Thoracic surgery pathological section image super-resolution reconstruction method

Abstract

The invention provides a thoracic surgery pathological section image super-resolution reconstruction method, and relates to the technical field of image processing, and the method specifically comprises the steps: collecting a high-resolution thoracic surgery pathological section image, and generating a low-resolution image; constructing a local directional dot matrix tensor and a response adjustment factor based on the low-resolution image, calculating an angle response aggregation kernel tensor, and performing segmented mapping in combination with a nonlinear response reconstruction function to complete directional fusion normalization so as to obtain a directional aggregation enhanced image; a robust local reference value and local texture energy are calculated in a neighborhood range, a structural strength weight is generated in combination with a variance balance parameter and a contrast balance parameter, a difference amplification item is constructed, image fusion with direction aggregation is enhanced through a residual mode, and a structure guide enhanced image is obtained; carrying out weighted mean and variance calculation to generate a difference regulation factor, and combining convolution smoothing and nonlinear amplitude limiting processing to obtain a structure contrast mapping image; and finally, inputting the image into an image reconstruction module to realize resolution improvement.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing, in particular to a thoracic surgery pathological section image super-resolution reconstruction method. BACKGROUND

[0002] At present, pathological section images are important basis for diagnosis and efficacy evaluation of thoracic surgery diseases, and their resolution and clarity directly affect the recognition accuracy of cell morphology, gland contour and nuclear chromatin characteristics. With the development of digital pathology, whole slide section (WSI) has become the core form of digital pathology, and the full-width scanning of tissue sections is usually realized through 20x or 40x objective lens. However, the 40x WSI data volume is huge, and the storage and transmission cost is extremely high, which limits its application in large-scale clinical and scientific research scenarios. Therefore, 20x WSI is often used as input and restored to an image quality close to 40x WSI through super-resolution method. However, direct magnification of 20x WSI has significant problems: the cell nucleus and cytoplasm boundary is blurred, resulting in unclear morphological characteristics, affecting tumor grading and micro-infiltration recognition; the gland contour and fiber structure appear broken or artifacts after magnification, which cannot accurately reflect the tissue continuity; the tissue texture shows coexistence of noise amplification and excessive smoothing at high magnification, which not only makes the homogeneous region unnatural, but also causes the loss of details in key areas, thereby affecting the reliability and stability of thoracic surgery pathological sections in fine diagnosis and computer-aided analysis.

[0003] Publication No. CN118967449A proposes a pathological section image super-resolution method based on diffusion model, which obtains multi-level features through a hybrid feature extraction module combined with global attention and local convolution, enhances image features in noise adding and denoising iteration using a diffusion transformer network, and finally generates a high-resolution section image through a feature fusion module. Publication No. CN120450963A proposes a medical image super-resolution method based on dynamic attention and implicit neural representation, which continuously expresses and reconstructs medical images by constructing a convolutional neural network and introducing a dynamic attention mechanism, while combining implicit neural representation.

[0004] However, the existing technology often fails to fully alleviate the problems of insufficient texture gradient and blurred edges in low-resolution images in the super-resolution reconstruction process of thoracic surgery pathological section images, lacks effective compensation mechanism for weakened information, and ignores the contrast imbalance and excessive enhancement problems that are prone to occur in tissue structure adhesion and overlapping areas, resulting in incomplete detail recovery, excessive enhancement or insufficient smoothing of homogeneous areas in the super-resolution reconstruction task. SUMMARY

[0005] The application provides a thoracic surgery pathological section image super-resolution reconstruction method, aiming to construct a thoracic surgery pathological section image super-resolution reconstruction framework with texture detail enhancement and structure continuity maintenance capability; the method gradually strengthens the edge profile and subtle features of low-contrast areas of the image by acquiring a low-resolution thoracic surgery pathological section image, combining a discrete polykernels tensor transformation, a local structure guided mapping and a structure contrast mapping module, finally inputs the enhancement result into an image reconstruction module to realize image resolution improvement and complete high-quality super-resolution reconstruction of the thoracic surgery pathological section image.

[0006] In order to achieve the above-mentioned purpose, the application provides the following technical scheme: a thoracic surgery pathological section image super-resolution reconstruction method, and the specific steps are as follows: Acquire a high-resolution thoracic surgery pathological section image, generate a corresponding low-resolution thoracic surgery pathological section image, and construct a thoracic surgery pathological section image dataset; According to the low-resolution thoracic surgery pathological section image, calculate a local directionality point array tensor and a response adjustment factor, construct a direction response polykernel tensor, and process the same by combining a structure adaptive threshold and a nonlinear response reconstruction mechanism to obtain a direction aggregation enhancement image; According to the direction aggregation enhancement image, calculate a robust local reference value in a neighborhood range, construct a local texture energy, calculate a structure intensity weight by combining a variance balance parameter and a contrast balance parameter, generate a difference amplification term, and fuse the same with the enhancement image in a residual manner to obtain a structure guided enhancement image; In the neighborhood range, calculate a weighted mean and a variance of the structure guided enhancement image, introduce a variance balance parameter and a contrast balance parameter to calculate a structure intensity weight, construct a difference regulation factor, perform convolution smoothing and nonlinear limiting processing, and fuse the same with the structure guided enhancement image to obtain a structure contrast mapping image; Input the structure contrast mapping image into an image reconstruction module to improve the resolution; Construct a thoracic surgery pathological section image super-resolution reconstruction model, train the model to convergence by taking a high-resolution image as a supervision signal, and realize super-resolution reconstruction of the thoracic surgery pathological section image.

[0007] Preferably, in the S1 step, the construction of the thoracic surgery pathological section image dataset comprises: firstly, full-width scanning of the tissue section is performed using a 40x objective lens to ensure the integrity and detail fidelity of the sample, and in the process, the samples of the section are rejected if there is abnormal staining distribution, focal plane deviation or blurred tissue structure, so as to obtain high-quality thoracic surgery pathological section images; then, standardization of the thoracic surgery pathological section images is performed, the whole section is cropped into fixed-size image blocks, high-resolution thoracic surgery pathological section images are generated, and the high-resolution thoracic surgery pathological section images are taken as label images; then, a bicubic interpolation algorithm is used to downsample the label image blocks to obtain corresponding low-resolution thoracic surgery pathological section images; further, normalization processing is uniformly performed on the low-resolution images and the high-resolution images to maintain the consistency and stability of the data distribution; finally, the low-resolution images and their corresponding high-resolution label images are paired to form a thoracic surgery pathological section image dataset for training and testing the super-resolution model, and a unified and standardized data input is provided for subsequent thoracic surgery pathological section image super-resolution reconstruction.

[0008] Preferably, the low-resolution thoracic surgery pathological section images generally have problems of unobvious texture gradient change, blurred transition of cell and tissue boundaries and background, and the traditional method is prone to response attenuation in the multi-directional feature extraction process, thereby causing missing of detail texture; based on such characteristics, the application proposes a discrete polykernels tensor transformation module, which constructs a local directional point array tensor and a response adjustment factor within the range of an angle set and a radius set, combines a structure adaptive threshold and a nonlinear response reconstruction mechanism, realizes aggregation and differential enhancement of multi-directional features, performs directional fusion normalization calculation, generates a direction aggregation and enhancement image, effectively alleviates the problems of insufficient texture gradient and blurred edges in the low-resolution image, and improves the clarity of the tissue texture and edge structure.

[0009] Further, in the S2 step, the direction aggregation and enhancement image is generated, and the specific process comprises: According to the low-resolution thoracic surgery pathological section image, a plurality of offset positions are determined in the center of the current position according to a preset angle set and a radius set, the Euclidean distance square of the feature vectors of the pixels in the current position and the offset positions in the channel dimension is calculated, and exponential mapping is performed in combination with a similarity decay factor to construct a local directional point array tensor; In the angle set and the radius set, the double-angle cosine of the angle is taken as an angle adjustment factor, and exponential decay is performed in combination with a radius and a scale decay parameter, while a direction gain coefficient is introduced for amplitude adjustment to construct a response adjustment factor; The local directional point array tensor and the response adjustment factor are multiplied item by item under all angle and radius combinations, and are accumulated at the corresponding pixel positions to construct a directional response aggregation kernel tensor; The difference between the current pixel and its horizontally adjacent pixel is taken as a horizontal difference item, and the difference between the current pixel and its vertically adjacent pixel is taken as a vertical difference item, and the absolute value of the difference between the two is accumulated and normalized in the whole image range to obtain a structure adaptive threshold value; At each pixel position, when the directional response aggregation kernel tensor is greater than the structure adaptive threshold value, a logarithmic enhancement term constructed by the directional response aggregation kernel tensor is superimposed on the eigenvalue of the low-resolution thoracic surgical pathology section image; when the directional response aggregation kernel tensor is less than or equal to the structure adaptive threshold value, an exponential decay term constructed by the directional response aggregation kernel tensor is applied to the eigenvalue of the low-resolution thoracic surgical pathology section image to obtain a nonlinear response reconstruction tensor. Under all combinations of radius and angle, the values of the nonlinear response reconstruction tensor at the corresponding offset positions are weighted and accumulated according to the aggregation adjustment factor to obtain the weighted aggregation response value of each pixel; the sum of the aggregation adjustment factors is taken as a normalization factor to normalize the weighted aggregation response value to obtain a directional aggregation enhancement image.

[0010] Further, in the S2 step, in the process of generating the direction aggregation enhancement image, first, in the angle and radius set, the Euclidean distance of the feature vector of the center pixel and each offset position in the channel dimension is calculated according to the low-resolution thoracic surgery pathology section image, and an exponential mapping is performed in combination with a similarity decay factor to construct a local direction point lattice tensor, which is used to accurately characterize the local difference characteristics of the pixels in different directions and distances, and can reflect the direction sensitivity of the local neighborhood and provide a basic representation for the subsequent direction response; then, the double-angle cosine of the angle is taken as the angle adjustment factor, combined with the radius and scale decay parameters and introducing the direction gain coefficient to form the response adjustment factor, which is used for adaptive weight allocation of the local difference in different directions and scales, highlighting the contribution of the main direction to the texture and edge and suppressing the interference of the non-main direction; under all angle and radius combinations, multiply them item by item and accumulate at the corresponding position to obtain the direction response aggregation kernel tensor, which is used to fuse the local response results under multiple directions and scales to form a comprehensive direction sensitive response map, providing global guidance for the enhancement of the salient region and texture structure; then, a structure adaptive threshold is constructed by the horizontal and vertical difference, which is used to distinguish the significant response and non-significant response according to the anisotropic distribution of the image structure, ensuring that the response enhancement is realized in the edge and texture area, while avoiding over-enhancement or false texture in the smooth area; at each pixel, when the aggregation kernel value is higher than the threshold, a logarithmic enhancement term constructed by the direction response is superimposed on the original feature, otherwise an exponential suppression is applied, thereby generating a nonlinear response reconstruction tensor, which is used to highlight the salient edges and texture details while effectively suppressing noise and reducing overshoot artifacts, ensuring the stability and reliability of the reconstruction result; finally, the nonlinear response reconstruction tensor is weighted and accumulated at all offset positions of the angle and radius, and the weight sum is normalized to obtain the direction aggregation enhancement image, which is clearer in texture level and more continuous in boundary level, and has strong noise robustness, which can be used as high-quality input data for subsequent identification and analysis modules, improving the reliability and accuracy of the overall system.

[0011] Preferably, the low-resolution thoracic surgery pathology section image is prone to local detail loss in the cell structure and low-contrast area, and traditional methods are difficult to balance between smooth background and fine structure, resulting in insufficient expression of tissue texture; based on such characteristics, the local structure guided mapping module is proposed, which calculates the structure strength weight by statistically analyzing the robust local reference value and local texture energy in the neighborhood, combines the variance balance parameter and the contrast balance parameter, adjusts the difference between the target pixel and the reference difference, and fuses with the direction aggregation enhancement image in a residual manner to obtain a structure guided enhancement image, thereby effectively alleviating the problem of detail weakening and improving the integrity and continuity of the overall tissue contour.

[0012] Further, in the S3 step, the structure-guided enhancement image is generated, and the specific process includes: According to the directional aggregation enhancement image, the feature values in the neighborhood range with the target position as the reference are sorted, the extreme value samples at both ends are removed according to a preset truncation ratio, the remaining part is accumulated and averaged, and a robust local reference value is obtained; Under the reference of the robust local reference value, the difference between each pixel value in the neighborhood of the directional aggregation enhancement image and the reference value is squared and accumulated, and then normalized according to the total number of neighborhood pixels to obtain a local texture energy; The local texture energy and the variance balance parameter are combined to form a suppression term, and the absolute value of the difference between the directional aggregation enhancement image and the robust local reference value is taken to form a regulation term together with the contrast balance parameter; the suppression term and the regulation term are multiplied to obtain a structure strength weight; The difference between the directional aggregation enhancement image and the robust local reference value is scaled by the structure strength weight to obtain a difference amplification term; On the basis of the difference amplification term, a nonlinear gain coefficient is introduced to modulate its amplitude, and a nonlinear voltage control processing is performed on the modulated difference amplification term combined with an activation function; the nonlinear amplitude-limited difference amplification term and the directional aggregation enhancement image are fused through a residual connection to obtain a structure-guided enhancement image.

[0013] Further, in the process of generating the structure-guided enhanced image in S3, first, the pixel feature values in the neighborhood range with the target position as the reference are sorted, and the extreme value samples at both ends are removed according to a preset truncation ratio, then the remaining part is accumulated and averaged to obtain a robust local reference value, which is used to provide a stable local reference to avoid interference of abnormal pixels on subsequent calculation; then, the square of the difference between each pixel value in the neighborhood and the robust local reference value is accumulated and normalized combined with the total number of neighborhood pixels to obtain a local texture energy, which is used to measure the texture complexity of the local region and provide a basis for weight design; then, the suppression term composed of the local texture energy and the variance balance parameter is combined with the adjustment term composed of the absolute value of the difference between the direction aggregation enhanced image and the robust local reference value and the contrast balance parameter, and the structure strength weight is obtained by multiplying the two terms, which is used to dynamically regulate the amplification degree of local difference and ensure the balance of detail enhancement; on this basis, the difference amplification term is obtained by scaling the difference between the direction aggregation enhanced image and the robust local reference value with the structure strength weight, which is used to highlight the significant local structural difference; finally, the non-linear gain coefficient is introduced on the basis of the difference amplification term for modulation, and the non-linear voltage control is completed through the tanh activation function, and the structure-guided enhanced image is obtained by adding the residual connection to the direction aggregation enhanced image, which can enhance the detail and edge information while maintaining the overall stability, providing enhanced input features for subsequent processing.

[0014] Preferably, low-resolution thoracic surgery pathology section images often have contrast imbalance and over-enhancement risk in tissue structure adhesion and overlapping areas, and traditional methods are prone to artifacts when processing, affecting image authenticity; based on such characteristics, the present application proposes a structure contrast mapping module, which calculates the weighted mean and variance of the structure-guided enhanced image in the neighborhood range to generate a difference control factor, and combines Gaussian convolution smoothing and non-linear clipping processing to realize balanced regulation of local contrast difference, avoid artifacts caused by over-enhancement while maintaining clear edges, and finally output a structure contrast mapping image to enhance the level of detail and contrast stability of the tissue region.

[0015] Further, in S4, the structure contrast mapping image is generated, and the specific process includes: Taking the current position as the reference, a neighborhood window is selected as the statistical range, the neighborhood weighting coefficient corresponding to the spatial distance between each pixel position in the neighborhood and the target pixel position is determined, the pixel values of the structure-guided enhanced image in the neighborhood are weighted and accumulated according to the neighborhood weighting coefficient determined by the distance from the center pixel, and the weighted local mean is obtained by normalization through the sum of the neighborhood weighting coefficients. In the neighborhood range, the difference between the pixel value of the structure guided enhanced image and the weighted local mean value is calculated, and the obtained difference is squared and accumulated one by one, and then normalized by the total number of neighborhood pixels to obtain a local change factor; The local change factor at the target position is obtained, and the local variance value is subjected to a denominator addition operation with a preset variance balance parameter to obtain an inhibition factor. The absolute value of the difference between the pixel value of the structure guided enhanced image at the current position and the weighted local mean value is obtained, and subjected to a denominator addition operation with a contrast balance parameter to form an adjustment factor. The inhibition factor and the adjustment factor are subjected to a product operation to obtain a structure intensity weight; The difference between the structure guided enhanced image and the weighted local mean value at the target position is obtained, and the absolute value of the difference is taken. The Euclidean norm of the difference is calculated in the channel dimension, and a contrast scale parameter is introduced to normalize the Euclidean norm. The result of taking the absolute value of the difference is added to the normalized Euclidean norm to obtain a comprehensive difference amplitude. The structure intensity weight is used as a modulation factor to scale the comprehensive difference amplitude to obtain a difference regulation factor; The difference regulation factor is subjected to a convolution operation using a Gaussian kernel function to obtain a smooth difference field. A nonlinear enhancement parameter is used to apply a nonlinear activation function to the smooth difference field to limit the amplitude and control the voltage. Finally, the structure guided enhanced image is superimposed and fused in a residual connection manner to obtain a structure contrast mapping image.

[0016] Further, in the step S4, in the process of generating the structure-contrast mapping image, first, in the target position, the pixel values of the structure-guided enhanced image in the neighborhood are weighted and accumulated according to the neighborhood weighting coefficients determined by the distance of the center pixel, and the weighted local mean value is obtained by normalization processing, which is used to represent the local background intensity and as a reference for subsequent difference calculation; then, in the neighborhood range, the difference between each pixel value of the structure-guided enhanced image and the corresponding weighted local mean value is squared and accumulated, and the local change factor is obtained by normalization processing combined with the total number of neighborhood pixels, which is used to describe the amplitude of local texture change and reflect the complexity of the region; on this basis, the absolute value of the difference between the local change factor and the relative weighted local mean value of the pixel value of the structure-guided enhanced image at the current position is taken as two factors, which are respectively adjusted by the variance balance parameter and the contrast balance parameter, and the two are multiplied to obtain the structure intensity weight, which is used to suppress the enhancement effect in smooth areas and amplify the structure difference in detail areas, so as to realize targeted regulation; further, the difference between the pixel value of the structure-guided enhanced image at the current position and the weighted local mean value is calculated, and the absolute value is taken to obtain the initial difference amplitude, and the Euclidean norm of the difference is calculated in the channel dimension, and the contrast scale parameter is introduced to normalize the Euclidean norm, and then added to the initial difference amplitude, and finally the structure intensity weight is introduced as a modulation factor to scale, to obtain the difference regulation factor, which is used to highlight the significant difference and suppress the non-key response; then the difference regulation factor is convolved and smoothed using a Gaussian kernel to obtain a smoothed difference field, which is used to eliminate local noise and maintain overall structural continuity; on this basis, a nonlinear enhancement parameter is introduced to apply a nonlinear activation function to the smoothed difference field to limit the amplitude, so that the high amplitude response is controlled to avoid artifacts or overshoot; finally, the smoothed difference field after amplitude limiting is fused with the structure-guided enhanced image in a residual connection manner to obtain the structure-contrast mapping image, which can effectively improve the local contrast and strengthen the tissue boundary and texture details, providing an optimized input for subsequent resolution enhancement.

[0017] Preferably, in the step S5, in the image reconstruction process, the structure-contrast mapping image after feature extraction and enhancement processing is convolved to extract intermediate feature representation; then, according to the preset upsampling ratio, an upsampling structure composed of convolution and pixel rearrangement is constructed to realize the enhancement of image resolution through feature channel expansion and spatial reorganization; finally, the high-dimensional features after upsampling are mapped to the image pixel domain by convolution operation, and the super-resolution thoracic surgery pathology slice image with rich texture details and excellent structure preservation is output.

[0018] Preferably, in the S6 step, for the thoracic surgery pathological section image, a thoracic surgery pathological section image super-resolution reconstruction model is constructed, and discrete polykernels tensor transformation, local structure guided mapping, structure contrast mapping and image reconstruction operations are sequentially performed; first, the low-resolution thoracic surgery pathological section image is input into a discrete polykernels tensor transformation module to realize fusion of direction information and enhancement of texture details, and a direction aggregation enhancement image is obtained; then, the direction aggregation enhancement image is input into a local structure guided mapping module to differentially magnify significant detail regions and output a structure guided enhancement image; next, the structure guided enhancement image is input into a structure contrast mapping module to balance and regulate local contrast, and a structure contrast mapping image is obtained; finally, the structure contrast mapping image is input into an image reconstruction module composed of convolution and pixel rearrangement to complete spatial resolution enhancement and output a super-resolution thoracic surgery pathological section image; by end-to-end integration of the above stages and using an iterative back propagation training mechanism with the optimization objective of minimizing L1 loss, the clarity of tissue texture and the integrity of boundary structure are effectively enhanced, so that high-quality super-resolution reconstruction of fine-grained structures in the thoracic surgery pathological section image is realized.

[0019] In the above technical solution, the present application provides technical effects and advantages: The present application uses a discrete polykernels tensor transformation module to input a low-resolution thoracic surgery pathological section image as input, construct a local directionality point array tensor, introduce a response adjustment factor, generate a direction response polykernels tensor, combine a structure adaptive threshold and a nonlinear response reconstruction mechanism, and obtain a direction aggregation enhancement image, thereby realizing synchronous enhancement of fine-grained texture and boundaries in multiple directions. This mechanism can effectively alleviate the blurring and breaking phenomena in complex tissue boundary regions, maintain local structure continuity and overall contour integrity, and significantly improve the clarity and stability of the reconstruction result at the boundary details.

[0020] Based on the direction aggregation enhancement image, the present application introduces a local structure guided mapping mechanism, performs statistics based on robust local reference values and local texture energy in a neighborhood range, combines a variance balance parameter and a contrast balance parameter to generate a structure strength weight, differentially magnifies significant detail regions, and fuses with the direction aggregation enhancement image in a residual connection manner to obtain a structure guided enhancement image. This mechanism can realize balanced processing between smooth regions and fine structure regions, avoid the problem of insufficient detail depiction in low-contrast regions in traditional methods, thereby enhancing local texture details and improving the level of expression of tissue structure.

[0021] The application proposes a structure contrast mapping mechanism, which guides and enhances the image input processing, calculates the weighted mean and local change factor in the neighborhood range, combines the difference control factor and Gaussian convolution smoothing operation, and limits the amplitude through a nonlinear activation function, and outputs the structure contrast mapping image, which can effectively avoid the contrast imbalance and excessive enhancement of the tissue adhesion and overlapping area, suppress the generation of artifacts, ensure the natural transition of the boundary area structure, and significantly improve the stability of the reconstructed image in the contrast level and the authenticity of the boundary details. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 is a flowchart of the thoracic surgery pathological section image super-resolution reconstruction method provided by the application.

[0023] Figure 2 is a structural diagram of the discrete polykaryon tensor transformation module provided by the application.

[0024] Figure 3 is a structural diagram of the local structure guided mapping module provided by the application.

[0025] Figure 4 is a structural diagram of the structure contrast mapping module provided by the application.

[0026] Figure 5 is a comparison diagram of the thoracic surgery pathological section image after super-resolution reconstruction and the low-resolution thoracic surgery pathological section image provided by the application. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0028] Please refer to the accompanying Figure 1 to the accompanying Figure 5 , the application provides a thoracic surgery pathological section image super-resolution reconstruction method.

[0029] Please refer to Figure 1 , the thoracic surgery pathological section image super-resolution reconstruction method in the embodiments of the application has the following specific steps.

[0030] S1, collect high-resolution thoracic surgery pathological section images, generate corresponding low-resolution thoracic surgery pathological section images, and construct a thoracic surgery pathological section image data set.

[0031] Furthermore, in step S1, a thoracic surgery pathology slice image dataset was constructed, specifically as follows: first, the original data of thoracic surgery pathology slice images obtained by scanning at 40× optical magnification were obtained, and invalid image samples with abnormal staining distribution, focal plane offset, and blurred tissue structure were excluded; then, standardized block processing was performed on the original data, and the entire slice image was cropped into image blocks of 512×512 pixels. During the cropping process, a sliding window method with a stride of 128 pixels was used to enhance sample diversity, and high-resolution thoracic surgery pathology slice images were obtained and uniformly saved in .png format; then, bicubic An interpolation algorithm downsamples the high-resolution image by a factor of 2 to generate corresponding low-resolution thoracic pathology slide image samples, which are then paired one by one with the corresponding high-resolution image. Next, pixel normalization is performed on the low-resolution and high-resolution images, normalizing the RGB channel pixel values ​​to the range [0, 255] and implementing a unified color standardization operation to maintain the consistency of image staining features. Finally, the above image samples are manually divided into a training set containing 650 pairs of samples and a test set containing 150 pairs of samples to construct a dataset for training and evaluating the super-resolution model of thoracic pathology slide images.

[0032] Furthermore, in step S2, a discrete core tensor transformation module is constructed. The specific steps of the module construction are as follows: Figure 2 As shown, the specific implementation of the module includes the following steps.

[0033] S2. Calculate the local directional lattice tensor and response adjustment factor based on low-resolution thoracic surgical pathology slice images, construct a directional response aggregation kernel tensor, and combine it with the structure adaptive threshold and nonlinear response reconstruction mechanism to obtain a directional aggregation enhanced image.

[0034] Based on low-resolution thoracic surgical pathology slice images, multiple offset positions are determined with the current position as the center according to a preset angle set and radius set. The square of the Euclidean distance of the feature vector of the pixel at the current position and the pixel at the offset position in the channel dimension is calculated, and an exponential mapping is performed in combination with a similarity attenuation factor to construct a local directional lattice tensor. Furthermore, during the implementation process, based on low-resolution thoracic surgical pathology slice images, under a preset radius set and angle set, for any radius and angle, the corresponding horizontal and vertical offsets are first calculated using cosine and sine functions, and rounded to obtain the pixel value at the offset position; the square of the Euclidean distance of the eigenvector of the current position pixel and the offset position pixel in the channel dimension is calculated, and exponential mapping is performed in combination with the similarity attenuation factor to construct a local directional lattice tensor. The mathematical model of the local directional lattice tensor is: ; in, For location The local directional lattice tensor at , is the radius, is the angle, For low-resolution thoracic surgery pathology slice images The pixel value at For low-resolution thoracic surgery pathology slice images at offset positions The pixel value at is the longitudinal offset, is the lateral offset, is the similarity attenuation factor, Represents exponential operation with the natural constant e as the base; In this embodiment, the radius Used to control the neighborhood range, , is a radius set that controls the neighborhood search range. ,angle Used to set the sampling direction, , It is an angle set used to set the angle distribution of neighborhood sampling, ensuring the capture of anisotropic tissue structure features in multi-angle space and achieving comprehensive characterization of edges and contours. , and Determine the position of the neighboring pixels relative to the target pixel, which is used to locate the corresponding neighboring pixel position at a certain angle and radius. , ; Similarity decay factor Controls the effect of spatial differences on local responses. The larger the value, the more obvious the neighborhood differences are amplified, and it is easier to highlight the fiber boundaries and lesion areas in pathological tissue sections. In this embodiment, Limited to the interval [0.3,1.0], and set , through the above parameter design, construct the local directional lattice distribution function , which can describe the similarity between the current position pixel and the offset position pixel at a certain angle and radius, ensuring the accurate expression of the directional characteristics of image details and effectively enhancing the texture continuity of pathological slices.

[0035] In the angle set and radius set, the double cosine of the angle is used as the angle adjustment factor, and the radius and scale attenuation parameters are combined to perform exponential attenuation. At the same time, the directional gain coefficient is introduced for amplitude adjustment to construct the response adjustment factor. Furthermore, in the implementation process, in the angle set and radius set, the direction adjustment term is constructed with the double angle cosine of the angle, and the radius and scale attenuation parameters are combined. An exponential decay term is formed, the direction adjustment term is multiplied by the exponential decay term, and the multiplication result is multiplied by a direction gain coefficient A proportional amplification or reduction is performed to obtain a response adjustment factor, and a mathematical model of the response adjustment factor is: ; wherein, the response adjustment factor, the direction gain coefficient, the scale decay parameter; In the embodiment, the direction gain coefficient is used to adjust the contribution degree of different directions, so that the edge structure of an important direction is more prominent, and the direction gain coefficient is set to =1.0. The scale decay parameter is used to control the weight decay of a far neighbor pixel, so that the interference of noise is suppressed and the structure feature of a near neighbor is highlighted. In the embodiment, the scale decay parameter is limited in the interval [2, 4], and the scale decay parameter is set to =3. Through the above parameter design, the response adjustment factor is constructed, and the important angle and the near neighbor information can be highlighted, so that the local structure saliency is enhanced.

[0036] The local directional point array tensor and the response adjustment factor are multiplied item by item under all angle and radius combinations, and are accumulated at corresponding pixel positions to construct a directional response aggregation kernel tensor. Further, in the implementation process, the product of the local directional point array response and the response adjustment factor is calculated item by item at each pixel position with all radii and angles as indexes, and the product is summed in the radius-angle plane to obtain the directional response aggregation kernel tensor, and a mathematical model of the directional response aggregation kernel tensor is: ; wherein, the directional response aggregation kernel tensor, which comprehensively reflects the response characteristics under multiple directions and multiple scales, can represent the overall structure strength of the pixel position, and provides a unified basis for subsequent threshold discrimination and nonlinear mapping.

[0037] A difference value between the current pixel and a horizontally adjacent pixel of the current pixel is taken as a horizontal difference term, a difference value between the current pixel and a vertically adjacent pixel of the current pixel is taken as a vertical difference term, and an absolute value of a difference between the horizontal difference term and the vertical difference term is accumulated and normalized in the whole image range to obtain a structure adaptive threshold. Further, in the implementation process, the horizontal difference term and the vertical difference term are calculated for the low-resolution thoracic surgery pathology section image, respectively, and then the absolute value of a difference between the horizontal difference term and the vertical difference term is obtained, and the sum is obtained in the whole image range and is normalized to obtain the structure adaptive threshold, and a mathematical model of the structure adaptive threshold is: ; in, is the structure adaptive threshold, which is used to distinguish significant structure areas from homogeneous areas. and are the height and width of low-resolution thoracic surgery pathology slice images, =64, =64, is the horizontal difference term, is the difference term in the vertical direction; In this embodiment, the horizontal differential term and the vertical difference The calculation formula is: ; ; in, For low-resolution thoracic surgery pathology slice images The pixel value at For low-resolution thoracic surgery pathology slice images The pixel value at .

[0038] At each pixel position, when the directional response aggregation kernel tensor is greater than the structure adaptive threshold, a logarithmic enhancement term constructed by the directional response aggregation kernel tensor is superimposed on the eigenvalues ​​of the low-resolution thoracic surgical pathology slice image. When the directional response aggregation kernel tensor is less than or equal to the structure adaptive threshold, an exponential decay term constructed by the directional response aggregation kernel tensor is applied on the eigenvalues ​​of the low-resolution thoracic surgical pathology slice image to obtain a nonlinear response reconstruction tensor. Furthermore, during implementation, when the directional response aggregation kernel tensor is greater than the structural adaptive threshold, a logarithmic enhancement term is superimposed on the eigenvalues ​​of the low-resolution thoracic surgical pathology slice image to perform nonlinear enhancement on the significant structural region; when the directional response aggregation kernel tensor is less than or equal to the structural adaptive threshold, an exponential attenuation term is subtracted from the eigenvalues ​​of the low-resolution thoracic surgical pathology slice image to perform nonlinear suppression on the homogeneous region, thereby obtaining a segmented nonlinear response reconstruction tensor; ; in, Reconstructing tensors for nonlinear responses can enhance significant structural areas and suppress homogeneous areas, thereby improving the structural visibility of pathological images. It is a logarithmic enhancement item, which is used to perform nonlinear enhancement on the significant structure area. It is an exponential attenuation term, which is used to perform exponential attenuation suppression in homogeneous areas, smooth the background area, and avoid artifacts caused by over-enhancement.

[0039] Under all combinations of radius and angle, the values ​​of the nonlinear response reconstruction tensor at the corresponding offset position are weighted and accumulated according to the aggregation adjustment factor to obtain the weighted aggregation response value of each pixel; the sum of the aggregation adjustment factors is used as a normalization factor to normalize the weighted aggregation response value to obtain a directional aggregation enhanced image; Furthermore, during the implementation process, under all combinations of radius and angle, the pixel values ​​of the nonlinear response reconstruction tensor at the offset position are taken respectively, and accumulated according to the aggregation adjustment factor to obtain the weighted aggregation response value; the aggregation adjustment factor gradually decays with increasing radius; in order to avoid distortion of the result as the scale range expands, all aggregation adjustment factors are summed over the radius and direction angle set to obtain the normalization factor; finally, the weighted aggregation response value is divided by the normalization factor to obtain the directional aggregation and the normalized directional aggregation enhanced image. The specific mathematical model is: ; in, Enhance the image for directional aggregation, is the normalization factor, Reconstruct tensors at offset locations for nonlinear responses The pixel value at is the polymerization regulatory factor; In this embodiment, the polymerization regulator and normalization factor The calculation formula is: ; ; In this embodiment, the polymerization regulator It is used to ensure that the neighboring position contributes more to the final result, so as to achieve the priority enhancement of the neighbor response and the suppression of the distant neighbor response, while keeping the local details clear and suppressing irrelevant interference. The normalization factor Used to ensure the stability of the output value range, through the above design, the resulting directional aggregation enhanced image can achieve significant improvement in cell edges, glandular contours and other complex tissue structure areas, while avoiding artifacts and over-enhancement in homogeneous areas, thereby improving the readability of pathological sections.

[0040] Furthermore, in step S3, a local structure guided mapping module is constructed. The specific construction steps of the module are as follows: Figure 3 As shown, the specific implementation of the module includes the following steps.

[0041] S3. Based on the direction aggregation enhanced image, the robust local reference value is calculated in the neighborhood range, the local texture energy is constructed, the structure strength weight is calculated by combining the variance balance parameter and the contrast balance parameter, and the difference amplification term is generated. After nonlinear modulation, it is fused with the enhanced image in a residual manner to obtain a structure-guided enhanced image.

[0042] Based on the direction aggregation enhancement image, the eigenvalues ​​in the neighborhood with the target position as the reference are sorted, and the extreme value samples at the high and low ends are eliminated according to the preset truncation ratio. The remaining parts are then accumulated and averaged to obtain a robust local benchmark value. Furthermore, in the implementation process, the enhanced image is aggregated according to the direction to obtain the target position. For reference, the eigenvalues ​​in its neighborhood are sorted in ascending order; the number of extreme value samples at both ends to be removed is determined according to the truncation ratio, and after removing the samples at the beginning and end of the sequence, the average value of the remaining part is selected to obtain the robust local benchmark. The calculation formula of the robust local benchmark value is: ; in, is the robust local reference value, is the total number of neighborhood pixels, is the number of extreme value samples at both ends, The number of neighboring pixels remaining after removing the extreme samples at the high and low ends is The first intensity samples; In this embodiment, the total number of neighborhood pixels is the number of pixels in a 5×5 neighborhood, i.e. 25, and the number of extreme value samples at both ends , To cut off the ratio, this embodiment will Limited to the interval [0.1,0.2], and set ,Through the above parameter design, a robust local benchmark value is constructed, which can ,suppress the interference of dyeing particles, broken tissues and cross-edge ,outliers on the statistics.

[0043] With reference to the robust local reference value, the difference between the pixel value in the neighborhood of the directional aggregation enhanced image and the reference value is squared and accumulated, and then normalized by the total number of neighborhood pixels to obtain the local texture energy. Furthermore, in the implementation process, with reference to the robust local benchmark, the difference between the pixel value of each pixel in the neighborhood of the directional aggregation enhanced image and the robust local benchmark is squared and accumulated and normalized by the total number of neighborhood pixels to obtain the local texture energy. The calculation formula of the local texture energy is: ; in, For local texture energy, for describing local structure complexity, the value is larger in significant structure area, and the value is smaller in homogeneous area, and the important structure is preferentially responded, For 5x5 neighborhood range, For direction aggregation enhancement image in neighborhood position The pixel value is enhanced;

[0044] The local texture energy and the variance balance parameter are combined to form the suppression term, and the absolute value of the difference between the direction aggregation enhancement image and the robust local reference value is combined to form the adjustment term; the suppression term and the adjustment term are multiplied to obtain the structure intensity weight; Further, in the implementation process, the local texture energy and the absolute value of the difference between the direction aggregation enhancement image and the robust local reference are introduced into the weight design; first, the suppression term formed by the local texture energy and the variance balance parameter is constructed, so that the weight of the smooth area where the local texture energy tends to zero is effectively lowered, avoiding excessive enhancement of the homogeneous area; then, the adjustment term formed by the absolute value of the difference between the direction aggregation enhancement image and the robust local reference and the contrast balance parameter is constructed; the two terms are multiplied to obtain the structure intensity weight, and the mathematical model of the structure intensity weight is: ; Wherein, The structure intensity weight increases with the increase of edge saliency, The variance balance parameter is The contrast balance parameter is In this embodiment, the variance balance parameter is used to suppress the false enhancement of homogeneous area, and the embodiment limits to the interval [0.01, 0.05], and sets The contrast balance parameter is used to limit the excessive enhancement of high-contrast single point, and the embodiment limits to the interval [0.05, 0.10], and sets Through the above parameter design, the structure intensity weight can realize enhancement in edge and texture and other significant structure areas, and can keep suppression in homogeneous area, so as to ensure the sensitivity of identification.

[0045] The difference between the direction aggregation enhancement image and the robust local reference value is scaled by the structure intensity weight to obtain the difference amplification term; Further, in the implementation process, the difference between the direction aggregation enhancement image and the robust local reference value is introduced into the structure intensity weight to modulate, so as to generate the difference amplification term, and the mathematical model of the difference amplification term is: ; wherein, is a difference amplification term, used to realize selective enhancement of key structures and adaptive suppression of background.

[0046] On the basis of the difference amplification term, a nonlinear gain coefficient is introduced to modulate the amplitude thereof, and a nonlinear voltage control process is performed on the modulated difference amplification term in combination with an activation function; the difference amplification term subjected to the nonlinear voltage control is fused with the directionally aggregated enhanced image in a residual connection manner to obtain a structure-guided enhanced image. Further, in the implementation process, a nonlinear gain coefficient is introduced to adjust the amplitude of the difference amplification term, which is realized by point-by-point weighting of the difference amplification term; subsequently, the adjusted difference amplification term is input to a tanh activation function to perform nonlinear voltage control on the value of each pixel position, so that it gradually tends to saturation in a high-amplitude region and remains linear in a low-amplitude region, thereby realizing amplitude limiting constraint as a whole; then, the difference amplification term subjected to the nonlinear voltage control is pixel-by-pixel superimposed on the directionally aggregated enhanced image in a residual connection manner to obtain a structure-guided enhanced image, and the mathematical model of the structure-guided enhanced image is: ; wherein, is the structure-guided enhanced image, is a tanh activation function used to limit high-amplitude input and reduce the risk of overshoot and artifacts, is a nonlinear gain coefficient used to adjust the amplitude of nonlinear response and ensure that the enhancement degree is adapted to different slice samples; In this embodiment, the nonlinear gain coefficient is calculated according to the following formula: ; wherein, denotes the 75th percentile of the absolute error between the directionally aggregated enhanced image and the robust local reference value, is a small constant to prevent division by zero and ensure numerical stability of the calculation, and is set to 0.000001.

[0047] Further, in the S4 step, a structure-contrast mapping module is constructed, and the specific construction steps of the module are shown in FIG. 4, and the specific implementation of the module includes the following steps. Figure 4

[0048] S4, weighted mean and variance calculation is performed on the structure-guided enhanced image within a neighborhood range, a variance balance parameter and a contrast balance parameter are introduced to calculate a structure intensity weight, a difference control factor is constructed, and after convolution smoothing and nonlinear amplitude limiting processing, the difference control factor is fused with the structure-guided enhanced image to obtain a structure-contrast mapping image.

[0049] ​With the current position as a reference, a neighborhood window is selected as a statistical range, a corresponding neighborhood weighting coefficient is determined according to the spatial distance between each pixel position in the neighborhood and the target pixel position, the pixel values of the structure-guided enhanced image in the neighborhood are weighted and accumulated according to the neighborhood weighting coefficient determined according to the distance from the center pixel, and the weighted local mean is obtained by normalizing the sum of the neighborhood weighting coefficients. Further, in the implementation process, first, a neighborhood window is set at each pixel position as a statistical range, and a corresponding neighborhood weighting coefficient is calculated according to the distance between each pixel in the neighborhood and the target pixel, for determining the contribution of the pixels in the neighborhood to the target pixel; the pixel values of the structure-guided enhanced image in the neighborhood are multiplied with the corresponding neighborhood weighting coefficients one by one, and all the products are accumulated; the accumulated result is normalized according to the sum of the neighborhood weighting coefficients to obtain a weighted local mean, and the mathematical model of the weighted local mean is: ; wherein, is the weighted local mean, is the neighborhood weighting coefficient, for controlling the contribution of the neighborhood pixels to the mean, is the pixel value of the structure-guided enhanced image at the neighborhood position . In the embodiment, the calculation formula of the neighborhood weighting coefficient is: ; wherein, is the Euclidean distance between the neighborhood position and the center position, for describing the spatial position relationship between pixels, is a spatial scale parameter; In the embodiment, the calculation formula of the Euclidean distance between the neighborhood position and the center position is: ; In the embodiment, the neighborhood weighting coefficient is used to control the contribution of the neighborhood pixels to the mean, the distance sensitivity is realized by constructing an exponential decay function, the spatial scale parameter is introduced to control the decay speed of the weight with the distance between the neighborhood pixels and the center position, in order to ensure that the neighborhood weighting coefficient effectively reflects the spatial position relationship in the neighborhood, the embodiment limits to [1.0, 1.5], and sets , through the above parameter design, the weighted local mean obtained can effectively reflect the spatial correlation of the pixels in the local range.

[0050] In the neighborhood range, the difference between the pixel value of the structure-guided enhanced image and the weighted local mean value is calculated, the obtained difference is squared one by one and then accumulated, and the local change factor is obtained by normalizing the total number of neighborhood pixels; Further, in the implementation process, in the neighborhood range, the difference between the pixel value of the structure-guided enhanced image and the corresponding weighted local mean value is squared, the result is accumulated, and then normalized by combining the total number of neighborhood pixels to obtain the local change factor, and the calculation formula of the local change factor is: ; Among them, is the local change factor, which indicates the change amplitude of the local region.

[0051] The local change factor at the target position is obtained, and the local variance value is added to the denominator with a preset variance balance parameter to form an inhibition factor; the absolute value of the difference between the pixel value of the structure-guided enhanced image at the current position and the weighted local mean value is obtained, and the contrast balance parameter is added to the denominator to form an adjustment factor; the inhibition factor and the adjustment factor are multiplied to obtain the structure intensity weight; Further, in the implementation process, the denominator part of the local change factor at the target position introduces a preset variance balance parameter, so that the local variance value is added to the variance balance parameter to obtain an inhibition factor for balancing the local difference amplitude; the pixel value of the structure-guided enhanced image at the current position is extracted, and the difference between the weighted local mean value at the corresponding position is calculated, the absolute value of the obtained difference is taken, and then the contrast balance parameter is added to the denominator to obtain an adjustment factor for adjusting the local difference sensitivity; the inhibition factor and the adjustment factor are combined by product operation to obtain the structure intensity weight at the target position, and the mathematical model of the structure intensity weight is: ; Among them, is the structure intensity weight, is the variance balance parameter, is the contrast balance parameter; In this embodiment, the variance balance parameter is used to avoid over-enhancement in the smooth area, and in this embodiment, the is limited to [0.01, 0.05], and the is set to 0.03. The contrast balance parameter is used to suppress the over-contrast of single points, and in this embodiment, the is limited to [0.05, 0.1], and the is set to 0.075. Through the above parameter settings, the structure intensity weight constructed can have a higher value at the nucleus and gland boundary and a lower value in the homogeneous region.

[0052] obtaining a difference between the structure-guided enhanced image and the weighted local mean at the target position, and taking an absolute value of the difference; calculating a Euclidean norm of the difference in the channel dimension, and introducing a contrast scale parameter to normalize the Euclidean norm; adding the result of taking the absolute value of the difference to the normalized Euclidean norm to obtain a comprehensive difference amplitude; and using the structure strength weight as a modulation factor to scale the comprehensive difference amplitude to obtain a difference regulation factor; Further, in the implementation process, at the target position, first, a difference between the pixel value of the structure-guided enhanced image and the corresponding weighted local mean is obtained, and an absolute value of the difference is taken to obtain an initial amplitude; a Euclidean norm operation is performed on the difference in the channel dimension to comprehensively describe the overall amplitude of the channel difference, and a contrast scale parameter is introduced to normalize the Euclidean norm, limit the numerical range, and realize scale consistency; the initial amplitude is added to the normalized Euclidean norm to obtain a comprehensive difference amplitude for more complete description of the local difference feature; finally, the comprehensive difference amplitude is multiplied by the structure strength weight to scale the comprehensive difference amplitude to generate a difference regulation factor, and the mathematical model of the difference regulation factor is: ; wherein, is the difference regulation factor, is the contrast scale parameter; In this embodiment, the contrast scale parameter is used to normalize the channel difference to ensure the stability of the enhancement, and the is limited to the interval [0.1, 0.3], and is set to 0.2. Through the above parameter setting, the difference regulation factor constructed can be effectively amplified in the significant structure area, and the suppression effect is maintained in the smooth area.

[0053] The difference regulation factor is convolved using a Gaussian kernel function to obtain a smooth difference field; a nonlinear activation function is applied to the smooth difference field to limit the amplitude control combined with a nonlinear enhancement parameter; and finally, the structure-guided enhanced image is superimposed and fused in a residual connection manner to obtain a structure contrast mapping image; Further, in the implementation process, the difference regulation factor is convolved using a Gaussian kernel to realize the smoothing of the difference in the local neighborhood to obtain a smooth difference field; then, the tanh activation function is applied to the smooth difference field to limit the amplitude control using a nonlinear enhancement parameter, so that the numerical stability is maintained in the high amplitude area, thereby avoiding the artifacts caused by excessive enhancement; finally, the amplitude-limited smooth difference field is superimposed to the structure-guided enhanced image through the residual connection to generate a structure contrast mapping image, and the mathematical model of the structure contrast mapping image is: ; wherein, is a structure contrast mapping image, is a nonlinear enhancement parameter, is a Gaussian kernel with is a standard deviation; In this embodiment, the nonlinear enhancement parameter controls the clipping intensity, and this embodiment limits to the interval [0.6, 1.0] and sets , the standard deviation is used to adjust the range and intensity of convolution smoothing, and this embodiment limits to [0.8, 1.2] and sets Through the above parameter settings, the structure contrast mapping image obtained can significantly improve the structure edge details while maintaining smooth consistency in the homogeneous background.

[0054] S5, input the structure contrast mapping image into an image reconstruction module to perform resolution enhancement.

[0055] Further, in the specific implementation process, the structure contrast mapping image is input into an image reconstruction module to perform resolution enhancement processing; the image reconstruction module specifically includes: first, performing convolution operation on the structure contrast mapping image to realize expansion of the channel dimension, and then completing the up-sampling of the spatial dimension through the pixel rearrangement operation, so as to realize 2 times super-resolution reconstruction.

[0056] S6, construct a thoracic surgery pathological section image super-resolution reconstruction model, and train the model to convergence with a high-resolution image as a supervision signal to realize super-resolution reconstruction of the thoracic surgery pathological section image.

[0057] Further, the thoracic surgery pathological section image super-resolution reconstruction model is constructed, and discrete polykernels tensor transformation, local structure guided mapping, structure contrast mapping and image reconstruction operations are sequentially performed. Firstly, the low-resolution thoracic surgery pathological section image is input into the discrete polykernels tensor transformation module, the directional response polykernels tensor is generated by constructing the local directional point lattice tensor and the response adjustment factor, the multi-directional feature fusion and detail enhancement of the low-resolution thoracic surgery pathological section image are carried out, the texture and edge structure are highlighted and the noise interference is suppressed, and a directional polykernels enhanced image is obtained. Subsequently, the directional polykernels enhanced image is input into the local structure guided mapping module, the robust local reference value and the local texture energy are calculated in the neighborhood range of the target position, the structure intensity weight is calculated by combining the variance balance parameter and the contrast balance parameter, the difference between the target pixel and the reference is amplified, and the amplified difference is superimposed on the directional polykernels enhanced image in the form of residual connection, so that the fine texture and low contrast area are finely described, and a structure guided enhanced image is output. Then, the structure guided enhanced image is input into the structure contrast mapping module, the weighted mean and variance are calculated in the neighborhood range, the structure intensity weight of the local difference perception is constructed, the difference between the target pixel and the neighborhood is scaled and adjusted, a difference regulation factor is obtained, and the difference regulation factor is inhibited and constrained by combining the tanh activation function, so that the balance and stability of the contrast are ensured, and a structure contrast mapping image is output. Finally, the structure contrast mapping image is input into the image reconstruction module, the 2-fold resolution is improved by convolution and pixel rearrangement, and a texture clear and structure continuous super-resolution thoracic surgery pathological section image is obtained. Through end-to-end integration of the above stages and using the iterative back propagation training mechanism with the optimization target of minimizing the L1 loss, the high-precision restoration and clear reconstruction of the fine-grained structure in the thoracic surgery pathological section image are realized.

[0058] Further, the thoracic surgery pathological section image super-resolution reconstruction model is constructed, and discrete polykernels tensor transformation, local structure guided mapping, structure contrast mapping and image reconstruction operations are sequentially performed. Firstly, the low-resolution thoracic surgery pathological section image is input into the discrete polykernels tensor transformation module, the directional response polykernels tensor is generated by constructing the local directional point lattice tensor and the response adjustment factor, the multi-directional feature fusion and detail enhancement of the low-resolution thoracic surgery pathological section image are carried out, the texture and edge structure are highlighted and the noise interference is suppressed, and a directional polykernels enhanced image is obtained. Subsequently, the directional polykernels enhanced image is input into the local structure guided mapping module, the robust local reference value and the local texture energy are calculated in the neighborhood range of the target position, the structure intensity weight is calculated by combining the variance balance parameter and the contrast balance parameter, the difference between the target pixel and the reference is amplified, and the amplified difference is superimposed on the directional polykernels enhanced image in the form of residual connection, so that the fine texture and low contrast area are finely described, and a structure guided enhanced image is output. Then, the structure guided enhanced image is input into the structure contrast mapping module, the weighted mean and variance are calculated in the neighborhood range, the structure intensity weight of the local difference perception is constructed, the difference between the target pixel and the neighborhood is scaled and adjusted, a difference regulation factor is obtained, and the difference regulation factor is inhibited and constrained by combining the tanh activation function, so that the balance and stability of the contrast are ensured, and a structure contrast mapping image is output. Finally, the structure contrast mapping image is input into the image reconstruction module, the 2-fold resolution is improved by convolution and pixel rearrangement, and a texture clear and structure continuous super-resolution thoracic surgery pathological section image is obtained. Through end-to-end integration of the above stages and using the iterative back propagation training mechanism with the optimization target of minimizing the L1 loss, the high-precision restoration and clear reconstruction of the fine-grained structure in the thoracic surgery pathological section image are realized.

[0059] Further, the thoracic surgery pathological section image dataset is input into the constructed super-resolution reconstruction model for processing, and experimental results are as shown in Figure 5 Figure 5 The super-resolution results of the thoracic surgery pathological section image reconstructed by the model proposed in the application are shown; it can be observed that the model can effectively maintain the detailed features of the tissue texture and structure boundary in the pathological section while improving the image resolution, avoiding the common blurring and structure distortion problems of the traditional interpolation method; compared with the original low-resolution image, the reconstruction result is significantly improved in terms of texture clarity, boundary sharpness and structure continuity, especially in the lesion area, fine blood vessels and cell boundary, which verifies the accuracy and robustness of the application in the thoracic surgery pathological section image super-resolution reconstruction task.

[0060] The above is only the preferred embodiment of the application, and it should be pointed out that for those skilled in the art, without departing from the concept of the application, several modifications and improvements can be made, which are all within the protection scope of the application.​

Claims

1. A thoracic surgery pathology section image super-resolution reconstruction method, characterized by: collecting high-resolution thoracic surgery pathology section images, generating corresponding low-resolution thoracic surgery pathology section images, and constructing a thoracic surgery pathology section image dataset; calculating local directional point array tensors and response adjustment factors from low-resolution thoracic surgery pathology section images, constructing directional response aggregation kernel tensors, and processing them in combination with a structure adaptive threshold and a nonlinear response reconstruction mechanism to obtain directional aggregation enhancement images; calculating robust local reference values within a neighborhood range based on the directional aggregation enhancement images, constructing local texture energy, calculating structure intensity weights in combination with variance balance parameters and contrast balance parameters, generating difference amplification terms, and fusing them with the enhancement images in a residual manner to obtain structure-guided enhancement images; performing weighted mean and variance calculations on the structure-guided enhancement images within a neighborhood range, introducing variance balance parameters and contrast balance parameters to calculate structure intensity weights, constructing difference regulation factors, performing convolution smoothing and nonlinear limiting processing, and fusing them with the structure-guided enhancement images to obtain structure contrast mapping images; inputting the structure contrast mapping images into an image reconstruction module for resolution enhancement; constructing a thoracic surgery pathology section image super-resolution reconstruction model, training the model to convergence using high-resolution images as supervision signals, and realizing super-resolution reconstruction of thoracic surgery pathology section images.

2. The thoracic surgery pathological section image super-resolution reconstruction method according to claim 1, characterized in that, obtain high-quality thoracic surgery pathology section image raw data, and remove image samples with abnormal staining distribution, focal plane deviation, or blurred tissue structure; perform standardization patching processing on the thoracic surgery pathology section image raw data, crop the entire section image into image blocks with fixed resolution, and generate high-resolution thoracic surgery pathology section images as label images; perform downsampling operation on the high-resolution thoracic surgery pathology section images using Bicubic interpolation algorithm to obtain low-resolution thoracic surgery pathology section image samples, and pair the low-resolution images with corresponding high-resolution label images; perform pixel normalization processing on the low-resolution images and high-resolution images, and divide them into training set and test set to construct a thoracic surgery pathology section image dataset.

3. The thoracic surgery pathological section image super-resolution reconstruction method according to claim 2, characterized in that, According to the low-resolution thoracic surgery pathology section images, determine a plurality of offset positions with the current position as the center, the preset angle set and the radius set, calculate the Euclidean distance square of the feature vectors of the current position pixels and the offset position pixels in the channel dimension, and construct the local directional point array tensor in combination with the similarity decay factor for exponential mapping; In the angle set and the radius set, take the double-angle cosine of the angle as the angle adjustment factor, combine the radius and the scale decay parameter for exponential decay, and introduce the direction gain coefficient for amplitude adjustment to construct the response adjustment factor; multiply the local directional point array tensor and the response adjustment factor item by item under all angle and radius combinations, and accumulate at the corresponding pixel position to construct the directional response aggregation kernel tensor.

4. The thoracic surgery pathological section image super-resolution reconstruction method according to claim 3, characterized in that, The difference between the current pixel and its horizontally adjacent pixel is taken as a horizontal difference item, and the difference between the current pixel and its vertically adjacent pixel is taken as a vertical difference item, the absolute value of the difference between the two is accumulated and normalized in the whole image range to obtain a structure adaptive threshold value; At each pixel position, when the direction response aggregation kernel tensor is greater than the structure adaptive threshold value, a logarithmic enhancement term constructed by the direction response aggregation kernel tensor is superimposed on the basis of the eigenvalue of the low-resolution thoracic surgical pathology section image; When the direction response aggregation kernel tensor is less than or equal to the structure adaptive threshold value, an exponential decay term constructed by the direction response aggregation kernel tensor is applied to the basis of the eigenvalue of the low-resolution thoracic surgical pathology section image to obtain a nonlinear response reconstruction tensor; Under all combinations of radius and angle, the values of the nonlinear response reconstruction tensor at the corresponding offset positions are weighted and accumulated according to the aggregation adjustment factor to obtain the weighted aggregation response value of each pixel; the sum of the aggregation adjustment factor is taken as a normalization factor to normalize the weighted aggregation response value to obtain a direction aggregation enhancement image.

5. The thoracic surgery pathological section image super-resolution reconstruction method according to claim 4, characterized in that, According to the direction aggregation enhancement image, the eigenvalues in the neighborhood range with the target position as the reference are sorted, the extreme value samples at both ends are removed according to a preset truncation ratio, and then the remaining part is accumulated and averaged to obtain a robust local reference value; Under the reference of the robust local reference value, the difference between each pixel value in the neighborhood of the direction aggregation enhancement image and the reference value is squared and accumulated, and then normalized according to the total number of neighborhood pixels to obtain a local texture energy.

6. The thoracic surgery pathological section image super-resolution reconstruction method according to claim 5, characterized in that, The local texture energy and the variance balance parameter are combined to form a suppression term, and the absolute value of the difference between the direction aggregation enhancement image and its robust local reference value is taken as a regulation term together with the contrast balance parameter; the suppression term and the regulation term are multiplied to obtain a structure intensity weight; The difference between the direction aggregation enhancement image and the robust local reference value is scaled by the structure intensity weight to obtain a difference amplification term; On the basis of the difference amplification term, a nonlinear gain coefficient is introduced to modulate its amplitude, and a nonlinear voltage control processing is performed on the modulated difference amplification term combined with an activation function; The direction aggregation enhancement image and the difference amplification term subjected to nonlinear amplitude limiting are fused through a residual connection to obtain a structure-guided enhancement image.

7. The thoracic surgery pathological section image super-resolution reconstruction method according to claim 6, characterized in that, Taking the current position as the reference, a neighborhood window is selected as the statistical range, and the spatial distance between each pixel position in the neighborhood and the target pixel position is used to determine the corresponding neighborhood weighting coefficient; the pixel values of the structure-guided enhancement image in the neighborhood are weighted and accumulated according to the neighborhood weighting coefficient determined by the distance from the center pixel, and the sum of the neighborhood weighting coefficients is normalized to obtain a weighted local mean value; In the neighborhood range, the difference between each pixel value of the structure-guided enhancement image and the weighted local mean value is calculated, and the obtained differences are squared one by one and then accumulated, and the total number of neighborhood pixels is normalized to obtain a local change factor.

8. The thoracic surgery pathological section image super-resolution reconstruction method according to claim 7, characterized in that, A local variation factor at the target position is obtained, and a local variance value is subjected to a denominator addition operation with a preset variance balance parameter to obtain an inhibition factor; a difference absolute value between a pixel value of the structure-guided enhancement image at the current position and a weighted local mean value is obtained, and subjected to a denominator addition operation with a contrast balance parameter to form an adjustment factor; the inhibition factor and the adjustment factor are subjected to a product operation to obtain a structure intensity weight; A difference between the structure-guided enhancement image at the target position and the weighted local mean value is obtained, and an absolute value of the difference is taken; a Euclidean norm of the difference is calculated in a channel dimension, and a contrast scale parameter is introduced to normalize the Euclidean norm; a result obtained by taking an absolute value of the difference is added to the normalized Euclidean norm to obtain a comprehensive difference amplitude; the structure intensity weight is used as a modulation factor to scale the comprehensive difference amplitude to obtain a difference regulation factor; A convolution operation is performed on the difference regulation factor using a Gaussian kernel function to obtain a smooth difference field; a nonlinear activation function is applied to the smooth difference field to limit the amplitude control in combination with a nonlinear enhancement parameter; Finally, the structure contrast mapping image is obtained by superimposing and fusing the structure-guided enhancement image in a residual connection manner.

Images