A method and system for image enhancement of traumatic neuropathic pain

Abstract

This invention discloses a method and system for image enhancement processing of traumatic neuropathic pain, relating to the fields of data processing and artificial intelligence. Based on multi-scale Gaussian difference edge extraction and structural tensor enhancement, this invention achieves precise preservation of blurred lesion boundaries; independently extracts structural features and optimal scale edge features to solve the problem of unbalanced feature responses caused by different lesion sizes; based on dual-path orthogonal constraints combined with spatial weight modulation, it achieves complete decoupling of lesion features from normal tissue; and utilizes edge features to generate a spatial attention mask to guide the reconstruction process while maintaining the continuity of nerve bundles and avoiding noise amplification. This achieves precise preservation of lesion details, effective suppression of noise, and structural reconstruction of the continuity of nerve bundles, providing clearer and more reliable imaging evidence for clinical diagnosis.

Description

Technical Field

[0001] This invention relates to the fields of data processing and artificial intelligence, and in particular to a method and system for enhancing images of traumatic neuropathic pain. Background Technology

[0002] Traumatic neuropathic pain is a type of chronic, intractable pain caused by peripheral nerve injury. Its clinical diagnosis largely depends on the visualization of pathological features such as nerve bundle rupture and microcalcifications by magnetic resonance imaging (MRI). However, due to limitations in imaging conditions, noise interference, and inter-tissue gray-level similarity, MRI images of traumatic neuropathic pain often present with problems such as blurred lesion boundaries, low local contrast, and unclear texture features.

[0003] Existing autoencoder-based deep learning methods often employ single-path encoding during feature extraction, making it difficult to simultaneously consider structural information and edge details. Furthermore, the lack of explicit modeling of the mutual exclusion between lesion and normal tissue features during feature fusion allows pathological features and noise signals to interfere with each other. In addition, existing decoding and reconstruction processes generally lack prior constraints on edges and anatomical structures, easily resulting in edge breaks or blurring at delicate neural bundle structures. While global enhancement methods can improve overall brightness and contrast, they also amplify noise in normal tissue areas, masking the true features of lesions. Summary of the Invention

[0004] To address these issues, this invention proposes an image enhancement method for traumatic neuropathic pain based on an improved autoencoder model. From multi-scale structural preprocessing, dual-channel edge-aware coding, lesion feature decoupling enhancement to edge-guided decoding and dynamic mask fine-tuning, it achieves precise preservation of lesion details, effective suppression of noise, and structural reconstruction of nerve bundle continuity, thereby providing clearer and more reliable image evidence for clinical diagnosis.

[0005] This invention is achieved through the following technical solution:

[0006] A method for image enhancement processing of traumatic neuropathic pain, comprising:

[0007] Step S1, training dataset construction; the training dataset consists of magnetic resonance imaging images of traumatic neuropathic pain. The labeled data required for supervised training is constructed by manually delineating pathological features independently in advance, and generating a binary lesion mask after cross-validation.

[0008] Step S2, multi-scale structure preprocessing; the preprocessed image is obtained based on adaptive weights combined with Gaussian difference operator and gradient tensor;

[0009] Step S3, Improved autoencoder model construction and training;

[0010] Step S4, enhanced output of traumatic neuropathic pain images; input the traumatic neuropathic pain MRI images into the trained improved autoencoder model, and output the enhanced images.

[0011] Furthermore, the preprocessed image representation obtained based on adaptive weights combined with the Gaussian difference operator and gradient tensor is as follows:

[0012]

[0013] In the formula, Preprocessed MRI images; The original MRI image. For the scale The Gaussian difference operator, It is the Frobenius norm. The gradient tensor of the original MRI image, for transpose, For adaptive weights, This is the structural reinforcement factor.

[0014] Furthermore, the adaptive weight calculation method is expressed as follows: ;in, To distinguish it from the scale index of s; The mean parameter in the weight calculation controls the center position of the target scale; The standard deviation parameter is used in weight calculations to control the width of the scale distribution.

[0015] Furthermore, the improved autoencoder model construction includes lesion region feature decoupling enhancement, enforced feature channel independence based on dual-path orthogonal constraint loss, and feature updating and decoupling combined with spatial weight modulation. The dual-path orthogonal constraint loss is expressed as...

[0016] In the formula, The loss is due to orthogonal constraints. Orthogonal weights within structural features; Align weights for structure-guided features; The first structural feature map One channel slice; The first structural feature map One channel slice; The first structural feature map One channel slice; To guide the feature map's first One channel slice; This refers to matrix inner product operations. This is the channel index, and its value range is... to ; To distinguish from Channel index; The number of channels in the structural feature map; This is the channel index of the structural feature map; For the first Channel normalization term for each channel.

[0017] Furthermore, the improved autoencoder model construction also includes an adaptive edge-guided decoder construction, which utilizes edge feature maps. Generate spatial attention mask And upsampling features are modulated by element-wise multiplication.

[0018] Furthermore, the improved autoencoder model construction also includes calculating the lesion perception contrast loss based on the location feature vector and the lesion sampling point set through cosine similarity calculation and exponential function, and calculating the multi-scale gradient consistency loss based on the gradient magnitude map through L1 norm.

[0019] Furthermore, the expression for the lesion perception contrast loss function is as follows:

[0020]

[0021] In the formula, Loss of contrast perception of lesions; This is the set of all pixel locations in the entire image region. Indicates the number of pixels in the lesion area; Represents the set of pixel locations in the lesion region; This is a cosine similarity calculation function; Temperature coefficient; To distinguish it from the pixel position index of p, it represents a specific coordinate point; To reconstruct the image The feature vector of the location; For real images in The feature vector of the location.

[0022] Furthermore, the improved autoencoder model construction includes dynamic mask enhancement fine-tuning, local enhancement of the reconstructed image based on a binary lesion mask, mixing the original image with the unsharpened masked enhanced version in the lesion region, and retaining the original reconstruction result in the normal region to obtain the lesion mask enhanced image.

[0023] The present invention also provides an image enhancement processing system for traumatic neuropathic pain, based on the image enhancement processing method for traumatic neuropathic pain described above, comprising:

[0024] The training dataset building module is used to acquire MRI images and build the labeled data required for supervised training.

[0025] The multi-scale structure preprocessing module is used to extract multi-scale edge features by applying the difference of Gaussian operator and weighting the contributions of different scales through adaptive weights. At the same time, it calculates the gradient tensor of the original image, and obtains the structural tensor components by taking its transpose and calculating the Frobenius norm. Then, it achieves the preservation of edge and texture details by weighted addition, avoiding damage to fine structures, and obtains the preprocessed MRI.

[0026] The improved autoencoder model construction and training module includes a dual-channel edge-aware encoder module, a lesion region feature decoupling enhancement module, an adaptive edge-guided decoder module, a lesion-aware contrast loss calculation module, a multi-scale gradient consistency loss calculation module, a dynamic mask enhancement fine-tuning module, and an improved autoencoder model iterative training and parameter update module.

[0027] The traumatic neuropathic pain image enhancement module is used to process new traumatic neuropathic pain MRI images into a trained improved autoencoder model to obtain the final image.

[0028] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing program instructions for a method of image enhancement processing for traumatic neuropathic pain. The program instructions for the method of image enhancement processing for traumatic neuropathic pain can be executed by one or more processors to implement the steps of the method of image enhancement processing for traumatic neuropathic pain as described above.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] 1. By combining multi-scale Gaussian difference edge extraction and structural tensor enhancement, and using adaptive weights to focus on the most relevant scale, we can achieve fine preservation of blurred lesion boundaries while suppressing noise.

[0031] 2. Independently extract structural features and optimal scale edge features, and fuse them in the encoding stage to solve the problem of uneven feature response caused by different lesion sizes, and improve the separability of lesions and normal tissues.

[0032] 3. By employing dual-path orthogonal constraints combined with spatial weight modulation, the lesion characteristics are completely decoupled from normal tissues, thereby enhancing the salience of the lesions.

[0033] 4. During the decoding stage, edge features are used to generate a spatial attention mask to guide the reconstruction process and maintain the continuity of the neural bundle. Dynamic mask non-sharpening enhancement is used to improve contrast only in the lesion area and avoid noise amplification. Attached Figure Description

[0034] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0035] Figure 1 This is a schematic flowchart of an image enhancement processing method for traumatic neuropathic pain according to an embodiment of this application;

[0036] Figure 2 These are the original MRI images according to embodiments of this application;

[0037] Figure 3 is an edge response image after multi-scale structural preprocessing according to an embodiment of the present application, wherein Figures 3(a)-(e) are edge response images with scales of 1-5, respectively.

[0038] Figure 4 The feature map is constructed by extracting multi-scale edge components according to the embodiments of this application;

[0039] Figure 5 It constructs a feature map based on the structural tensor components according to the embodiments of this application;

[0040] Figure 6 These are preprocessed MRI images according to embodiments of this application;

[0041] Figure 7 These are edge detection performance results under different total scale numbers according to the embodiments of this application;

[0042] Figure 8 This is a heatmap of the spatial attention mask according to an embodiment of this application;

[0043] Figure 9 The image is a reconstructed image according to an embodiment of this application;

[0044] Figure 10 This is a comparison diagram of the distribution of lesion perception contrast loss according to an embodiment of this application;

[0045] Figure 11 It is a lesion mask enhancement image according to an embodiment of this application;

[0046] Figure 12 This is a comparison of the Dice coefficients of different methods in the lesion area. Detailed Implementation

[0047] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0048] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0049] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. The illustrations only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the shape, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0050] See Figure 1 A method for image enhancement processing of traumatic neuropathic pain includes the following steps:

[0051] S1. Construction of training dataset

[0052] The training dataset consists of magnetic resonance imaging (MRI) images of traumatic neuropathic pain. The raw data was acquired using an MRI scanner, strictly adhering to medical ethical standards. With informed consent from the patients, case data exhibiting clear clinical symptoms of traumatic neuropathic pain, such as nerve tract rupture and microcalcifications, were selected. Figure 2 The image shown is the original MRI image.

[0053] Furthermore, the labeled data required for supervised training is constructed by manually delineating the lesion area in advance, including pathological features such as nerve bundle rupture areas and microcalcifications. After cross-validation, a binary lesion mask is generated, with the size consistent with the original MRI image. In the mask, the lesion area is marked as 1, and normal tissue is marked as 0.

[0054] The constructed training dataset contains 1200 paired samples, which are divided into training set, validation set and test set in an 8:1:1 ratio.

[0055] S2, Multi-scale structural pretreatment

[0056] MRI images often show blurred boundaries of lesion areas, and the texture of normal tissue is easily affected by noise. Conventional normalization methods such as Min-Max scaling can destroy the fine structural features of the image, causing the loss of key pathological information in subsequent enhancement processes and failing to effectively preserve lesion details.

[0057] Based on the original MRI image, multi-scale edge features are extracted using the difference of Gaussian operator, and the contributions of different scales are weighted using adaptive weights. Simultaneously, the gradient tensor of the original image is calculated, and its transpose is combined with the Frobenius norm to obtain the structural tensor components. Then, weighted summation is used to preserve edge and texture details, avoiding damage to fine structures, resulting in a preprocessed MRI, represented as:

[0058]

[0059] In the formula, The preprocessed MRI image has a size of [size missing]. ; The original MRI image, with a size of ; The height of the original MRI image; The width of the original MRI image;

[0060] For the scale The difference of Gaussian operator is used to extract edge features at a specific scale, which is specifically implemented by processing images with Gaussian filters of different standard deviations. This is a scale index, with values ​​ranging from 1 to S; The total number of scales is a preset parameter that controls the coverage of multi-scale analysis, such as... It is usually set to 3. or To cover multi-scale edges ranging from minute lesions to coarse large structures, balancing computational efficiency and feature integrity; The structural enhancement factor controls the intensity of texture retention, such as... =0.7; The Frobenius norm is used to calculate the norm of a matrix to quantify structural strength. This is the Sobel gradient operator, used to calculate the spatial gradient of an image; The gradient tensor of the original MRI image is specifically calculated using the Sobel operator; for transpose; These are adaptive weights used to weight contributions at different scales, and their calculation method is expressed as follows: ;in, To distinguish it from the scale index of s; The mean parameter in the weight calculation controls the center position of the target scale, for example, through an empirical formula. Calculated; The standard deviation parameter in the weight calculation controls the width of the scale distribution; for example, it can be set to a fixed value. .

[0061] It should be noted that when processing images using Gaussian filters with different standard deviations to implement the difference in Gaussian operator, the original image is first convolved with two Gaussian kernels of different standard deviations to obtain two images with different degrees of blur. Then, these two blurred images are subtracted pixel by pixel to obtain the edge response map at that scale. The Gaussian kernel with a smaller standard deviation retains more details, while the Gaussian kernel with a larger standard deviation has a stronger smoothing effect. The difference between the two can highlight the edge features within a specific scale range. Throughout the process, zero padding is used to keep the output image size unchanged.

[0062] It should also be noted that the gradient tensor is obtained by calculating the Sobel operator. At that time, Sobel convolution kernels in the horizontal and vertical directions are used to perform two-dimensional convolution operations on the original image. The horizontal convolution kernel detects the vertical edge and generates the horizontal gradient component, and the vertical convolution kernel detects the horizontal edge and generates the vertical gradient component. Then, these two gradient components are combined into a tensor with two channels, where each pixel position contains the gradient values ​​in the horizontal and vertical directions, forming a complete image gradient representation.

[0063] It should also be noted that in MRI scans for traumatic neuropathic pain, lesion boundaries are often blurred and subject to significant noise interference; conventional normalization can disrupt fine structures. The term represents the multi-scale edge extraction components, through adaptive weights. Weighted, emphasizing the edge responses at the most relevant scales, and the weights Based on Gaussian distribution, Gaussian difference is usually only used for edge detection. Here, by combining it with structural tensor, it can achieve the technical effect of improving the local contrast of the lesion area, so that subsequent modules can more accurately separate pathological features, such as capturing microcalcifications at a small scale and preserving the continuity of nerve bundles at a large scale. Multi-scale weighting not only suppresses noise, but also enhances the topological coherence of blurred boundaries.

[0064] It should also be noted that, The term characterizes the structural tensor components. Considering that normal tissue texture in neuropathic MRI is easily contaminated by noise, the local structural strength is quantified by utilizing the Frobenius norm of the structural tensor and the structural enhancement coefficient. By controlling the intensity of texture enhancement, the texture inside the lesion is significantly enhanced while preserving the edges. In particular, the use of the Frobenius norm, which is commonly used in gradient tensors, can amplify the anisotropic features of the pathological region and improve the signal-to-noise ratio of texture details.

[0065] Figure 3 shows the edge response images of multi-scale structures after preprocessing at different scales. Figures 3(a)-(e) are the edge response images at scales of 1-5, respectively. Figure 4 Construct feature maps for multi-scale edge extraction components; Figure 5 To construct feature maps for the structural tensor components; Figure 6 A preprocessed MRI image according to an embodiment of this application.

[0066] The scientific validity of scale parameter selection in multi-scale preprocessing was verified by evaluating edge detection performance under different total scale settings (S). The impact of scale parameters 1 to 5 on edge detection performance was tested, assessing precision (the proportion of correctly detected edge pixels), recall (the proportion of real edges detected), and F1 score (a comprehensive balance indicator). The line graph shows that when the scale parameter is 3, all three indicators reach their peak simultaneously, with the F1 score being the highest. Scale 1 is sensitive to noise, resulting in the lowest precision. Scale 3 achieves the best balance between preserving neural bundle continuity and enhancing microcalcification point edges, thus improving the topological coherence retention rate of blurred boundaries. Figure 7 The results show the edge detection performance under different total scale numbers.

[0067] S3. Improved Autoencoder Model Construction and Training

[0068] S301, Dual-channel edge-aware encoder construction

[0069] In MRI analysis of neuropathic pain, lesion structural features such as microcalcifications and nerve bundles have similar grayscale characteristics. Conventional encoders using a single convolution path cannot effectively separate these features from noise, which can easily lead to feature confusion and reduce the identification of pathological areas.

[0070] This invention employs a dual-channel encoding path to independently extract structural features and edge features based on the optimal scale. Channel splicing and fusion are then used to improve the separation of lesion structural features from noise. The specific steps are as follows:

[0071] 1) Structural channel feature extraction

[0072] Based on preprocessed MRI images, a dimensionality-reduced structural feature map is output through convolution operations with a specified number of layers, achieving spatial size reduction and structural feature encoding. In one embodiment, the obtained structural feature map is defined as follows: This process is implemented using a 3×3 convolution kernel, represented as... ;

[0073] in, For a convolution operation module containing d layers of 3×3 convolution kernels, in In this item, the preset step size for each layer is 2, and the fill size is 1; The number of convolutional layers controls the spatial downsampling rate, e.g., d=4; This is a structural feature map with dimensions of [size missing]. The convolution operation reduces the spatial size.

[0074] 2) Calculation of optimal scale Gaussian difference

[0075] Based on the preprocessed MRI images, the L2 norm of the Gaussian difference is calculated for each scale in a preset scale set. The edge response intensity of each scale is then processed and evaluated using the Gaussian difference operator. The scale with the largest norm is selected as the optimal scale, and the index of the optimal scale that produces the maximum response is output, denoted as:

[0076]

[0077] In the formula, The optimal scale index is the scale that produces the maximum edge response;

[0078] It is an L2 norm;

[0079] In the first The index operator for finding the maximum value on a scale.

[0080] 3) Edge channel feature extraction

[0081] Based on the preprocessed MRI image and the optimal scale index, edge features are extracted using the optimal scale Gaussian difference operator, and spatial downsampling is performed through max pooling to output the edge feature map, represented as:

[0082]

[0083] In the formula, This is the optimal scale Gaussian difference operator;

[0084] For max pooling operations, in In this project, the spatial dimensions are downsampled from H×W to H / 16×W / 16 by pooling with a window size of 2×2 and a step size of 2 four times.

[0085] This is an edge feature map with a size of [size missing]. Pooling operations reduce the size of the space.

[0086] 4) Dual-channel feature fusion

[0087] The structural feature map output from the structural channel and the edge feature map output from the edge channel are concatenated along the channel dimension, represented as follows: Based on this, a fused coding feature map is generated, and the resulting fused coding feature map is defined as follows: The size is ;

[0088] in, This is a channel splicing operation that connects two feature maps along the channel dimension.

[0089] It should be noted that the lesion structure of microcalcifications has highly similar grayscale characteristics to the normal tissue of nerve bundles, and noise interference can further obscure the feature boundaries, making it difficult for conventional single-path encoders to effectively separate pathological information and structural feature maps. and edge feature map Acquired through dual independent channels, and via This approach achieves adaptive optimal scale selection, driving edge feature extraction and resolving the issue of uneven response in multi-scale lesions. Furthermore, it provides an optimal scale index. Instead of simply selecting the strongest edge, it assesses the overall response intensity through the L2 norm and adaptively focuses on the scale most relevant to the lesion. For example, microcalcifications correspond to small scales and nerve bundle breaks correspond to large scales. Under noise interference, it can enhance the topological coherence of blurred boundaries. Therefore, the optimal scale-driven dual-channel mechanism can achieve "unentanglement" of the feature space under the gray-scale similarity constraint of MRI. That is, edge features serve as spatial guiding cues, indirectly amplifying the local anisotropy of the lesion, rather than just enhancing the edge intensity.

[0090] S302, Decoupling and Enhancement of Lesion Area Features

[0091] Conventional bottleneck layers ignore the mutual exclusion between lesion areas and normal tissue features, and do not make full use of multi-source feature information, resulting in insufficient specificity of pathological features and inability to effectively distinguish pathological patterns.

[0092] This invention utilizes fused coded feature maps as guidance, forces feature channel independence through dual-path orthogonal constraint loss, and combines spatial weight modulation for feature updating and decoupling, eliminating the coupling between pathological features and normal tissues, and enhancing the specific characterization and local contrast of lesion areas. The specific steps are as follows:

[0093] 1) Feature-guided projection

[0094] The fused encoded features are projected onto the structural feature space to generate a guiding feature map, specifically achieved through a 1×1 convolution operation, as shown below. The resulting guided feature map is defined as follows: The size is ;

[0095] in, For a 1×1 convolution operation, in In this item, the number of output channels is .

[0096] 2) Dual-path orthogonal constraint loss

[0097] Channel slices based on structural feature maps and guiding feature maps are evaluated using matrix inner product operations to assess internal orthogonality and alignment losses, and orthogonality constraint losses are output. By calculating the bipath orthogonality loss of structural and guiding features, the independence of representations between channels is enforced, as expressed below:

[0098]

[0099] In the formula, The loss is due to orthogonal constraints. The first structural feature map Each channel slice, with a size of [size missing] ; The first structural feature map Each channel slice, with a size of [size missing] ; The first structural feature map Each channel slice has a size of H / 16×W / 16; To guide the feature map's first Each channel slice, with a size of [size missing] ; This refers to matrix inner product operations. This is the channel index, and its value range is... to ; To distinguish from Channel index; The number of channels in the structural feature map; This is the channel index of the structural feature map; For structural features, orthogonal weights are used, such as... ; Align weights for structure-guided features, such as, ; For the first The channel normalization term for each channel is calculated as follows: .

[0100] It should be noted that, regarding the issue of mutual exclusion of features between lesions and normal tissues in MRI for neuropathic pain, orthogonal constraint loss... Implement dual-path constraints, The term, as an internal orthogonal term, forces independence between structural feature channels to reduce redundancy. The term is used as an external alignment term, combined with the guiding feature map. Achieve channel-level semantic alignment, while incorporating channel normalization terms. Adaptively balancing the contribution weights of different channels, in MRI with blurred lesion boundaries, the external alignment term is adjusted through channel normalization. Dynamically adjusting feature importance can effectively suppress noise-dominated channels while strengthening pathologically sensitive channels, such as those responding to microcalcifications. This decoupling process not only eliminates feature coupling but also improves the signal-to-noise ratio of the lesion region, a result that conventional methods cannot achieve. This is because conventional orthogonal constraints are only used for feature dimensionality reduction. However, this invention, combined with guided features, achieves "semantic-guided decoupling" in medical images, i.e., guided feature maps. Derived from the fusion coding features, it carries multi-scale edge information, enabling orthogonal loss to distinguish between pathological abnormalities and normal variations in weakly textured regions of MRI.

[0101] It should also be noted that setting the preferred value and This is to prioritize ensuring the independence of internal channels and avoid contamination of pathological features by normal tissue, and then use external alignment to ensure that the decoupling features are consistent with the edge guidance. In the processing of MRI images of neuropathic pain, it can enhance the local contrast of lesions. After the internal orthogonal term eliminates channel correlation, the external alignment term injects the spatial prior of edge features into the structural features through matrix inner product, so that subsequent feature updates can more accurately locate the nerve bundle break area, thereby producing a synergistic effect of "redundancy removal-enhancement". That is, the feature channels form sparse representation under orthogonal constraints, while the edge guidance focuses on key pathological areas, improving the targeting of decoupling.

[0102] 3) Feature decoupling and enhanced update

[0103] Based on the structural feature map, orthogonal constraint loss and its gradient, features are updated along the orthogonal loss gradient direction through gradient update and spatial weight modulation of Sigmoid activation. Guided features are fused to enhance lesion representation, and a decoupled feature map is output, represented as:

[0104]

[0105] In the formula, For the decoupled feature map, the size is ; The decoupling strength coefficient controls the magnitude of feature separation, for example, =0.05; The enhancement coefficient of the lesion, such as, ; Use the Sigmoid activation function; This is element-wise multiplication.

[0106] It should be noted that orthogonal constraint loss Dependency Structure Feature Map Channel slices are structural feature maps. The function, therefore, utilizes orthogonal constraint loss. Structural feature map Calculate the partial derivative for gradient update.

[0107] It should be noted that, The term represents the orthogonal decoupling term. Through gradient backpropagation of the dual-path orthogonal constraint loss, it forces structural feature channels to be independent of each other and aligned with edge-guided features, effectively eliminating the coupling between pathological features and normal tissues, improving the separability of the feature space, and... The term represents the lesion enhancement term. It utilizes the Sigmoid spatial weights generated by the fusion features to specifically amplify the response intensity of the lesion-sensitive area on the orthogonally decoupled features. The orthogonal decoupling term and the lesion enhancement term form a "decoupling-enhancement" cascade mechanism. Orthogonal decoupling first eliminates feature redundancy and separates pathological patterns. The lesion enhancement then strengthens the local contrast of the nerve bundle break area and microcalcification points through edge-guided spatial modulation. While reducing channel correlation, it improves the signal-to-noise ratio of the lesion boundary, achieving synergistic optimization of decoupling thoroughness and lesion saliency.

[0108] S303, Adaptive Edge-Guided Decoder Construction

[0109] Conventional deconvolution operations can lead to blurred edges during image reconstruction, especially in the reconstruction of fine nerve bundle structures in traumatic neuropathic pain MRI, which is not effective in preserving key anatomical features of lesion boundaries and affects diagnostic accuracy.

[0110] This invention utilizes edge feature maps extracted by the encoder to generate a spatial attention mask, and guides feature modulation during the decoding upsampling process to enhance the reconstruction accuracy and clarity of key anatomical structures such as lesion boundaries. The specific steps are as follows:

[0111] 1) Edge feature space transformation

[0112] Perform a 1×1 convolution operation on the edge feature map output by the encoder to compress its channel dimension to 1, resulting in the spatial attention base map, represented as follows. The resulting spatial attention baseline graph is defined as follows: The size is ;

[0113] in, For a 1×1 convolution operation, in In this term, one convolution kernel is used by default, and there is no bias term.

[0114] 2) Spatial attention mask generation

[0115] Applying the Sigmoid activation function to the spatial attention base map generates a spatial attention mask. The output mask, with values ​​in the range [0,1], is represented as follows: The obtained spatial attention mask is defined as follows: The size is The value range is [0,1]; where, This is the Sigmoid activation function.

[0116] 3) Feature upsampling reconstruction

[0117] Perform a deconvolution operation with a stride of 2 on the decoupled feature map to gradually restore the spatial resolution, and output an upsampled feature map, represented as follows: The obtained upsampled feature map is defined as follows: The size is ;

[0118] in, For deconvolution modules, in The project includes 4 layers of deconvolution operations, each using a 3×3 convolution kernel with a stride of 2 and padding of 1.

[0119] 4) Edge-guided feature modulation

[0120] Based on the upsampled feature map and the upsampled mask, the reconstructed image is output through element-wise multiplication, as follows:

[0121]

[0122] In the formula, For bilinear upsampling operation, in In this project, the mask spatial resolution is presampled from H / 16×W / 16 to H×W;

[0123] This is element-wise multiplication;

[0124] To reconstruct the image, the size is .

[0125] It should be noted that, in order to overcome the defect of blurred edges caused by conventional deconvolution in MRI reconstruction of neuropathic pain, this invention utilizes edge feature maps. Generate spatial attention mask Furthermore, by modulating upsampling features through element-wise multiplication, the problem of anatomical distortion during reconstruction is resolved. This is achieved even under noise interference in traumatic neuropathic MRI. Instead of simply binarizing the edge map, it generates continuous weights through Sigmoid activation, which can achieve smooth enhancement of the edge transition zone. That is, the high-weighted region of the lesion boundary enhances the gradient response, while the low-weighted region of homogeneous tissue suppresses reconstruction artifacts, so that the continuity of the nerve bundle is maintained during the decoding process. In MRI reconstruction tasks, it is equivalent to "anatomical prior". The multi-scale information of the edge features is guided by the mask for upsampling, and key diagnostic areas, such as nerve bundle connection points, are restored first, avoiding the homogenization of the global reconstruction.

[0126] It should also be noted that the spatial attention base map By compressing the channel dimension through 1×1 convolution and then applying Sigmoid activation, a spatial attention mask is obtained. This allows the mask to incorporate the statistical properties of multi-channel edge responses. This operation is designed to improve the reconstruction accuracy of lesion boundaries, due to the edge feature map... Derived from multi-scale Gaussian difference, its channels contain edge information of different scales, and the resulting spatial attention mask is compressed. It can adaptively fuse thick and thin edges, enhance local contrast in the blurred neural bundle breakage area, and suppress over-enhancement in the sharp calcification point, achieving a balance between "structure fidelity" and "noise suppression", making the model more robust in low signal-to-noise ratio regions.

[0127] In one embodiment, the thermal map of the spatial attention mask is as follows: Figure 8 As shown, the reconstructed image is as follows Figure 9 As shown.

[0128] S304, Calculate the contrast loss of lesion perception.

[0129] Conventional mean squared error loss functions ignore the differences in local pathological features in image reconstruction tasks, and cannot enhance the feature differentiation between lesion areas and non-lesion areas, which can easily lead to insufficient contrast in key pathological areas of the reconstructed image.

[0130] This invention constructs a set of sampling points in the lesion region and extracts local feature vectors. It employs a lesion-perception contrast loss mechanism to maximize the consistency between the lesion location features and the actual annotations while minimizing the similarity to non-lesion regions, thereby improving the model's ability to identify pathological features. The specific steps are as follows:

[0131] 1) Construct a set of sampling points for the lesion area

[0132] Based on the binary lesion mask, the pixel locations with a mask value of 1 are selected to output the set of lesion sampling points. The coordinates of all lesion pixel locations are extracted and represented as follows:

[0133]

[0134] In the formula, It is a binary lesion mask with a size of The lesion area is pre-labeled by manual binarization.

[0135] The pixel coordinates are the position. , ;

[0136] This is a set of sampling points for the lesion area, containing the pixel locations of all lesions.

[0137] 2) Extracting location feature vectors

[0138] Based on the reconstructed image, the real image, and pixel locations, a location feature vector is output through local windowing operations, represented as follows:

[0139]

[0140]

[0141] In the formula, This is a real image, with dimensions of [size missing]. Specifically, this is achieved through manual annotation, generating noise-free, high-contrast reference images by manually annotating the original traumatic neuropathic MRI.

[0142] To reconstruct the image The feature vector of the location, with size ;

[0143] For real images in The feature vector of the location, with size ;

[0144] For image extraction operations, Take the location as the center × Neighborhood;

[0145] The neighborhood size, i.e., the size of the local window, is used to define the size of the local window. The side length of the square area centered on the location, for example, .

[0146] 3) Construct the contrastive loss function

[0147] Based on the location feature vector and the set of lesion sampling points, the perceptual contrast loss of the lesion is calculated using cosine similarity calculation and exponential function. This maximizes the similarity between the lesion location and the features at the same location in the real image, while minimizing the similarity with other locations, as expressed below:

[0148]

[0149] In the formula, This is the set of all pixel locations in the entire image region. Indicates the number of pixels in the lesion area;

[0150] This is a cosine similarity calculation function; The temperature coefficient controls the sharpness of the similarity distribution, for example, =0.1; It is an exponential function; To distinguish it from the pixel position index of p, it represents a specific coordinate point; Loss of perception contrast for lesions.

[0151] It should be noted that this is achieved by constructing a set of sampling points in the lesion area. It also extracts local feature vectors, and then uses cosine similarity and temperature coefficient to... By implementing a contrastive learning mechanism to maximize the consistency between lesion location features and real annotations, the system can capture contextual information of small lesions such as microcalcifications in noisy MRI. By comparing the similarity between the lesion location and the overall image location, the loss function can not only enhance the contrast of the lesion area, but also suppress the noise response in normal tissue, thereby strengthening the spatial consistency of the pathological structure. That is, small microcalcified lesions form compact clusters in the feature space, while normal tissue is pushed away.

[0152] It should also be noted that a temperature coefficient is preferably set. This is due to the sharpness of the similarity distribution, temperature coefficient Values ​​less than 1 make the similarity distribution sharper, which can improve the discriminative power of lesion features. In the blurred boundary areas of neuropathic pain MRI, low similarity values ​​are more effective. Magnifying subtle differences forces the model to focus on the consistency within the lesion, such as the texture pattern in the nerve bundle damage area.

[0153] like Figure 10As shown, a comparative analysis of the lesion perception contrast loss distribution was conducted. Kernel density estimation was used to compare the distribution characteristics of the two techniques in lesion perception contrast loss. The horizontal axis represents the loss value (dimensionless), and the vertical axis represents the probability density (dimensionless). The red distribution represents the conventional technique, an enhancement technique based on a conventional autoencoder, employing a standard convolutional autoencoder architecture containing a single-channel encoding path and a deconvolutional decoding path. The blue distribution represents the technique of this invention. Analysis shows that the loss distribution of the technique of this invention is generally shifted to the left, and its mean position (blue dashed line) is significantly lower than that of the conventional technique (red dashed line), indicating that the technique of this invention has better overall feature consistency in the lesion region. Furthermore, the distribution curve of the technique of this invention is steeper and more concentrated, indicating that it performs more stably when processing different lesion samples. Experiments demonstrate that the lesion perception contrast loss mechanism, by constructing a set of lesion sampling points, maximizes the consistency between the lesion location and the true annotation in the feature space, while minimizing the similarity with non-lesion regions. This mechanism is particularly effective for enhancing small lesions such as microcalcifications in neuropathic MRI, enabling the model to more accurately enhance the local contrast of pathological areas.

[0154] S305. Calculate the multi-scale gradient consistency loss.

[0155] In MRI of traumatic neuropathic pain, the continuity of nerve bundle structures is crucial for diagnosis. However, conventional reconstruction methods lack effective constraints in gradient space, leading to breaks or discontinuities at the edges of nerve bundles, which affects the observation of subtle pathological features.

[0156] This invention calculates the difference in gradient magnitude between the reconstructed image and the real image at multiple scales and applies an L1 loss constraint to force the model to maintain the continuity of the neural tract edge structure at different resolutions, as expressed below:

[0157] 1) Constructing multi-scale downsampling sequences

[0158] Based on the input image, average pooling downsampling is used to output a downsampled image of a specified factor, realizing multiple downsampling operations on the reconstructed image and the real image, as shown below:

[0159]

[0160] In the formula, The input image has a size of [size missing]. Specifically refers to reconstructing images or real image ;

[0161] This is an average pooling downsampling operation with a downsampling factor of k;

[0162] The image is a downsampled image with a downsampling factor of k, specifically referring to the reconstructed image. or real image The downsampled image;

[0163] This is the downsampling factor. The value 1 represents the original scale, 2 represents half resolution, and 4 represents quarter resolution.

[0164] It should be noted that nerve bundles need to maintain coherence across resolutions in MRI. Single-scale constraints can easily overlook global structures, while multi-scale analysis can improve the ability to repair subtle breaks. Characterizing output size , Characterized by 2×2 pooling with a stride of 2, output size , Characterized by 4×4 pooling with a stride of 4, the output size is... Edge continuity is enforced through multi-scale gradient constraints.

[0165] 2) Calculate the gradient magnitude at each scale.

[0166] Based on the downsampled image, a gradient magnitude map is output through horizontal and vertical Sobel convolutions, represented as follows:

[0167]

[0168] In the formula, This is a horizontal Sobel operator with a 3×3 convolution kernel at the horizontal level. ;

[0169] It is a vertical Sobel operator with a 3×3 convolution kernel. ;

[0170] This is a two-dimensional convolution operation;

[0171] express Gradient magnitude map of scale, specifically referring to the reconstructed image. or real image The downsampled image.

[0172] 3) Calculate multi-scale L1 loss

[0173] Based on the gradient magnitude map, a multi-scale gradient consistency loss is output through L1 norm operation. This loss calculates the consistency difference between the reconstructed and real images at multiple scales, and is expressed as follows:

[0174]

[0175] In the formula, To reconstruct the image The gradient magnitude map at different scales will reconstruct the image. Substitution get;

[0176] For real images in A gradient magnitude map of the scale, representing the real image. Substitution get;

[0177] It is an L1 norm;

[0178] This is a multi-scale gradient consistency loss.

[0179] It should be noted that, in The calculation of gradient magnitude differences in downsampled sequences and the enforcement of multi-scale edge consistency using the L1 norm stem from the need for continuity of nerve bundle structures in MRI for neuropathic pain. Under the noise interference of traumatic neuropathic pain MRI, multi-scale analysis can enhance the global topological coherence of nerve bundles, coarse-scale constraints on the large-scale structural continuity of nerve bundle orientation, and fine-scale repair of micro-fractures within bundle injuries. The gradient magnitude calculated by the Sobel operator... By directly quantifying edge errors using L1 loss, the model maintains edge smoothness at different resolutions, making the reconstructed images more anatomically accurate in key areas of the neural bundle.

[0180] S306, Dynamic Mask Enhancement Fine-tuning

[0181] In contrast-enhanced MRI images of traumatic neuropathic pain, lesion areas such as nerve bundle damage or microcalcifications often have blurred boundaries and low contrast, while normal tissue is easily affected by noise. Conventional global enhancement methods, such as histogram equalization, will simultaneously increase the intensity of noise and normal areas, which can easily lead to the lesion features being masked or the boundaries being further blurred, making it impossible to achieve targeted contrast enhancement.

[0182] Local enhancement of the reconstructed image is performed based on a binary lesion mask. The original image and the unsharpened masked enhanced version are blended in the lesion region, while the original reconstruction result is preserved in the normal region, resulting in a lesion mask-enhanced image, represented as follows:

[0183]

[0184] In the formula, Enhance images by masking lesions;

[0185] This is an unsharpened masking enhancement function used to improve the local contrast of the input image. It is achieved by enhancing high-frequency details, and its calculation method is expressed as follows: ;

[0186] As a binary lesion mask, using edge feature maps Thresholding is used to generate a value of 1, which represents a lesion area and a value of 0, which represents a normal tissue area.

[0187] The weighting coefficient controls the balance between the original reconstructed image and the enhanced version in the lesion region, and its value ranges from [0, 1]. =0.6;

[0188] The gain coefficient for unsharpened masking controls the enhancement intensity of high-frequency details, such as... =1.2;

[0189] For Gaussian filtering operations, its subscript This is the standard deviation parameter of the Gaussian kernel, which controls the smoothing degree and is used to extract low-frequency components of the image, such as... Typically set to To balance noise suppression and detail preservation.

[0190] It should be noted that, This represents a normal tissue mask. 1 represents a matrix of all 1s with the same size as the mask. When 1 is subtracted from the binary mask, an element-wise operation is actually performed on two matrices of the same size. In the original mask, the lesion position is 1, which becomes 0 after subtraction. The normal tissue position is 0, which becomes 1 after subtraction. The generated new matrix forms a reverse mask, which is used to preserve the original reconstructed image in the normal area and avoid noise generated by enhancement.

[0191] It should also be noted that binary lesion masking This is generated by processing edge feature maps. First, multi-channel edge feature maps are generated. The values ​​of each channel are averaged to obtain a single-channel response map. Then, the response map is normalized to make the values ​​distributed within a standard range. Then, an appropriate threshold is set, and pixels in the normalized map that are above the threshold are marked as lesion areas, and pixels that are below the threshold are marked as normal tissue, forming a preliminary mask. Finally, the mask is enlarged to the original image size through interpolation.

[0192] It should be noted that, in conjunction with binary lesion masking Unsharpened masking enhancement function In the lesion area, the original reconstruction and enhanced version are blended, while the original image is preserved in the normal area. In low-contrast lesions, unsharpened masking is used to extract low-frequency components through Gaussian filtering, and then the gain coefficient is applied. Enhancing high-frequency details can improve local contrast at blurred boundaries, while also improving the binary lesion mask. This ensures that enhancement is applied only to the pathological area, avoiding the amplification of noise in normal tissue, thus achieving "sharpening of pathological perception." Specifically, the enhancement of high-frequency details at the edge of the nerve bundle is confined to the lesion area, achieved by combining a mixed weighting coefficient. Balancing details and noise makes small structures such as microcalcifications more prominent.

[0193] Enhanced images of lesions masked as follows Figure 11 As shown.

[0194] S307, Improved Autoencoder Modulus Iterative Training and Parameter Update

[0195] The model training employs a phased iterative optimization strategy, with the following specific steps:

[0196] First, initialize the weight parameters of the dual-channel edge-aware encoder, the lesion region feature decoupling enhancement model, and the adaptive edge-guided decoder;

[0197] Then, the image is reconstructed through forward propagation, and the lesion perception contrast loss and multi-scale gradient consistency loss are calculated sequentially by combining the real labeled image. The two are weighted and summed to form the total loss function with a weight ratio of 1:0.7.

[0198] Then, the Adam optimizer was used to backpropagate and update the network parameters with a learning rate of 0.001. 16 sets of data were input in each batch. Mask enhancement fine-tuning was performed dynamically during training. The binary lesion mask generated by the edge feature map was applied to the reconstructed image to improve the contrast of the lesion area.

[0199] The iteration termination condition meets two criteria simultaneously: first, the total loss of the validation set decreases by less than 0.1% for 15 consecutive epochs, indicating that the model has converged; second, the optimal weights are retained during the plateau period of the validation loss through an early stopping mechanism, and the final model saves the parameters when the test set performance is optimal. The total number of training iterations is controlled within 2000 to avoid the risk of overfitting.

[0200] In one embodiment, the Dice coefficients of different methods in the lesion region are compared, such as Figure 12As shown, this study evaluates the accuracy of different image enhancement methods in segmenting lesion regions of traumatic neuropathic pain. By comparing histogram equalization, conventional autoencoders, U-Net, wavelet enhancement, and the method of this invention, the Dice coefficient was used as a quantitative indicator (range 0-1, with values ​​closer to 1 indicating a higher degree of overlap between the segmentation result and the lesion region labeled by the doctor). Experimental data show that the method of this invention achieves the highest Dice coefficient value, significantly outperforming other methods. Histogram equalization leads to segmentation deviation due to excessive noise enhancement; conventional autoencoders are limited by feature confusion; U-Net, while offering some improvement, is still affected by edge blurring; wavelet enhancement is sensitive to microcalcifications but fails to maintain sufficient continuity of nerve bundles. This invention, through multi-scale edge preservation and feature decoupling mechanisms, improves segmentation accuracy and can accurately separate pathological features from normal tissue.

[0201] S4, Enhanced image output for traumatic neuropathic pain;

[0202] The improved autoencoder model, which has been trained, inputs MRI images of traumatic neuropathic pain and outputs enhanced images. The specific process is as follows:

[0203] Edge and texture features are extracted through multi-scale structural preprocessing to generate preprocessed MRI images;

[0204] Subsequently, a dual-channel edge-aware encoder separates structural features from optimal-scale edge features, and the coupling between pathological features and normal tissue is eliminated through lesion region feature decoupling enhancement.

[0205] The decoding stage utilizes an edge-guided spatial attention mask to modulate the upsampling process and reconstruct a high-resolution image.

[0206] Dynamic mask enhancement fine-tuning is applied to perform unsharpened mask enhancement on the lesion area, and the lesion mask enhancement image is output as the final enhanced image.

[0207] In this implementation, the following is achieved:

[0208] 1. By combining multi-scale Gaussian difference edge extraction and structural tensor enhancement, and using adaptive weights to focus on the most relevant scale, we can achieve fine preservation of blurred lesion boundaries while suppressing noise.

[0209] 2. Independently extract structural features and optimal scale edge features, and fuse them in the encoding stage to solve the problem of uneven feature response caused by different lesion sizes, and improve the separability of lesions and normal tissues.

[0210] 3. By employing dual-path orthogonal constraints combined with spatial weight modulation, the lesion characteristics are completely decoupled from normal tissues, thereby enhancing the salience of the lesions.

[0211] 4. During the decoding stage, edge features are used to generate a spatial attention mask to guide the reconstruction process and maintain the continuity of the neural bundle. Dynamic mask non-sharpening enhancement is used to improve contrast only in the lesion area and avoid noise amplification.

[0212] This invention also proposes an image enhancement processing system for traumatic neuropathic pain, based on the image enhancement processing method for traumatic neuropathic pain described above, comprising:

[0213] The training dataset building module is used to acquire MRI images and build the labeled data required for supervised training.

[0214] The multi-scale structure preprocessing module is used to extract multi-scale edge features by applying the difference of Gaussian operator and weighting the contributions of different scales through adaptive weights. At the same time, it calculates the gradient tensor of the original image, and obtains the structural tensor components by taking its transpose and calculating the Frobenius norm. Then, it achieves the preservation of edge and texture details by weighted addition, avoiding damage to fine structures, and obtains the preprocessed MRI.

[0215] The improved autoencoder model construction and training module includes a dual-channel edge-aware encoder module, a lesion region feature decoupling enhancement module, an adaptive edge-guided decoder module, a lesion-aware contrast loss calculation module, a multi-scale gradient consistency loss calculation module, a dynamic mask enhancement fine-tuning module, and an improved autoencoder model iterative training and parameter update module.

[0216] The traumatic neuropathic pain image enhancement module is used to process new traumatic neuropathic pain MRI images into a trained improved autoencoder model to obtain the final image.

[0217] Furthermore, embodiments of the present invention also propose a computer-readable storage medium storing program instructions for a method of image enhancement processing for traumatic neuropathic pain. These program instructions can be executed by one or more processors to implement the steps of the method of image enhancement processing for traumatic neuropathic pain as described above.

[0218] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for image enhancement of traumatic neuropathic pain, characterized by, include: Step S1, training dataset construction; The training dataset consists of magnetic resonance imaging images of traumatic neuropathic pain. The labeled data required for supervised training is constructed by manually delineating pathological features independently beforehand, and a binary lesion mask is generated after cross-validation. Step S2, multi-scale structure preprocessing; the preprocessed image is obtained based on adaptive weights combined with Gaussian difference operator and gradient tensor; Step S3, Improved autoencoder model construction and training; The improved autoencoder model construction includes lesion region feature decoupling enhancement, forced feature channel independence based on dual-path orthogonal constraint loss, and feature update and decoupling combined with spatial weight modulation; the dual-path orthogonal constraint loss is based on channel slices of structural feature map and channel slices of guiding feature map, evaluates internal orthogonality and external alignment loss through matrix inner product operation, and outputs orthogonal constraint loss, expressed as: ； In the formula, The loss is due to orthogonal constraints. The first structural feature map One channel slice; The first structural feature map One channel slice; The first structural feature map One channel slice; To guide the feature map's first One channel slice; This refers to matrix inner product operations. Orthogonal weights within structural features; for structure-guided feature alignment weights; for the first channel normalization term for the jth channel; The feature update and decoupling combined with spatial weight modulation includes spatial weight modulation based on structural feature map, orthogonal constraint loss and its gradient, through gradient update and Sigmoid activation, updating features along the orthogonal loss gradient direction, and fusing guided features to enhance lesion representation, and outputting a decoupled feature map; Step S4, enhanced output of traumatic neuropathic pain images; input the traumatic neuropathic pain MRI images into the trained improved autoencoder model, and output the enhanced images.

2. The method of claim 1, wherein the method is for enhancing images of neuropathic pain. The improved autoencoder model includes a dual-channel edge-aware encoder, which independently extracts structural features and edge features based on the optimal scale, and then fuses them through channel splicing.

3. The method of claim 1, wherein the method is for enhancing images of neuropathic pain. The improved autoencoder model construction also includes an adaptive edge-guided decoder construction, which uses edge feature maps to generate a spatial attention mask and modulates upsampled features through element-wise multiplication.

4. The method of claim 3, wherein the image enhancement processing is performed on the image of the patient's face. The improved autoencoder model construction also includes calculating the lesion perception contrast loss based on location feature vectors and lesion sampling point sets through cosine similarity calculation and exponential function.

5. The method for image enhancement processing of traumatic neuropathic pain according to claim 4, characterized in that, The improved autoencoder model construction also includes calculating multi-scale gradient consistency loss. By calculating the difference in gradient magnitude between the reconstructed image and the real image at multiple scales, and applying L1 loss constraints, the model is forced to maintain the continuity of the neural tract edge structure at different resolutions.

6. The method of claim 1, wherein the method is for enhancing images of neuropathic pain. The improved autoencoder model construction includes dynamic mask enhancement fine-tuning, local enhancement of the reconstructed image based on a binary lesion mask, mixing the original image with the unsharpened masked enhanced version in the lesion region, and retaining the original reconstruction result in the normal region to obtain the lesion mask enhanced image.

7. The method of claim 1, wherein the method is for enhancing images of traumatic neuropathic pain. The improved autoencoder model training employs a staged iterative optimization strategy, with the following specific steps: Initialize the weight parameters of the dual-channel edge-aware encoder, the lesion region feature decoupling enhancement model, and the adaptive edge-guided decoder; The image is reconstructed by forward propagation, and the lesion perception contrast loss and multi-scale gradient consistency loss are calculated sequentially by combining the real labeled image. The two are weighted and summed to form the total loss function. The Adam optimizer is used to update network parameters through backpropagation. During training, mask enhancement fine-tuning is performed dynamically, and the binary lesion mask generated by the edge feature map is applied to the reconstructed image to improve the contrast of the lesion area.

8. A traumatic neuropathic pain image enhancement processing system, based on the traumatic neuropathic pain image enhancement processing method as described in any one of claims 1 to 7, comprising: The training dataset building module is used to acquire MRI images and build the labeled data required for supervised training. The multi-scale structure preprocessing module is used to extract multi-scale edge features by applying the difference of Gaussian operator and weighting the contributions of different scales through adaptive weights. At the same time, it calculates the gradient tensor of the original image, and obtains the structural tensor components by taking its transpose and calculating the Frobenius norm. Then, it achieves the preservation of edge and texture details by weighted addition, avoiding damage to fine structures, and obtains the preprocessed MRI. The improved autoencoder model construction and training module includes a dual-channel edge-aware encoder module, a lesion region feature decoupling enhancement module, an adaptive edge-guided decoder module, a lesion-aware contrast loss calculation module, a multi-scale gradient consistency loss calculation module, a dynamic mask enhancement fine-tuning module, and an improved autoencoder model iterative training and parameter update module. The traumatic neuropathic pain image enhancement module is used to process new traumatic neuropathic pain MRI images into a trained improved autoencoder model to obtain the final image.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions for a method of image enhancement processing for traumatic neuropathic pain, which can be executed by one or more processors to implement the steps of the method of image enhancement processing for traumatic neuropathic pain as described in any one of claims 1 to 7.

Images