classification capability of the model on the focus can be enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and data processing technology, and specifically relates to an artificial intelligence-based clinical prognosis assessment method for anti-NMDAR encephalitis. Background Art

[0002] Anti-NMDAR encephalitis, an acutely progressive autoimmune encephalitis, presents with highly heterogeneous clinical manifestations (e.g., psychiatric and behavioral abnormalities, epilepsy, movement disorders), and rapidly progresses. Accurate prognostic assessment is crucial for developing personalized immunotherapy and intensive care plans, directly impacting the patient's neurological recovery and survival rate. However, traditional assessment methods rely heavily on physician experience and single biomarkers (e.g., cerebrospinal fluid antibody titers), lacking the ability to dynamically integrate multidimensional clinical data (imaging, laboratory parameters, and medical records). This results in highly subjective predictions and delayed timeliness, making it difficult to meet the clinical needs of early intervention. Existing methods, particularly for critical scenarios such as prognostic stratification of critically ill patients and relapse risk warning, have significant blind spots, necessitating the use of intelligent technologies to overcome these assessment bottlenecks.

[0003] Although artificial intelligence has made progress in prognostic models for fields such as oncology, existing technologies face the following challenges when applied to NMDAR encephalitis: Extensive multimodal fusion: Conventional concatenation or averaging strategies ignore the spatial-temporal correlations between MRI lesions, EEG rhythms, and serum markers in encephalitis, resulting in loss of critical cross-correlation information; Lesion feature degradation: Isotropic filtering oversmoothes the edges of small inflammatory lesions (e.g., marginal gyral edema) during denoising, weakening the discriminative power of imaging features; Model initialization mismatch: Random initialization of convolutional kernels struggles to adapt to the spatial heterogeneity of encephalitis lesions (e.g., morphological differences between basal ganglia and cortical lesions), reducing feature extraction efficiency; and Inflexible pooling: Homogenized pooling fails to prioritize the fine structural features of lesion edges, resulting in loss of critical information for classification. These deficiencies collectively limit the model's ability to interpret the complex pathological features of anti-NMDAR encephalitis, necessitating the development of a dedicated technical framework for lesion perception. Summary of the Invention

[0004] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions: The present invention provides an artificial intelligence-based method for evaluating clinical prognosis of anti-NMDAR encephalitis, comprising the following steps: S1. Obtain multimodal neuroimaging data from patients; S2. Use tensor decomposition and fusion strategy to perform fusion preprocessing on multimodal neuroimaging data, preserve cross-modal spatial correlation through low-rank constraints, and obtain the fused output tensor; S3. obtaining an output image after diffusion filtering by performing lesion-aware anisotropic diffusion filtering on the fused output tensor; S4. constructing a clinical prognosis evaluation model for anti-NMDAR encephalitis, the model comprising a dynamic convolution module, a spatial feature decoupling module, a gated pooling module and a fully connected classifier; the output image after diffusion filtering is input into the clinical prognosis evaluation model for anti-NMDAR encephalitis for training, and the training process is optimized by using an Adam adaptive optimizer, and finally a trained model is obtained; S5. inputting the output image after diffusion filtering to be evaluated into the trained model to obtain an evaluation classification result.

[0005] Further, step S1 specifically comprises: acquiring multi-modal neural image data of a patient by a magnetic resonance imaging system, manually labeling the multi-modal neural image data, and the labeling categories include good prognosis and poor prognosis; the multi-modal neural image data includes T1 weighted sequence, T2 weighted sequence and FLAIR sequence, and the multi-modal neural image data is stored in DICOM format; a cross-modal registration technology is used to align the three types of sequences to MNI152 standard space, and finally generate three-dimensional image data with consistent spatial dimensions; The three-dimensional image data with consistent spatial dimensions is defined as a four-dimensional image tensor , including height , width , depth and mode number four dimensions.

[0006] Further, step S2 specifically comprises: S21. Tensor decomposition fusion: performing product operation of specific dimensions between multi-modal factor matrix and four-dimensional image tensor, and combining residual tensor to realize feature fusion, the feature fusion operation is based on the factor matrix corresponding to the height, width and depth three dimensions, and finally generates a fused output tensor, which is expressed as follows: , wherein, represents the fused output tensor; represents n-mode tensor product operation, ; represents the factor matrix of the first dimensional space, corresponding to the height dimension of the four-dimensional image tensor; represents the factor matrix of the second dimensional space, corresponding to the width dimension of the four-dimensional image tensor; represents the factor matrix of the third dimensional space, corresponding to the depth dimension of the four-dimensional image tensor; represents the residual tensor; S22. Optimize the factor matrix: The factor matrix is ​​solved by minimizing the reconstruction error and imposing a sparsity constraint. The solution utilizes the Frobenius norm to control the difference between the original image tensor and the fused output tensor. The L21 norm of the factor matrix is ​​used to achieve row sparsification. The sparsity control coefficient is used to adjust the intensity of feature selection in the lesion region.

[0007] Furthermore, step S3 specifically includes: S31. Based on the lesion probability map generated by the pre-trained U-Net model, the original gradient amplitude tensor is weighted enhanced. The edge enhancement coefficient is used to adjust the gradient amplification strength of the lesion area to generate an enhanced gradient amplitude tensor. The formula is as follows: , in, Represents the gradient magnitude tensor after fusion; represents the spatial gradient operator; represents the lesion probability map; represents the edge enhancement coefficient; represents element-wise multiplication; represents the enhanced gradient magnitude tensor; It means multiplying the edge enhancement coefficient in scalar format with each element in the lesion probability map in matrix format; S32. Based on the enhanced gradient amplitude tensor, a nonlinear diffusion coefficient function is constructed. The nonlinear diffusion coefficient function uses the global gradient amplitude median, edge threshold parameter, and edge sharpness parameter to calculate the diffusion coefficient value corresponding to each gradient amplitude. The formula is as follows: , in, Represents the gradient magnitude variable, which is the enhanced gradient magnitude tensor The specific value of each element in; represents the median function; represents the edge threshold parameter; represents the edge sharpness parameter; represents the diffusion coefficient function; S33. Anisotropic Diffusion: Solve the diffusion equation guided by the enhanced gradient amplitude to filter the fused image. This process is iterative within the virtual diffusion time and terminates when the maximum difference between two adjacent iterations is less than the set threshold. The final output is the diffusion-filtered image.

[0008] Furthermore, step S33 specifically includes: Output image after diffusion filtering By numerically solving partial differential equations Get, when initialized, Output image after diffusion filtering of iterations Set to the fused output tensor , and then perform iterative solution to define the discretized virtual diffusion time , and define the step size as , updated using the explicit Euler method, the update method is expressed as: , in, Indicates the The output image after the diffusion filtering of the iterations; Indicates the The output image after the diffusion filtering of the iterations; Indicates step size; definition The mathematical model of the diffusion process .

[0009] Furthermore, in step S4, the dynamic convolution module gives the convolution kernel lesion prior knowledge by initializing the dynamic convolution kernel parameters, uses anisotropic depthwise separable convolution to decompose the feature extraction process along the three orthogonal directions of x-axis, y-axis, and z-axis, and uses the directional attention map to enhance the main response of the lesion: Dynamic convolution kernel parameter initialization: Calculate the average value of the diffusion filtered images of healthy samples and diseased samples respectively to obtain the healthy template and lesion template. Subtract the lesion template from the healthy template to obtain the basic convolution kernel parameters. Calculate the Sigmoid weighted expected value of the diffusion image difference between random healthy samples and random diseased samples and multiply it by the scaling factor to generate the lesion difference enhancement term. Add the basic convolution kernel parameters and the lesion difference enhancement term to obtain the final convolution kernel initial parameter tensor. Anisotropic depthwise separable convolution: Calculate the output image after diffusion filtering along the three-dimensional spatial coordinate axis The partial derivative of the gradient field tensor in the corresponding direction is obtained to represent the rate of change of the image in that direction. For the The gradient field tensor in the direction, Represents the directional variables of the x-axis, y-axis, and z-axis in three-dimensional space; uses the temperature-regulated Softmax function and adopts the temperature parameter to control the sharpness of the weight distribution, converting the gradient field tensors in the three directions into the corresponding directional attention weight map; the formula is as follows: , in, express Attention weight map of direction; represents the temperature parameter; Indicates the difference from The direction variables of the x-axis, y-axis, and z-axis in the three-dimensional space, Representation Sum Index Traverse the three spatial directions of x-axis, y-axis and z-axis; Indicates along The gradient field tensor of the direction; Represents the natural exponential function; the output image after diffusion filtering along three directions The weighted input image performs a one-dimensional convolution operation, using the corresponding trainable convolution kernel parameters in each direction to generate the feature tensor of the convolution output in that direction; the feature tensors output by the convolution along the x-axis, y-axis, and z-axis are spliced ​​along the feature dimension to generate a multi-directional convolution feature tensor.

[0010] Furthermore, in step S4, the spatial feature decoupling module decouples the lesion morphological features through 3D convolution and curvature-driven spatial attention mechanism, specifically: performing a three-dimensional convolution operation on the multi-directional convolution feature tensor, using the Gaussian error linear unit activation function combined with nonlinearity, and at the same time, using the Laplace operator to calculate the curvature of the input feature, generating a spatial attention map through convolution adjustment and Sigmoid function, and multiplying the multi-directional convolution feature tensor and the spatial attention map element by element to obtain the decoupled spatial feature tensor.

[0011] Furthermore, in step S4, the gated pooling module performs gated weighted down-sampling on the feature map through the lesion difference-aware pooling operation to retain the lesion sensitive area information, specifically: using the lesion template generated in the dynamic convolution kernel parameter initialization stage to calculate the similarity between the local area of ​​the decoupled spatial feature tensor and the lesion template; using the sharpness adjustment parameter to control the sharpness of the weight distribution, combining the similarity with the feature curvature of the decoupled spatial feature tensor to generate an adaptive gating weight; using the gating weight to weight the local area features of the decoupled spatial feature tensor, and then performing the maximum pooling operation to finally obtain the pooled feature tensor.

[0012] Furthermore, in step S4, the fully connected classifier includes a fully connected layer and an output layer to achieve feature dimensionality reduction and high-level abstract feature extraction, specifically: the pooled feature tensor is flattened into a one-dimensional feature vector, the one-dimensional feature vector is passed through multiple fully connected layers for data transmission, and finally a compact low-dimensional feature vector is output; each fully connected layer is composed of trainable weights and bias parameters, and uses the ReLU nonlinear activation function for feature transformation; the compact low-dimensional feature vector passes through the output layer to obtain a prognostic evaluation classification category; the output layer is also a fully connected layer, using the Softmax function.

[0013] Furthermore, in step S4, the loss function uses a cross-entropy loss function to measure the difference between the prognosis category probability distribution predicted by the model and the true label. The cross-entropy loss function quantifies the degree of classification error and serves as the optimization target.

[0014] The advantages of the present invention are: The present invention captures high-order information through low-rank constraints and residual tensors, retains cross-modal spatial correlation, and uses the sparse constraints of the factor matrix to achieve lesion-sensitive feature selection, effectively avoiding the loss of lesion boundary information in traditional channel splicing methods; in view of the diffuse distribution characteristics of encephalitis lesions, an anisotropic diffusion filtering mechanism guided by gradient field is adopted, and a gradient enhancement strategy weighted by lesion probability map is used to retain the edges of tiny lesions while suppressing noise, overcoming the lesion blurring problem caused by conventional isotropic filtering; convolution kernel initialization parameters are generated based on the differences between healthy and diseased samples, and the convolution kernel is given prior knowledge of the lesion. By calculating the weighted expectation of the differences between healthy and diseased samples, the convolution kernel obtains a stronger feature response in the lesion area, accelerates model convergence, and avoids the blindness of traditional random initialization; an adaptive gated pooling mechanism based on lesion template similarity is adopted, and the contribution weight of the pooling area is dynamically adjusted to strengthen the extraction of lesion-related features, enhance feature contribution in areas with high lesion similarity, and retain details at the edge of the lesion, overcoming the problem of information loss in traditional pooling operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0016] Figure 1 is a flow chart of the steps of the method of the present invention; Figure 2 These are the T1-weighted sequence, T2-weighted sequence, and FLAIR sequence images of the present invention; Figure 3 To investigate the impact of different feature fusion methods on the accuracy of prognostic assessment; Figure 4 The lesion probability map and the enhanced gradient amplitude map of the present invention; Figure 5 Characteristic graphs of the healthy template and the lesion template of the present invention; Figure 6 It is a feature map after the convolution operation is performed using the initial parameter tensor of the convolution kernel in the present invention; Figure 7 This is a comparison chart of training losses for different convolution kernel initialization methods of the present invention; Figure 8 The robustness of different methods of the present invention under noise conditions. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments derived by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0018] Example 1 In this embodiment, Figure 1 As shown, the present invention provides an artificial intelligence-based method for evaluating clinical prognosis of anti-NMDAR encephalitis, which specifically includes the following steps: S1. Obtain multimodal neuroimaging data from patients; The patient's multimodal neuroimaging data were collected by magnetic resonance imaging (MRI) system, including T1-weighted sequence, T2-weighted sequence and FLAIR sequence, such as Figure 2 shown.

[0019] All sequences were acquired using a three-dimensional whole-brain scan, with uniform scanning parameters set to: 1 mm slice thickness, 256×256 matrix size, 1 mm³ isotropic voxel resolution to ensure spatial alignment. T1 sequences were acquired using a fast gradient echo sequence with a TR / TE ratio of 8.2 ms / 3.2 ms and a flip angle of 12°, highlighting the contrast between gray and white matter anatomical structures and enabling precise identification of brain tissue morphology. T2 sequences were acquired using a fast spin echo sequence with a TR / TE ratio of 4800 ms / 110 ms, highlighting the signal from cerebrospinal fluid and edema within the lesion area, enhancing the visualization of fluid-related lesions. FLAIR sequences used a fluid-attenuated inversion recovery technique with a TR / TE ratio of 9000 ms / 140 ms / 2250 ms to suppress high signal interference from cerebrospinal fluid and improve the detection rate of intraparenchymal lesions.

[0020] During the scan, the patient's head is fixed to a custom coil, and axial, sagittal, and coronal images are acquired simultaneously. The raw data is stored in DICOM format, and cross-modal registration technology is used to align the three sequences to the MNI152 standard space, ultimately generating 3D image data with consistent spatial dimensions. This cross-modal registration technology can be a pre-trained cross-modal image registration model, such as a pre-trained generative adversarial network cross-modal image registration model.

[0021] Furthermore, the three-dimensional image data with consistent spatial dimensions are defined as a four-dimensional image tensor, which contains four dimensions, corresponding to the height, width, depth and modality number of the image, and is used to represent multimodal brain structure information. The four-dimensional image tensor is defined as , the height of the four-dimensional image tensor is , the width of the four-dimensional image tensor is , the depth of the four-dimensional image tensor is , the number of modes of the four-dimensional image tensor is .

[0022] Furthermore, in order to enable the model to better learn and identify lesion characteristics, the collected imaging data is annotated. This annotation work is completed by professional radiologists, who assign a unique label to each data sample through manual annotation, such as labeling it into two categories: good prognosis and poor prognosis.

[0023] S2. Use tensor decomposition and fusion strategy to perform fusion preprocessing on multimodal neuroimaging data, preserve cross-modal spatial correlation through low-rank constraints, and obtain the fused output tensor; Neuroimaging data has three-dimensional spatial heterogeneity and multimodal complementary characteristics. Conventional channel splicing will ignore the spatial correlation between modalities and easily lead to loss of lesion boundary information.

[0024] This paper adopts a tensor decomposition and fusion strategy to preserve cross-modal spatial correlation through low-rank constraints. The specific steps are as follows: 1) Perform tensor decomposition and fusion Feature fusion is achieved by multiplying the multimodal factor matrix with the four-dimensional image tensor in a specific dimension and combining it with the residual tensor. The feature fusion operation is based on the factor matrices corresponding to the three dimensions of height, width, and depth, and finally generates the fused output tensor, which is expressed as: , Where, is the output tensor after fusion; is the n-module tensor product operation, that is, is a 1-module tensor product operation, is a 2-module tensor product operation, It is a 3-module tensor product operation; For the The factor matrix of the dimensional space, that is, is the factor matrix of the first dimension space, corresponding to the height dimension of the four-dimensional image tensor, is the factor matrix of the second-dimensional space, corresponding to the width dimension of the four-dimensional image tensor, is the factor matrix of the third-dimensional space, corresponding to the depth dimension of the four-dimensional image tensor; It is a residual tensor used to capture high-order residual information in the multimodal image tensor that cannot be fully represented by low-rank tensor decomposition.

[0025] 2) Optimize factor matrix The factor matrix is ​​solved by minimizing the reconstruction error and imposing a sparsity constraint optimization problem. The Frobenius norm is used to control the difference between the original image tensor and the fused output tensor, and the L21 norm of the factor matrix is ​​used to achieve row sparsification. The sparsity control coefficient is used to adjust the intensity of the lesion area feature selection. The optimization objective function is expressed as: , Where, is the Frobenius norm, The term is used to control the reconstruction error; Represents a variable and Perform minimization operation to characterize the and The value of , which minimizes the value of the objective function, is a set of factor matrices, i.e. ; For the The factor matrix of the dimensional space; is the L21 norm, Characterize the L21 norm of the factor matrix to achieve row sparsification; is the sparsity control coefficient, which adjusts the feature selection strength of the lesion area, such as .

[0026] It should be noted that the optimization process of the objective function uses an iterative algorithm to solve, such as gradient descent or alternating least squares, and finally obtains the set of optimal factor matrices and the residual tensor .

[0027] It should also be noted that when performing tensor decomposition and fusion, the factor matrix explicitly models the spatial correlation of the depth, height, and width dimensions of the four-dimensional image tensor through a low-rank structure shared across modalities, effectively achieving spatial correlation preservation.

[0028] It should also be noted that the residual tensor It is a variable learned during the optimization process, not predefined, and is obtained by minimizing the optimization objective function and the factor matrix Solve the optimization problem together In the residual tensor As a free variable, used to capture the 4D image tensor The high-order residual information that cannot be fully represented by the low-rank tensor decomposition, such as noise, subtle lesion features or nonlinear correlation between modalities, is used to reduce the reconstruction error, improve the integrity of feature fusion, and avoid the loss of lesion boundary information due to low-rank constraints in areas with weak spatial correlation between modalities. In addition, the output tensor after fusion is During the calculation of By means of residual compensation, the fused output tensor Closer to the four-dimensional image tensor .

[0029] It should also be noted that the L21 norm of the factor matrix The row vectors of the factor matrix are forced to be sparse, the feature weights of non-lesion areas are automatically suppressed, and lesion-sensitive feature selection is achieved while retaining cross-modal spatial correlation, overcoming the boundary information loss problem of channel splicing.

[0030] It should also be noted that in the optimization objective function, Characterize the reconstruction error term and use the Frobenius norm to measure the four-dimensional image tensor and the fused output tensor The difference between them ensures that the fusion result is faithful to the original data. Characterize the sparse constraint term, apply the L21 norm to the factor matrix, and control the coefficients by sparsity The feature weights of non-lesion areas are forced to approach zero, thereby highlighting the lesion area.

[0031] S3. performing lesion-aware anisotropic diffusion filtering on the fused output tensor to obtain a diffusion-filtered output image; Neuroimaging data are easily affected by noise, and conventional isotropic filtering can blur tiny lesions.

[0032] In view of the diffuse distribution characteristics of encephalitis lesions, the present invention adopts a gradient field-guided diffusion filtering mechanism to suppress noise while preserving the lesion edges. The specific steps are as follows: 1) Calculate the enhanced gradient amplitude Based on the lesion probability map generated by the pre-trained U-Net model, the original gradient amplitude tensor is weighted enhanced, and the edge enhancement coefficient is used to adjust the gradient amplification intensity of the lesion area to generate an enhanced gradient amplitude tensor, which is expressed as: , Where, is the fused gradient amplitude tensor, representing the fused output tensor The gradient magnitude tensor of ; It is a spatial gradient operator, a mathematical differential operator that represents the gradient and is used to calculate the three-dimensional directional derivative of each element. In discrete image data, it is implemented through numerical difference approximation to quantify local changes in the image, thereby achieving edge detection and feature enhancement. is the lesion probability map, generated by the pre-trained U-Net model, with a range of , represents the probability that each voxel belongs to a lesion; is the edge enhancement coefficient, which adjusts the gradient amplification intensity of the lesion area, such as ; is element-wise multiplication; is the enhanced gradient magnitude tensor; In one embodiment, the edge enhancement coefficient in scalar format is multiplied by each element in the lesion probability map in matrix format. Figure 4 As shown, the left side is the lesion probability map, and the right side is the enhancement gradient amplitude map.

[0033] It should be noted that the fused gradient amplitude tensor Must be based on the fused output tensor Calculation is not based on the original image to preserve the cross-modal spatial correlation. The calculation process is as follows: first, the fused output tensor Applying the spatial gradient operator , get the three-dimensional directional derivative of each element, then modulate the gradient vector of each element, that is, calculate its Euclidean norm, and get the fused gradient amplitude tensor .

[0034] It should also be noted that the lesion probability map Generated using a pre-trained U-Net model. U-Net is an encoder-decoder architecture widely used in medical image segmentation that can capture multi-scale contextual information. The pre-trained U-Net model is trained on an independent dataset and uses annotated lesion masks and a loss function to optimize weights. After training, the model is generalized to this task to generate lesion probability maps. , in the lesion probability map In the generation process of , the input data is the neuroimaging data before diffusion filtering, which can be T1, T2, FLAIR sequence, or the fusion representation of T1, T2, FLAIR sequence. Preferably, the fused output tensor is selected. As the input of the pre-trained U-Net model, U-Net further outputs a probability map with the same spatial dimension as the input, namely the lesion probability map , where each value is normalized to .

[0035] It should also be noted that in Middle, lesion probability map It can achieve weighted amplification of the gradient of the lesion area and preserve the edges of tiny lesions.

[0036] 2) Define edge-aware diffusion coefficient Based on the enhanced gradient amplitude tensor, a nonlinear diffusion coefficient function is constructed. The nonlinear diffusion coefficient function uses the global gradient amplitude median, edge threshold parameter and edge sharpness parameter to calculate the diffusion coefficient value corresponding to each gradient amplitude, which is expressed as: , Where, is the gradient magnitude variable, which is the enhanced gradient magnitude tensor The specific value of each element in , that is, the enhanced gradient magnitude tensor Each spatial position of takes an independent value; is the median function, The global median gradient magnitude used to calculate the enhanced gradient magnitude tensor; is the edge threshold parameter, which controls the boundary between edge and non-edge areas, such as ; It is the edge sharpness parameter. The larger the value, the sharper the edge will be. For example, ; is the diffusion coefficient function, and the output range is .

[0037] It should be noted that in the diffusion coefficient function, the gradient amplitude variable As an input variable, the diffusion intensity of each position is dynamically adjusted. When it is large, it indicates the edge of the lesion. becomes smaller, suppressing diffusion to preserve the edge, when hours, representing the flat area, It becomes larger, promotes diffusion to smooth noise, makes the diffusion intensity depend on the local gradient, and avoids lesion blurring.

[0038] 3) Perform anisotropic diffusion The diffusion equation guided by the enhanced gradient amplitude is solved to filter the fused image. The process is iteratively performed within the virtual diffusion time and terminated when the maximum difference between two adjacent iteration results is less than the set threshold. Finally, the diffusion filtered image is output.

[0039] Specifically, the output image after diffusion filtering is By numerically solving partial differential equations Get, when initialized, Output image after diffusion filtering of iterations Set to the fused output tensor , and then perform iterative solution to define the discretized virtual diffusion time as , and define the step size as , updated using the explicit Euler method, the update method is expressed as: , Where, For the The output image after the diffusion filtering of the iterations; For the The output image after the diffusion filtering of the iterations; is the step length.

[0040] Furthermore, the iteration is performed within the virtual diffusion time, and the iteration is terminated when the maximum difference between two adjacent iteration results is less than the set threshold. The iteration termination condition is: , is the number of iterations, Represents the maximum value function.

[0041] Furthermore, the final output image after diffusion filtering is .

[0042] It should be noted that it is possible to define The mathematical model of the diffusion process , represents the output image after diffusion filtering With virtual diffusion time In the iterative update process of the output image after diffusion filtering, the diffusion process is realized by numerically solving the partial differential equation.

[0043] It should be noted that the iterative process is an iterative algorithm for numerically solving the diffusion equation. The core is to approximate the solution of the partial differential equation through discrete time steps and stop when the solution is stable, that is, the iteration stops when the change is less than a threshold.

[0044] It should also be noted that the iterative process ensures that the output image after diffusion filtering Smooth and retain edge, virtual diffusion time It has no physical meaning and only controls the number of iterations.

[0045] It should also be noted that the fused gradient amplitude tensor In the calculation process, the lesion probability map The weighted gradient enhancement mechanism enables the diffusion process to automatically reduce the smoothing intensity at the edge of tiny lesions, overcoming the lesion blurring problem of conventional filtering.

[0046] S4. Constructing an anti-NMDAR encephalitis clinical prognosis assessment model, the model comprising a dynamic convolution module, a spatial feature decoupling module, a gated pooling module, and a fully connected classifier; the output image after diffusion filtering is input into the anti-NMDAR encephalitis clinical prognosis assessment model for training, and the training process is optimized using the Adam adaptive optimizer to ultimately obtain a trained model; S401. Build a prognostic assessment model framework The prognostic evaluation model constructed in the present invention adopts a cascade feature learning architecture, which consists of a dynamic convolution module, a spatial feature decoupling module, a gated pooling module and a fully connected classifier.

[0047] The model input is the output image after diffusion filtering First, the convolution kernel is given prior knowledge of the lesion through dynamic convolution kernel parameter initialization. Then, anisotropic depthwise separable convolution is used to decompose the feature extraction process along the three orthogonal directions of x, y, and z axes, and the directional attention map is used to enhance the main response of the lesion. The spatial feature extraction layer is connected to decouple the morphological characteristics of the lesion through 3D convolution and curvature-driven spatial attention mechanism. The feature map is gated and weighted down-sampled through the lesion difference-aware pooling operation to retain the information of the sensitive area of ​​the lesion. The pooled features are passed through the fully connected layer to achieve high-order abstraction and dimensionality reduction. The output layer performs prognostic classification calculations.

[0048] S402, dynamic convolution kernel parameter initialization Traditional convolution kernel random initialization is difficult to adapt to the spatial heterogeneity of lesions in encephalitis images.

[0049] The present invention generates initialization parameters based on the difference between healthy and diseased samples, so that the convolution kernel has prior knowledge of the lesion area. The specific steps are as follows: 1) Calculate healthy and diseased templates Calculate the average value of the diffusion filtered images of healthy samples and diseased samples respectively to obtain the healthy template and the lesion template. Subtract the lesion template from the healthy template to obtain the basic convolution kernel parameters, which are expressed as: , Where, is the basic convolution kernel parameter; is the number of healthy samples; is the number of diseased samples; For the Diffusion filtered image of a healthy sample; For the Diffusion filtered image of a diseased sample.

[0050] It should be noted that The item calculates the average features of the healthy sample image and represents the healthy template. The average feature of the diseased sample image is calculated by the term to represent the lesion template. In one embodiment, the feature map of the healthy template and the lesion template is as follows: Figure 5 shown.

[0051] 2) Calculate the difference-sensitive weighted terms The sigmoid weighted expected value of the diffusion image difference between a random healthy sample and a random diseased sample is calculated and multiplied by the scaling factor to generate the lesion difference enhancement term. The lesion difference enhancement term has a higher weight in the area with significant difference between healthy and diseased samples, which is expressed as: , Where, It is the lesion difference reinforcement item; is a scaling factor that controls the enhancement strength, e.g. ; is the expectation operator, Items are used to calculate cross-sample means; is the diffusion image of a random diseased sample; is the diffusion image of a random healthy sample; is the Sigmoid activation function, Used to implement difference-sensitive weighting.

[0052] It should be noted that the diffusion image of random diseased samples Diffusion images with random healthy samples The acquisition process is to randomly extract samples from the data set, and random sampling makes the lesion difference strengthen the item Generalize to the entire dataset and strengthen the feature response in areas with significant differences between healthy and diseased.

[0053] 3) Synthesize initial convolution kernel parameters The basic convolution kernel parameters are added to the lesion difference enhancement term to obtain the final convolution kernel initial parameter tensor, which has a stronger feature response in the lesion area and is expressed as: , Where, is the initial parameter tensor of the convolution kernel.

[0054] It should be noted that in the lesion difference enhancement item During the calculation process, The larger the value of , the higher the weight is in the area with significant difference between healthy and sick people. The Sigmoid activation function is used to activate the output, ensuring that the weight range is , so that the initial convolution kernel obtains a higher feature response weight in the lesion area, overcoming the blindness problem of random initialization.

[0055] In one embodiment, if a 1×1 convolution operation is used, the feature map after the convolution operation using the initial parameter tensor of the convolution kernel is as follows: Figure 6 shown.

[0056] S403, Anisotropic Depthwise Separable Convolution Conventional 3D convolution is computationally redundant and ignores brain anisotropy.

[0057] The present invention adopts direction-aware convolution kernel and performs weighted fusion through independent convolution paths in three orthogonal directions. The specific process is as follows: 1) Calculate the gradient field in each direction Calculate the output image after diffusion filtering along the three-dimensional space coordinate axis The partial derivative of the gradient field tensor in the corresponding direction is obtained to represent the rate of change of the image in that direction. For the The gradient field tensor in the direction, Variables representing the directions of the x-axis, y-axis, and z-axis in three-dimensional space.

[0058] 2) Generate directional attention map By using the temperature-regulated Softmax function and the temperature parameter to control the sharpness of the weight distribution, the gradient field tensors in the three directions are converted into corresponding directional attention weight maps to enhance the main direction response of the lesion, which is expressed as: , Where, for Direction attention weight map, used to enhance the response to the main direction of the lesion; is the temperature parameter that controls the sharpness of the weight distribution, e.g. =0.5; is the natural exponential function; To distinguish The direction variables of the x-axis, y-axis, and z-axis in the three-dimensional space, Representation Sum Index Traverse the three spatial directions of x-axis, y-axis and z-axis; For the The gradient field tensor of the direction; is a normalized exponential function that converts the gradient strength into a probability distribution.

[0059] It should be noted that The term is the calculation format of the Softmax function, which is calculated by dividing the gradient strength in each direction by It is converted into a probability distribution to realize the normalization of the directional attention weight so that the sum of the weights is 1, and the response in the direction of the maximum gradient is enhanced, that is, the response in the main direction of the lesion is enhanced.

[0060] It should also be noted that The term represents the exponential transformation, which is used to amplify the gradient difference, and the temperature parameter Control distribution sharpness, temperature parameters Smaller values ​​lead to more concentrated weights in the direction of high gradients.

[0061] 3) Perform direction-specific convolution The output image after diffusion filtering is processed along three directions respectively. The weighted input image performs a one-dimensional convolution operation, using the corresponding trainable convolution kernel parameters in each direction to generate the feature tensor of the convolution output in that direction, expressed as: , Where, Indicates along Directional 1D convolution operation; For direction The corresponding trainable convolution kernel parameters, whose initial value is the convolution kernel initial parameter tensor ; The item representation uses the attention weight map to represent the output image after diffusion filtering weighted; for The feature tensor output by the directional convolution.

[0062] 4) Merge multi-directional features The feature tensors output by convolution along the x-axis, y-axis, and z-axis are concatenated along the feature dimension to generate a multi-directional convolution feature tensor, which is expressed as: , Where, Represents a tensor concatenation operation; is the multi-directional convolution feature tensor; It is the feature tensor output by convolution along the x-axis direction, representing the lesion structure information in the x-axis dimension; It is the feature tensor output by convolution along the y-axis direction, representing the lesion structure information in the y-axis dimension; It is the feature tensor output by convolution along the z-axis direction, representing the lesion structure information in the z-axis dimension.

[0063] It should be noted that anisotropic depthwise separable convolution uses 1D convolution in three orthogonal directions instead of standard 3D convolution, which significantly reduces the number of parameters. Through directional decomposition and possible output channel adjustment, the number of parameters is reduced to 1 / 9 (in terms of as an example), reducing computational complexity while retaining anisotropic characteristics.

[0064] It should also be noted that the direction-specific convolution operation is performed through the attention weight map Adaptively enhancing the main direction response of the lesion can preserve the anisotropic structural characteristics while significantly reducing the computational complexity.

[0065] S404, spatial feature extraction A three-dimensional convolution operation is performed on the input feature tensor, and the Gaussian error linear unit activation function is combined with nonlinearity. At the same time, the Laplace operator is used to calculate the curvature of the input feature. The spatial attention map is generated through convolution adjustment and Sigmoid function to strengthen the lesion boundary. Then, the convolution output feature is multiplied element-by-element with the spatial attention map to obtain the decoupled spatial feature tensor, which represents the spatial structural information such as the lesion morphology, expressed as: , Where, It is the decoupled spatial feature tensor, representing spatial structural information such as lesion morphology; is the spatial attention map, which enhances the lesion boundary through the Laplace operator, with a value range of (0, 1) and is calculated as ; is the Gaussian error linear unit activation function; is the spatial convolution kernel weight, used for feature mapping; It is a 3D convolution operation, used to process three-dimensional spatial data; is the multi-directional convolution feature tensor; It is a 1×1×1 convolution operation for feature channel adjustment; Represents the second-order partial derivative about the x-axis. Similarly, represents the second-order partial derivative about the y-axis, represents the second-order partial derivative about the z-axis.

[0066] It should be noted that Characterize the Laplacian operator and calculate the curvature of the input features to locate the edge of the lesion.

[0067] S405. Pooling operation for lesion difference perception Conventional pooling operations treat all regions equally and may easily overlook the local specificity of encephalitis lesions; The present invention adopts an adaptive gated pooling mechanism based on lesion template similarity, and enhances lesion-related feature extraction by dynamically adjusting the contribution weight of the pooling region. The specific steps are as follows: 1) Calculate local lesion similarity Using the lesion template generated in the dynamic convolution kernel parameter initialization stage, the similarity between the local area of ​​the feature map and the lesion template is calculated as the basis of the gating weight, which is expressed as: , Where, For the Local similarity tensor at a spatial location; is the spatial position coordinate index; is the decoupled spatial feature tensor China-Israel Spatial location-centered local blocks; is the basic convolution kernel parameter; is the tensor inner product operation; is the Frobenius norm.

[0068] 2) Generate gated weight map The sharpness adjustment parameter is used to control the sharpness of the weight distribution, and the similarity and feature curvature are combined to generate the adaptive gating weight, which is expressed as: , Where, For the The gating weight value at the position; is the decoupled spatial feature tensor; is the decoupled spatial feature tensor The absolute value tensor of the Laplace operator, which represents the absolute value of the curvature of the input feature; Indicates in Decoupled spatial feature tensor at the spatial position The absolute value of the Laplace operator is used to quantify the strength of local boundary curvature; Adjust the parameter for sharpness. The smaller the value, the sharper the weight distribution. For example, .

[0069] 3) Perform gated weighted pooling After weighting the local area features with the gated weights, the maximum pooling operation is performed, which can be expressed as: , Where, is the feature tensor after pooling in the The value of the spatial position; is the input index set corresponding to the output position in the height dimension, that is, the index range of the pooling window in the height direction; is the set of input indices corresponding to the output positions in the width dimension, that is, the index range of the pooling window in the width direction; is the input index set corresponding to the output position in the depth dimension, that is, the index range of the pooling window in the depth direction; Indicates in Height index belongs to , No. Width index belongs to , No. Depth index belongs to The maximum value operation is taken within the range.

[0070] It should be noted that the gate weight value The similarity between the local area and the lesion template and the boundary curvature information are integrated, so that the pooling process enhances the feature contribution in the area with high lesion similarity, retains the details at the edge of the lesion where the boundary curvature is large, and adjusts the parameters by sharpness. Adjust the sensitivity of lesion region selection to overcome the information loss problem of traditional pooling.

[0071] Furthermore, the pooled feature tensor It is obtained by iteratively applying a gated weighted pooling operation to each position of the output feature map, that is, is the feature tensor after pooling elements.

[0072] S406, fully connected layer feature processing For the pooled feature tensor Perform multiple fully connected layer processing to achieve feature dimensionality reduction and high-level abstract feature extraction.

[0073] Specifically, the feature tensor after pooling First, it is flattened into a one-dimensional feature vector to meet the input requirements of the fully connected layer. The one-dimensional feature vector contains the lesion-related spatial features extracted from the original neuroimaging; The one-dimensional feature vector is then passed through multiple fully connected layers for data transmission. Each fully connected layer consists of trainable weights and bias parameters, and uses the ReLU nonlinear activation function for feature transformation to enhance the expressiveness of the model and extract higher-level abstract features.

[0074] The fully connected layer gradually reduces the feature dimension, achieves feature compression and de-redundancy, and finally outputs a compact low-dimensional feature vector.

[0075] It should be noted that the number and size of fully connected layers can be adjusted according to task requirements to ensure that key lesion information is retained while reducing the risk of overfitting, thereby improving the generalization performance of the model. Preferably, the number of fully connected layers is set to 3, the number of neurons in the first fully connected layer is the same as the number of elements in the one-dimensional feature vector, the number of neurons in the second fully connected layer is set to 200, and the number of neurons in the third fully connected layer is set to 100.

[0076] S407, output layer realizes prognosis assessment classification Based on the feature vector processed by the fully connected layer, the clinical prognosis evaluation and classification of anti-NMDAR encephalitis are realized.

[0077] Specifically, the output layer is also composed of a fully connected layer, and the number of neurons corresponds to the number of prognosis categories. For example, the binary classification task includes two categories: good prognosis and poor prognosis, and the number of neurons in the output layer is 2.

[0078] The output layer receives the feature vector output by the fully connected layer as input, calculates the original score of each category through linear transformation, and uses the Softmax function to convert these scores into probability distributions, indicating the confidence that the input sample belongs to each prognostic category.

[0079] S408. Loss calculation and parameter update The cross entropy loss function is used to measure the difference between the prognostic category probability distribution predicted by the model and the true label. The cross entropy loss function quantifies the degree of classification error and serves as the optimization target.

[0080] During the training iteration, the backpropagation algorithm is used to calculate the gradient of the loss with respect to the parameters of each layer of the network, including the gradient of the convolution kernel weights and the fully connected layer parameters; Then, the Adam adaptive optimizer is used to update all trainable parameters based on the gradient to gradually minimize the loss function and improve the model accuracy; Training continues until a stopping condition is met, such as when the loss value changes by less than a preset threshold in consecutive iterations, indicating convergence, or when a maximum number of training rounds, such as 1000 rounds, is reached.

[0081] S5. Input the output image after diffusion filtering to be evaluated into the trained model to obtain the evaluation classification result.

[0082] Example 2 In this embodiment, Figure 3 As shown in the figure, in order to verify the effectiveness of the multimodal neuroimaging tensor fusion strategy, the performance differences of four fusion methods in the prognosis assessment task were compared experimentally. The bar graph shows the accuracy comparison of channel splicing, simple averaging, conventional tensor decomposition and the technology of the present invention. The experimental results show that the fusion method of the technology of the present invention is significantly better than other methods. The residual tensor captures high-order information in the original image that cannot be fully represented by low-rank decomposition, and the sparse constraint of the factor matrix is ​​used to realize feature selection of the lesion area.

[0083] Example 3 In this embodiment, Figure 7 As shown in the figure, the influence of the convolution kernel initialization strategy on the convergence performance of the model is analyzed, and the loss changes of random initialization, basic convolution kernel initialization and the technology of the present invention during the training process are compared. The curve of the technology of the present invention shows the fastest decline rate and the lowest stable loss value. Its advantage comes from the lesion difference enhancement term. By calculating the weighted expectation of the difference between healthy and diseased samples, the convolution kernel obtains a stronger feature response in the lesion area. The separation point of the three curves in the figure appears in the early stage of training, which proves that the initialization strategy of the technology of the present invention gives the model prior knowledge of lesions, effectively accelerates the feature extraction process, avoids the blindness of random initialization, and provides more reliable lesion characterization capabilities for prognosis evaluation.

[0084] Example 4 This paper summarizes the examples and analyzes the stability of different methods in a noisy environment. The experiment simulates the common image quality fluctuation problems in clinical practice, such as Figure 8 As shown in the figure, the changing trend of the accuracy of each method as the noise level increases is shown by a broken line graph. The horizontal axis represents the noise intensity and the vertical axis represents the classification accuracy. The curve of the method of the present invention shows a slight decline and remains stable at a high level, while the curves of other methods show an obvious landslide. The experimental results show that the lesion-aware anisotropic diffusion filtering operation adopted by the present invention has the effect of enhancing robustness. Through the gradient enhancement strategy of the weighted lesion probability map, the filtering process automatically reduces the smoothing intensity at the edge of the tiny lesion, avoiding feature confusion caused by noise amplification. At the same time, the nonlinear diffusion coefficient function effectively distinguishes the lesion edge from the flat area, suppressing random noise interference while retaining key anatomical structures.

[0085] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for clinical prognosis assessment of anti-NMDAR encephalitis based on artificial intelligence, characterized in that: The following steps are involved: S1. Obtain multimodal neuroimaging data from patients; S2. Use tensor decomposition and fusion strategy to perform fusion preprocessing on multimodal neuroimaging data, preserve cross-modal spatial correlation through low-rank constraints, and obtain the fused output tensor; S3. performing lesion-aware anisotropic diffusion filtering on the fused output tensor to obtain a diffusion-filtered output image; S4. Constructing an anti-NMDAR encephalitis clinical prognosis assessment model, the model comprising a dynamic convolution module, a spatial feature decoupling module, a gated pooling module, and a fully connected classifier; the output image after diffusion filtering is input into the anti-NMDAR encephalitis clinical prognosis assessment model for training, and the training process is optimized using the Adam adaptive optimizer to ultimately obtain a trained model; S5. Input the output image after diffusion filtering to be evaluated into the trained model to obtain the evaluation classification result.

2. The artificial intelligence-based anti-NMDAR encephalitis clinical prognosis assessment method according to claim 1, characterized in that: Step S1 specifically includes: Multimodal neuroimaging data from patients was collected using a magnetic resonance imaging system and manually annotated with good and poor prognosis categories. The multimodal neuroimaging data included T1-weighted sequences, T2-weighted sequences, and FLAIR sequences, and was stored in DICOM format. Cross-modal registration technology was used to align the three sequences to the MNI152 standard space, ultimately generating three-dimensional imaging data with consistent spatial dimensions. Define three-dimensional image data with consistent spatial dimensions as a four-dimensional image tensor , including height ,width ,depth and the number of modes Four dimensions.

3. The artificial intelligence-based anti-NMDAR encephalitis clinical prognosis assessment method according to claim 2, characterized in that: Step S2 specifically includes: S21. Tensor decomposition and fusion: Feature fusion is achieved by multiplying the multimodal factor matrix with the four-dimensional image tensor in a specific dimension and combining it with the residual tensor. The feature fusion operation is based on the factor matrices corresponding to the height, width, and depth dimensions, and ultimately generates a fused output tensor. The formula is as follows: ， in, Represents the fused output tensor; represents the n-module tensor product operation, ; Represents the factor matrix of the first dimension, corresponding to the height dimension of the four-dimensional image tensor; Represents the factor matrix of the second dimension, corresponding to the width dimension of the four-dimensional image tensor; Represents the factor matrix of the third dimension, corresponding to the depth dimension of the four-dimensional image tensor; represents the residual tensor; S22. Optimize the factor matrix: The factor matrix is ​​solved by minimizing the reconstruction error and imposing a sparsity constraint. The solution utilizes the Frobenius norm to control the difference between the original image tensor and the fused output tensor. The L21 norm of the factor matrix is ​​used to achieve row sparsification. The sparsity control coefficient is used to adjust the intensity of feature selection in the lesion region.

4. The artificial intelligence-based anti-NMDAR encephalitis clinical prognosis assessment method according to claim 3, characterized in that: Step S3 specifically includes: S31. Based on the lesion probability map generated by the pre-trained U-Net model, the original gradient amplitude tensor is weighted enhanced. The edge enhancement coefficient is used to adjust the gradient amplification strength of the lesion area to generate an enhanced gradient amplitude tensor. The formula is as follows: ， in, Represents the gradient magnitude tensor after fusion; represents the spatial gradient operator; represents the lesion probability map; represents the edge enhancement coefficient; represents element-wise multiplication; represents the enhanced gradient magnitude tensor; It means multiplying the edge enhancement coefficient in scalar format with each element in the lesion probability map in matrix format; S32. Based on the enhanced gradient amplitude tensor, a nonlinear diffusion coefficient function is constructed. The nonlinear diffusion coefficient function uses the global gradient amplitude median, edge threshold parameter, and edge sharpness parameter to calculate the diffusion coefficient value corresponding to each gradient amplitude. The formula is as follows: ， in, Represents the gradient magnitude variable, which is the enhanced gradient magnitude tensor The specific value of each element in; represents the median function; represents the edge threshold parameter; represents the edge sharpness parameter; represents the diffusion coefficient function; S33. Anisotropic Diffusion: Solve the diffusion equation guided by the enhanced gradient amplitude to filter the fused image. This process is iterative within the virtual diffusion time and terminates when the maximum difference between two adjacent iterations is less than the set threshold. The final output is the diffusion-filtered image.

5. The artificial intelligence-based anti-NMDAR encephalitis clinical prognosis assessment method according to claim 4, characterized in that: Step S33 specifically includes: Output image after diffusion filtering By numerically solving partial differential equations Get, when initialized, Output image after diffusion filtering of iterations Set to the fused output tensor , and then perform iterative solution to define the discretized virtual diffusion time , and define the step size as , updated using the explicit Euler method, the update method is expressed as: ， in, Indicates the The output image after the diffusion filtering of the iterations; Indicates the The output image after the diffusion filtering of the iterations; Indicates step size; definition The mathematical model of the diffusion process .

6. The artificial intelligence-based anti-NMDAR encephalitis clinical prognosis assessment method according to claim 5, characterized in that: In step S4, the dynamic convolution module gives the convolution kernel lesion prior knowledge by initializing the dynamic convolution kernel parameters, uses anisotropic depth-separable convolution to decompose the feature extraction process along the three orthogonal directions of x-axis, y-axis, and z-axis, and uses the directional attention map to enhance the main response of the lesion: Dynamic convolution kernel parameter initialization: Calculate the average value of the diffusion filtered images of healthy samples and diseased samples respectively to obtain the healthy template and lesion template. Subtract the lesion template from the healthy template to obtain the basic convolution kernel parameters. Calculate the Sigmoid weighted expected value of the diffusion image difference between random healthy samples and random diseased samples and multiply it by the scaling factor to generate the lesion difference enhancement term. Add the basic convolution kernel parameters and the lesion difference enhancement term to obtain the final convolution kernel initial parameter tensor. Anisotropic depthwise separable convolution: Calculate the output image after diffusion filtering along the three-dimensional spatial coordinate axis The partial derivative of the gradient field tensor in the corresponding direction is obtained to represent the rate of change of the image in that direction. For the The gradient field tensor in the direction, Represents the directional variables of the x-axis, y-axis, and z-axis in three-dimensional space; uses the temperature-regulated Softmax function and adopts the temperature parameter to control the sharpness of the weight distribution, converting the gradient field tensors in the three directions into the corresponding directional attention weight map; the formula is as follows: ， in, express Attention weight map of direction; represents the temperature parameter; Indicates the difference from The direction variables of the x-axis, y-axis, and z-axis in the three-dimensional space, Representation Sum Index Traverse the three spatial directions of x-axis, y-axis and z-axis; Indicates along The gradient field tensor of the direction; Represents the natural exponential function; the output image after diffusion filtering along three directions The weighted input image performs a one-dimensional convolution operation, using the corresponding trainable convolution kernel parameters in each direction to generate the feature tensor of the convolution output in that direction; the feature tensors output by the convolution along the x-axis, y-axis, and z-axis are spliced ​​along the feature dimension to generate a multi-directional convolution feature tensor.

7. The artificial intelligence-based anti-NMDAR encephalitis clinical prognosis assessment method according to claim 6, characterized in that: In step S4, the spatial feature decoupling module decouples the lesion morphological features through 3D convolution and curvature-driven spatial attention mechanism, specifically: performing a three-dimensional convolution operation on the multi-directional convolution feature tensor, using the Gaussian error linear unit activation function combined with nonlinearity, and at the same time, using the Laplace operator to calculate the curvature of the input feature, generating a spatial attention map through convolution adjustment and Sigmoid function, and multiplying the multi-directional convolution feature tensor and the spatial attention map element by element to obtain the decoupled spatial feature tensor.

8. The artificial intelligence-based anti-NMDAR encephalitis clinical prognosis assessment method according to claim 7, characterized in that: In step S4, the gated pooling module performs gated weighted down-sampling on the feature map through the lesion difference-aware pooling operation to retain the lesion sensitive area information, specifically: using the lesion template generated in the dynamic convolution kernel parameter initialization stage to calculate the similarity between the local area of ​​the decoupled spatial feature tensor and the lesion template; using the sharpness adjustment parameter to control the sharpness of the weight distribution, combining the similarity with the feature curvature of the decoupled spatial feature tensor to generate an adaptive gating weight; using the gating weight to weight the local area features of the decoupled spatial feature tensor, and then performing the maximum pooling operation to finally obtain the pooled feature tensor.

9. The artificial intelligence-based anti-NMDAR encephalitis clinical prognosis assessment method according to claim 8, characterized in that: In step S4, the fully connected classifier includes a fully connected layer and an output layer to achieve feature dimensionality reduction and high-level abstract feature extraction. Specifically, the pooled feature tensor is flattened into a one-dimensional feature vector, which is passed through multiple fully connected layers for data transmission, and finally outputs a compact low-dimensional feature vector; each fully connected layer consists of trainable weights and bias parameters, and uses a ReLU nonlinear activation function for feature transformation; The compact low-dimensional feature vector is passed through the output layer to obtain a prognostic assessment classification category; The output layer is also a fully connected layer and uses the Softmax function.

10. The artificial intelligence-based anti-NMDAR encephalitis clinical prognosis assessment method according to claim 9, characterized in that: In step S4, the loss function uses the cross entropy loss function to measure the difference between the prognosis category probability distribution predicted by the model and the true label. The cross entropy loss function quantifies the degree of classification error and serves as the optimization target.

Images