Tumor Image Processing System Based on Image Enhancement

Abstract

The present invention discloses a tumor image processing system based on image enhancement, comprising an image acquisition module, an image enhancement module, a tumor classification training module, and a tumor image processing module. The present invention belongs to the field of image processing, and specifically refers to a tumor image processing system based on image enhancement. This solution introduces an adaptive decomposition network and an attention mechanism to reduce illumination interference and highlight tumor boundaries. It uses illumination contrast attention combined with a scale parameter to avoid noise amplification caused by over-enhancement. It reconstructs high-quality reflectance images, retaining tumor boundaries and texture details while suppressing noise, thereby improving the accuracy of subsequent tumor image classification. It also introduces dynamic similarity estimation and local sparsity constraints, utilizing both local features of tumor boundaries and textures and global features of image distribution, based on dynamic noise control and weight adjustment, to adapt to the complex relationships between tumor image categories, thereby improving tumor image classification results.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a tumor image processing system based on image enhancement. Background Art

[0002] A tumor image processing system is a computerized system based on image processing and machine learning technologies that assists medical professionals in processing tumor images. It processes, enhances, and analyzes input medical images to aid physicians in decision-making. However, typical tumor image processing systems suffer from illumination interference, weak tumor boundary recognition capabilities, and insufficient coordinated optimization of illumination and reflectance reconstruction, leading to low accuracy in subsequent tumor image classification. These systems also suffer from interference from outliers, resulting in unclear boundaries for tumor classification and difficulty in simultaneously considering both tumor region boundary information and overall image distribution, leading to poor tumor image classification results. Summary of the Invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a tumor image processing system based on image enhancement. In view of the problem that general tumor image processing systems have illumination interference, weak tumor boundary recognition ability, insufficient collaborative optimization of illumination and reflection reconstruction, and thus low accuracy of subsequent tumor image classification, this solution introduces an adaptive decomposition network and an attention mechanism to effectively separate the illumination distribution and texture detail information in tumor images, reduce illumination interference, and highlight tumor boundaries; use illumination contrast attention combined with a ratio parameter to dynamically adjust the illumination distribution according to the characteristics of the tumor area, improve the contrast of the tumor area, and avoid noise amplification caused by excessive enhancement; combine with a multi-resolution context-aware network And the multi-scale illumination attention module reconstructs high-quality reflectance images, retains tumor boundaries and texture details, and suppresses noise at the same time, thereby improving the accuracy of subsequent tumor image classification; in view of the problem that general tumor image processing systems have unclear tumor classification boundaries due to interference from outliers, and it is difficult to simultaneously take into account the boundary information of the tumor area and the overall image distribution, which leads to poor tumor image classification effect, this scheme introduces dynamic similarity estimation and local sparsity constraints to make tumor classification boundaries clearer; at the same time, it utilizes the local features of tumor boundaries and textures and the global features of image distribution, based on dynamic noise control and weight adjustment, to adapt to the complex relationship between tumor image categories, thereby improving the tumor image classification effect.

[0004] The technical solution adopted by the present invention is as follows: the tumor image processing system based on image enhancement provided by the present invention includes an image acquisition module, an image enhancement module, a tumor classification training module and a tumor image processing module;

[0005] The image acquisition module acquires a historical tumor image dataset;

[0006] The image enhancement module enhances the illumination distribution and boundary texture of the tumor image through an adaptive decomposition network and a multi-scale optimization strategy to perform image enhancement;

[0007] The tumor classification training module performs tumor classification training;

[0008] The tumor image processing module assists in classifying tumor images collected in real time.

[0009] Furthermore, in the image acquisition module, the historical tumor image dataset includes tumor images and tumor types.

[0010] Furthermore, the image enhancement module specifically includes the following contents:

[0011] Decomposition unit: An adaptive decomposition network is introduced to separate illumination and reflection information through an attention mechanism, decomposing the tumor image into an illumination map and a reflection map. The illumination map represents the illumination distribution of the tumor image, and the reflection map represents the tumor boundary and texture detail information of the tumor image. The decomposition is represented as follows: ;reflection Figure 1 Causative loss Expressed as: ; Light map smoothness loss Expressed as: ; Lighting gradient consistency loss Expressed as: ;in, is the original tumor image, x and y are the image coordinates; It is a reflection map; It is a light map; is the noise map; and They are the low-quality and high-quality reflection maps after decomposition; is the L1 norm; is the gradient; c is the weight for adjusting gradient consistency; Denoise(·) is the denoising result obtained based on DnCNN; It is a weight map generated based on the attention mechanism; and They are low-quality light maps and high-quality light maps;

[0012] Light adjustment unit: adjusts the light intensity through the ratio parameter r, introduces light contrast attention, and inputs the light map Ls output by the decomposition module. The adaptive light enhancement is expressed as: ; Lighting adjustment unit loss function Expressed as: ;in, is the illumination enhancement output; MSE(·) is the mean square error; Lk is the ideal illumination map; and They are the predicted and target gradient illumination maps respectively; It is the tumor region feature extracted by convolutional neural network; It is a weight map generated based on illumination contrast attention;

[0013] Reflection restoration unit; uses a multi-resolution context-aware network to integrate multi-scale context information when restoring the reflection map; extracts multi-scale features of the illumination map based on the multi-scale illumination attention module, generates dynamic weights, and introduces context information to guide the reconstruction of the reflection map. Expressed as: ;Reflection recovery unit loss function express: ; The final output is a synthetic enhanced tumor image , expressed as: ;in, It is a multi-scale illumination attention module; is the original reflectance map; Rh is the ideal reflectance map; SSIM(·) is the structural similarity index; CF is the contextual information of the tumor region extracted by the region growing algorithm; is the recovery weight; It is based on a multi-resolution context-aware network; is a residual network; is the residual coefficient.

[0014] Furthermore, the tumor classification training module specifically includes the following contents:

[0015] Dynamic similarity graph learning unit; introduce robust dynamic similarity estimation, add noise suppression terms to the initial similarity formula, and construct the initial similarity between image features , expressed as: ; Introduce the constraint of local sparsity and optimize the similarity matrix, which can be expressed as: ; Weight matrix update, through reweighted optimization to reduce the influence of outliers, expressed as: ;in, and are two image features, i and j are image feature indices; is the Gaussian kernel parameter; N is the total number of image features; is the similarity regularization parameter; S is the similarity matrix; A is the weight matrix; is element-wise multiplication; is the square of the Frobenius norm; is the category center; is a parameter that controls the degree of noise suppression; is the constraint that controls local sparsity; and is the adjustment parameter;

[0016] Laplace rank constraint unit; construct weighted Laplace matrix , expressed as: ; Rank constraint means: ; Using eigenvalue optimization, it is expressed as: ; Where D is the degree matrix, which represents the total connection weight of each sample; is the weighted similarity matrix, is a matrix element; are the elements of the degree matrix; rank(·) is the rank constraint; C is the number of tumor classes; tr(·) is the trace of the matrix; is the feature regularization parameter; F is the classification matrix; T is the matrix transpose;

[0017] Dual-feature regularization unit; feature extraction, introducing category sparsity constraints, constructing the projection matrix P, expressed as: , the constraints are expressed as: ; The adaptive regression feature matrix W is expressed as: ;in, is the unit moment of dimension m; is the projection regularization parameter; X is the feature vector extracted from the tumor image; 1 refers to a column vector of all 1s; is the weighted norm; is a parameter that controls feature sparsity; is the reference matrix of global features; is a parameter that controls global consistency; is the ideal characteristic matrix; b is the bias;

[0018] Adaptive regression optimization unit; dynamically adjust the adaptive regression feature matrix W and bias b to adapt to the complex relationship between categories in tumor images; the bias b update is expressed as: ; The adaptive regression feature matrix update is expressed as: ; Where H is the weight matrix; is the regularization term;

[0019] The total angle optimization unit is integrated; dynamic similarity graph learning, dual-feature regularization, Laplace rank constraint and adaptive regression are unified to adapt to the tumor image classification task; a dynamic weight balancing mechanism is introduced, and the overall optimization is expressed as: ;in, is the loss weight;

[0020] Weight adaptation dynamic adjustment unit; adjustment based on the first k minimum distances of each image feature , introducing the category distribution weight, expressed as: ;in, and are the distances between the i1th image and the k+1th minimum distance image and the j1th image respectively; is the weight adjustment coefficient of the i1th image; is the category balance factor; N1 is the total number of images;

[0021] Eigen decomposition optimization unit; used to obtain the projection matrix P and classification matrix F; the projection matrix decomposition optimization is expressed as: ; Obtain the eigenvectors corresponding to the first m minimum eigenvalues ​​through eigendecomposition; The classification matrix decomposition optimization is expressed as: ;

[0022] Convergence judgment unit; pre-set the accuracy threshold and maximum training times; divide the image enhanced data set into a training set and a test set, start training the tumor classification training module based on the training set, when the similarity matrix S and the classification matrix F converge or reach the maximum training times, the training iteration stops and the classification result is obtained; when the classification accuracy of the trained tumor classification training module for the test set is higher than the accuracy threshold, the tumor classification training module is established.

[0023] Furthermore, the tumor image processing module collects tumor image data in real time, and after being processed by the image enhancement module, inputs the data into the tumor classification training module, and outputs the classification result output by the tumor classification training module as the output of the tumor image processing module.

[0024] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0025] (1) In view of the problem that general tumor image processing systems have illumination interference, weak tumor boundary recognition ability, and insufficient coordinated optimization of illumination and reflection reconstruction, which leads to low accuracy of subsequent tumor image classification, this scheme introduces an adaptive decomposition network and attention mechanism to effectively separate the illumination distribution and texture detail information in tumor images, reduce illumination interference, and highlight tumor boundaries; use illumination contrast attention combined with a scale parameter to dynamically adjust the illumination distribution according to the characteristics of the tumor area, improve the contrast of the tumor area, and avoid noise amplification caused by excessive enhancement; combine a multi-resolution context-aware network and a multi-scale illumination attention module to reconstruct a high-quality reflection map, retain the tumor boundary and texture details, and suppress noise; thereby improving the accuracy of subsequent tumor image classification.

[0026] (2) In view of the problem that the general tumor image processing system has unclear boundaries for tumor classification due to the interference of outliers, and it is difficult to take into account the boundary information of the tumor area and the overall image distribution at the same time, which leads to poor tumor image classification effect, this scheme introduces dynamic similarity estimation and local sparsity constraints to make the tumor classification boundaries clearer; at the same time, it utilizes the local features of tumor boundaries and textures and the global features of image distribution, based on dynamic noise control and weight adjustment, to adapt to the complex relationship between tumor image categories, thereby improving the tumor image classification effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic diagram of the principle of the tumor image processing system based on image enhancement provided by the present invention;

[0028] Figure 2 Schematic diagram of the principle of tumor classification training module.

[0029] 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. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0031] In the description of the present invention, it should be understood that terms such as "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they should not be understood as limiting the present invention.

[0032] Example 1, see Figure 1 The tumor image processing system based on image enhancement provided by the present invention includes an image acquisition module, an image enhancement module, a tumor classification training module and a tumor image processing module;

[0033] The image acquisition module collects historical tumor image data sets and sends the data to the image enhancement module;

[0034] The image enhancement module enhances the illumination distribution and boundary texture of the tumor image through an adaptive decomposition network and a multi-scale optimization strategy, and performs image enhancement; and sends the data to the tumor classification training module;

[0035] The tumor classification training module performs tumor classification training; and sends the data to the tumor image processing module;

[0036] The tumor image processing module assists in classifying tumor images collected in real time.

[0037] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the image acquisition module, the historical tumor image data set includes tumor images and tumor types; the tumor types include skin tumors, breast tumors, thyroid tumors, head and neck tumors, and soft tissue tumors.

[0038] Example 3, see Figure 1 This embodiment is based on the above embodiment, and the image enhancement module specifically includes the following contents:

[0039] Decomposition unit: An adaptive decomposition network is introduced to separate illumination and reflection information through an attention mechanism, decomposing the tumor image into an illumination map and a reflection map. The illumination map represents the illumination distribution of the tumor image, and the reflection map represents the tumor boundary and texture detail information of the tumor image. The decomposition is represented as follows: ;reflection Figure 1 Causative loss Expressed as: ; Light map smoothness loss Expressed as: ; Lighting gradient consistency loss Expressed as: ;in, is the original tumor image, x and y are the image coordinates; It is a reflection map; It is a light map; is the noise map; and They are the low-quality and high-quality reflection maps after decomposition; is the L1 norm; is the gradient; c is the weight for adjusting gradient consistency; Denoise(·) is the denoising result obtained based on DnCNN; It is a weight map generated based on the attention mechanism, which is used to suppress the reflection information in the illumination map; and They are low-quality light maps and high-quality light maps;

[0040] Light adjustment unit: Enhance the light distribution to make the tumor area brighter and easier to identify, while avoiding noise amplification caused by over-enhancement. Adjust the light intensity through the scale parameter r, introduce light contrast attention, and input the light map Ls output by the decomposition module. The adaptive light enhancement is expressed as: ; Lighting adjustment unit loss function Expressed as: ;in, is the illumination enhancement output; MSE(·) is the mean square error; Lk is the ideal illumination map; and They are the predicted and target gradient illumination maps respectively; It is the tumor region feature extracted by convolutional neural network; It is a weight map generated based on illumination contrast attention, which is used to preserve edge details and suppress overly bright areas;

[0041] Reflection restoration unit; suppresses noise, restores tumor boundaries and internal texture details, and enhances image quality; uses a multi-resolution context-aware network to fuse multi-scale context information when restoring the reflectance map, enhancing boundary and texture details; extracts multi-scale features of the illumination map based on the multi-scale illumination attention module, generates dynamic weights, and introduces context information to guide the reconstruction of the reflectance map. Expressed as: ;Reflection recovery unit loss function express: ; The final output is a synthetic enhanced tumor image , expressed as: ;in, It is a multi-scale illumination attention module; is the original reflectance map; Rh is the ideal reflectance map; SSIM(·) is the structural similarity index; CF is the contextual information of the tumor region extracted by the region growing algorithm; is the recovery weight; It is based on a multi-resolution context-aware network; is a residual network; is the residual coefficient.

[0042] By performing the above operations, this solution introduces an adaptive decomposition network and attention mechanism to address the problems of illumination interference, weak tumor boundary recognition ability, and insufficient coordinated optimization of illumination and reflection reconstruction in general tumor image processing systems, which in turn leads to low accuracy in subsequent tumor image classification. This effectively separates the illumination distribution and texture detail information in tumor images, reduces illumination interference, and highlights tumor boundaries. It uses illumination contrast attention combined with a scale parameter to dynamically adjust the illumination distribution according to the characteristics of the tumor area, enhance the contrast of the tumor area, and avoid noise amplification caused by excessive enhancement. It combines a multi-resolution context-aware network and a multi-scale illumination attention module to reconstruct high-quality reflection maps, retain tumor boundaries and texture details, and suppress noise, thereby improving the accuracy of subsequent tumor image classification.

[0043] Example 4, see Figure 1 and Figure 2 This embodiment is based on the above embodiment, and the tumor classification training module specifically includes the following contents:

[0044] Dynamic similarity graph learning unit; constructs a dynamic similarity matrix between samples to capture the fuzzy boundaries between different categories in tumor images; introduces robust dynamic similarity estimation, adds noise suppression terms to the initial similarity formula, and constructs the initial similarity between image features , expressed as: ; Introduce the constraint of local sparsity and optimize the similarity matrix, which can be expressed as: ; Weight matrix update, through reweighted optimization to reduce the influence of outliers, expressed as: ;in, and are two image features, i and j are image feature indices; is the Gaussian kernel parameter; N is the total number of image features; is the similarity regularization parameter; S is the similarity matrix; A is the weight matrix; is element-wise multiplication; is the square of the Frobenius norm; is the category center; is a parameter that controls the degree of noise suppression; is the constraint that controls local sparsity; and is the adjustment parameter;

[0045] Laplace rank constraint unit; ensure that the number of connected components of the similarity matrix is ​​equal to the number of tumor categories C, making the classification boundaries clearer; construct a weighted Laplace matrix , expressed as: ; Rank constraint means: ; Using eigenvalue optimization, it is expressed as: ; Where D is the degree matrix, which represents the total connection weight of each sample; is the weighted similarity matrix, is a matrix element; are the elements of the degree matrix; rank(·) is the rank constraint; C is the number of tumor classes; tr(·) is the trace of the matrix; is the feature regularization parameter; F is the classification matrix; T is the matrix transpose;

[0046] Dual-feature regularization unit; jointly extract information from local and global features of tumor image data to optimize classification performance; feature extraction, introduce category sparsity constraints, and construct the projection matrix P, which is expressed as: , the constraints are expressed as: ; The adaptive regression feature matrix W is expressed as: ;in, is the unit moment of dimension m; is the projection regularization parameter; X is the feature vector extracted from the tumor image; 1 refers to a column vector of all 1s; is the weighted norm; is a parameter that controls feature sparsity; is the reference matrix of global features; is a parameter that controls global consistency; is the ideal characteristic matrix; b is the bias;

[0047] Adaptive regression optimization unit; dynamically adjust the adaptive regression feature matrix W and bias b to adapt to the complex relationship between categories in tumor images; the bias b update is expressed as: ; The adaptive regression feature matrix update is expressed as: ; Where H is the weight matrix; It is a regularization term used to suppress feature overfitting;

[0048] The total angle optimization unit is integrated; dynamic similarity graph learning, dual-feature regularization, Laplace rank constraint and adaptive regression are unified to adapt to the tumor image classification task; a dynamic weight balancing mechanism is introduced, and the overall optimization is expressed as: ;in, is the loss weight;

[0049] Weight adaptation dynamic adjustment unit; used to adaptively control the weight of noise points in tumor images; adjusted according to the first k minimum distances of each image feature , introducing the category distribution weight, expressed as: ;in, and are the distances between the i1th image and the k+1th minimum distance image and the j1th image respectively; is the weight adjustment coefficient of the i1th image; is the category balance factor; N1 is the total number of images;

[0050] Eigen decomposition optimization unit; used to obtain the projection matrix P and classification matrix F; the projection matrix decomposition optimization is expressed as: ; Obtain the eigenvectors corresponding to the first m minimum eigenvalues ​​through eigendecomposition; The classification matrix decomposition optimization is expressed as: ;

[0051] Convergence judgment unit; pre-set the accuracy threshold and maximum training times; divide the image enhanced data set into a training set and a test set, start training the tumor classification training module based on the training set, when the similarity matrix S and the classification matrix F converge or reach the maximum training times, the training iteration stops and the classification result is obtained; when the classification accuracy of the trained tumor classification training module for the test set is higher than the accuracy threshold, the tumor classification training module is established.

[0052] By performing the above operations, this solution introduces dynamic similarity estimation and local sparsity constraints to make the tumor classification boundaries clearer due to the interference of outliers in general tumor image processing systems, making it difficult to simultaneously consider the boundary information of the tumor area and the overall image distribution, which in turn leads to poor tumor image classification results. At the same time, it utilizes the local features of tumor boundaries and textures and the global features of image distribution, based on dynamic noise control and weight adjustment, to adapt to the complex relationship between tumor image categories, thereby improving the tumor image classification effect.

[0053] Example 5, see Figure 1 This embodiment is based on the above embodiment. The tumor image processing module collects tumor image data in real time, and after being processed by the image enhancement module, it is input into the tumor classification training module, and the classification result output by the tumor classification training module is output as the tumor image processing module.

[0054] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0055] While the embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that various changes, modifications, substitutions, and alterations can be made to the embodiments without departing from the principles and spirit of the invention.

[0056] The present invention and its embodiments are described above. This description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A tumor image processing system based on image enhancement, characterized by: The system includes an image acquisition module, an image enhancement module, a tumor classification training module, and a tumor image processing module; The image acquisition module acquires a historical tumor image dataset; The image enhancement module enhances the illumination distribution and boundary texture of the tumor image through an adaptive decomposition network and a multi-scale optimization strategy to perform image enhancement; The tumor classification training module uses dynamic similarity graph learning, Laplace rank constraint and dual feature regularization to perform tumor classification training; The tumor image processing module assists in classifying tumor images collected in real time; The image enhancement module includes the following contents: a decomposition unit; An adaptive decomposition network is introduced to separate illumination and reflection information through an attention mechanism, decomposing the tumor image into an illumination map and a reflection map. The illumination map represents the illumination distribution of the tumor image, and the reflection map represents the tumor boundary and texture details of the tumor image. The decomposition is represented as follows: ;Reflection map consistency loss Expressed as: ; Light map smoothness loss Expressed as: ; Lighting gradient consistency loss Expressed as: ;in, is the original tumor image, x and y are the image coordinates; It is a reflection map; It is a light map; is the noise map; and They are the low-quality and high-quality reflection maps after decomposition; is the L1 norm; is the gradient; c is the weight for adjusting gradient consistency; Denoise(·) is the denoising result obtained based on DnCNN; and They are low-quality light maps and high-quality light maps; It is a weight map generated based on the attention mechanism; The image enhancement module specifically includes the following contents: Decomposition unit; Light adjustment unit: adjusts the light intensity through the ratio parameter r, introduces light contrast attention, and inputs the light map Ls output by the decomposition unit. The adaptive light enhancement is expressed as: ; Lighting adjustment unit loss function Expressed as: ;in, is the illumination enhancement output; MSE(·) is the mean square error; Lk is the ideal illumination map; and They are the predicted and target gradient illumination maps respectively; It is the tumor region feature extracted by convolutional neural network; It is a weight map generated based on illumination contrast attention; Reflection restoration unit; uses a multi-resolution context-aware network to integrate multi-scale context information when restoring the reflection map; extracts multi-scale features of the illumination map based on the multi-scale illumination attention module, generates dynamic weights, and introduces context information to guide the reconstruction of the reflection map. Expressed as: ;Reflection recovery unit loss function express: ; The final output is a synthetic enhanced tumor image , expressed as: ;in, It is a multi-scale illumination attention module; is the original reflectance map; Rh is the ideal reflectance map; SSIM(·) is the structural similarity index; CF is the contextual information of the tumor region extracted by the region growing algorithm; is the recovery weight; It is a multi-resolution context-aware network; is a residual network; is the residual coefficient.

2. The tumor image processing system based on image enhancement according to claim 1, characterized in that: The tumor classification training module specifically includes the following contents: Dynamic similarity graph learning unit; introduce robust dynamic similarity estimation, add noise suppression terms to the initial similarity formula, and construct the initial similarity between image features , expressed as: ; Introduce the constraint of local sparsity and optimize the similarity matrix, which can be expressed as: ; Weight matrix update, through reweighted optimization to reduce the influence of outliers, expressed as: ;in, and are two image features, i and j are image feature indices; is the Gaussian kernel parameter; N is the total number of image features; is the similarity regularization parameter; S is the similarity matrix; is the weight matrix; is element-wise multiplication; is the square of the Frobenius norm; is the category center; is a parameter that controls the degree of noise suppression; is the constraint that controls local sparsity; and is the adjustment parameter; Laplace rank constraint unit; construct weighted Laplace matrix , expressed as: ; Rank constraint means: ; Using eigenvalue optimization, it is expressed as: ; Where D is the degree matrix, which represents the total connection weight of each sample; is the weighted similarity matrix, is a matrix element; are the elements of the degree matrix; rank(·) is the rank constraint; C is the number of tumor classes; tr(·) is the trace of the matrix; is the feature regularization parameter; F is the classification matrix; T is the matrix transpose; Dual-feature regularization unit; feature extraction, introducing category sparsity constraints, constructing the projection matrix P, expressed as: , the constraints are expressed as: ; The adaptive regression feature matrix W is expressed as: ;in, is the unit moment of dimension m; is the projection regularization parameter; X is the feature vector extracted from the tumor image; 1 refers to a column vector of all 1s; is the weighted norm; is a parameter that controls feature sparsity; is the reference matrix of global features; is a parameter that controls global consistency; is the ideal characteristic matrix; Adaptive regression optimization unit; dynamically adjust the adaptive regression feature matrix W and bias b to adapt to the complex relationship between categories in tumor images; the bias b update is expressed as: ; The adaptive regression feature matrix update is expressed as: ; ; Where H is the weighting matrix; is the regularization term; The total angle optimization unit is integrated; dynamic similarity graph learning, dual-feature regularization, Laplace rank constraint and adaptive regression are unified to adapt to the tumor image classification task; a dynamic weight balancing mechanism is introduced, and the overall optimization is expressed as: ;in, is the loss weight; Weight adaptation dynamic adjustment unit; adjustment based on the first k minimum distances of each image feature , introducing the category distribution weight, expressed as: ; ;in, and are the distances between the i1th image and the k+1th minimum distance image and the j1th image respectively; is the weight adjustment coefficient of the i1th image; is the category balance factor; N1 is the total number of images; Eigen decomposition optimization unit; used to obtain the projection matrix P and classification matrix F; the projection matrix decomposition optimization is expressed as: ; Obtain the eigenvectors corresponding to the first m minimum eigenvalues ​​through eigendecomposition; The classification matrix decomposition optimization is expressed as: ; Convergence judgment unit; pre-set the accuracy threshold and maximum training times; divide the image enhanced data set into a training set and a test set, start training the tumor classification training module based on the training set, when the similarity matrix S and the classification matrix F converge or reach the maximum training times, the training iteration stops and the classification result is obtained; when the classification accuracy of the trained tumor classification training module for the test set is higher than the accuracy threshold, the tumor classification training module is established.

3. The tumor image processing system based on image enhancement according to claim 2, characterized in that: In the image acquisition module, the historical tumor image dataset includes tumor images and tumor types.

4. The tumor image processing system based on image enhancement according to claim 3, characterized in that: The tumor image processing module collects tumor image data in real time, and after being processed by the image enhancement module, inputs it into the tumor classification training module, and outputs the classification result output by the tumor classification training module as the output of the tumor image processing module.

Images