A thoracic surgery pathological section image super-resolution reconstruction method

Abstract

The present application provides a thoracic surgery pathological section image super-resolution reconstruction method, which relates to the technical field of image processing, and the specific steps include: collecting high-resolution thoracic surgery pathological section images and generating low-resolution images; constructing local directionality dot matrix tensor and response adjustment factor based on the low-resolution images, calculating angle response aggregation kernel tensor, and combining nonlinear response reconstruction function for segmented mapping, completing directionality fusion normalization, and obtaining direction aggregation enhanced image; calculating robust local reference value and local texture energy within the neighborhood range, combining variance balance parameter and contrast balance parameter to generate structure strength weight, constructing difference amplification term and fusing with the direction aggregation enhanced image through residual method, and obtaining structure guided enhanced image; calculating weighted mean and variance, generating difference regulation factor, combining convolution smoothing and nonlinear limiting processing, and obtaining structure contrast mapping image; finally, inputting the image reconstruction module to realize resolution improvement.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more particularly to a method for super-resolution reconstruction of thoracic surgical pathology slide images. Background Technology

[0002] Currently, pathological slide images are crucial for the diagnosis and treatment evaluation of thoracic surgical diseases. Their resolution and clarity directly affect the accuracy of identifying cell morphology, glandular contours, and nucleoplasmic features. With the development of digital pathology, whole-field-of-slice (WSI) has become the core form of digitized pathology, typically achieved through full-frame scanning of tissue sections using 20× or 40× objectives. However, 40× WSI data is massive, with extremely high storage and transmission costs, limiting its application in large-scale clinical and research scenarios. Therefore, 20× WSI is often used as input, and super-resolution methods are employed to restore image quality close to that of 40× WSI. However, directly magnifying 20× WSI presents significant problems: blurred boundaries between cell nuclei and cytoplasm lead to unclear morphological features, affecting tumor grading and microinvasive identification; glandular contours and fibrous structures show breaks or artifacts after magnification, failing to accurately reflect tissue continuity; tissue texture exhibits both noise amplification and excessive smoothing at high magnification, making homogeneous areas unnatural and causing loss of detail in key areas, thus affecting the reliability and stability of thoracic surgical pathological slides in refined diagnosis and computer-aided analysis.

[0003] Publication No. CN118967449A proposes a super-resolution method for pathological slice images based on a diffusion model. It obtains multi-level features by combining global attention and local convolution through a hybrid feature extraction module, enhances image features through iterative noise addition and denoising using a diffusion transformer network, and finally generates high-resolution slice images through a feature fusion module. Publication No. CN120450963A proposes a super-resolution method for medical images based on dynamic attention and implicit neural representation. It 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, existing technologies often fail to adequately alleviate the problems of insufficient texture gradient and blurred edges in low-resolution images during super-resolution reconstruction of thoracic surgical pathology slides. They lack effective compensation mechanisms for weakened information and ignore the contrast imbalance and over-enhancement problems that easily occur in areas of tissue adhesion and overlap. This results in incomplete detail recovery, over-enhancement of homogeneous areas, or insufficient smoothing in super-resolution reconstruction tasks. Summary of the Invention

[0005] This invention provides a method for super-resolution reconstruction of thoracic surgical pathology slide images, aiming to construct a super-resolution reconstruction framework for thoracic surgical pathology slide images with the ability to enhance texture details and maintain structural continuity. The method acquires low-resolution thoracic surgical pathology slide images, and combines discrete kernel tensor transformation, local structure-guided mapping, and structure contrast mapping modules to gradually enhance the edge contours and subtle features of low-contrast areas of the image. Finally, the enhancement results are input into the image reconstruction module to improve the image resolution and complete high-quality super-resolution reconstruction of thoracic surgical pathology slide images.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for super-resolution reconstruction of thoracic surgical pathological slide images, the specific steps of which are as follows:

[0007] High-resolution thoracic surgical pathology slide images are acquired, corresponding low-resolution thoracic surgical pathology slide images are generated, and a thoracic surgical pathology slide image dataset is constructed.

[0008] Based on low-resolution thoracic surgical pathology slide images, the local directional lattice tensor and response regulation factor are calculated, a directional response aggregation kernel tensor is constructed, and then processed by a structure adaptive threshold and nonlinear response reconstruction mechanism to obtain a directional aggregation enhanced image.

[0009] Based on the direction aggregation enhancement image, robust local reference values ​​are calculated in the neighborhood range to construct local texture energy. The structure strength weight is calculated by combining variance balance parameters and contrast balance parameters to generate difference amplification terms. After nonlinear modulation, the amplification terms are fused with the enhancement image in a residual manner to obtain a structure-guided enhancement image.

[0010] Within the neighborhood, the weighted mean and variance of the structure-guided enhancement image are calculated. The variance balance parameter and contrast balance parameter are introduced to calculate the structure strength weight. A difference regulation factor is constructed. After convolution smoothing and nonlinear limiting processing, it is fused with the structure-guided enhancement image to obtain a structure contrast mapping image.

[0011] The structural contrast mapping image is input into the image reconstruction module for resolution enhancement.

[0012] A super-resolution reconstruction model for thoracic surgical pathology slide images was constructed. The model was trained until convergence using high-resolution images as supervision signals to achieve super-resolution reconstruction of thoracic surgical pathology slide images.

[0013] Preferably, in step S1, the construction of the thoracic surgical pathology slide image dataset includes: firstly, using a 40× objective lens to perform a full-frame scan of the tissue slides to ensure the integrity and detail fidelity of the samples. During this process, slide samples with abnormal staining distribution, focal plane shift, or blurred tissue structure are removed, thereby obtaining high-quality thoracic surgical pathology slide images; subsequently, the thoracic surgical pathology slide images are subjected to standardized block processing, cropping the entire slide into fixed-size image blocks to generate high-resolution thoracic surgical pathology slide images, which are then used as label images; next, a bicubic interpolation algorithm is used to downsample the label image blocks to obtain corresponding low-resolution thoracic surgical pathology slide images; further, the low-resolution images and high-resolution images are uniformly normalized to maintain the consistency and stability of data distribution; finally, the low-resolution images are paired with their corresponding high-resolution label images to form a thoracic surgical pathology slide image dataset for training and testing the super-resolution model, providing a unified and standardized data input for subsequent super-resolution reconstruction of thoracic surgical pathology slide images.

[0014] Preferably, low-resolution thoracic surgical pathology slide images generally suffer from indistinct texture gradient changes and blurred transitions between cell and tissue boundaries and the background. Traditional methods are prone to response attenuation during multi-directional feature extraction, resulting in the loss of detailed textures. Based on these characteristics, this invention proposes a discrete clustering tensor transformation module. By constructing local directional lattice tensors and response adjustment factors within the angle and radius sets, and combining structural adaptive thresholding and nonlinear response reconstruction mechanisms, it achieves the aggregation and differential enhancement of multi-directional features. Directional fusion and normalization calculations are then performed to generate directional aggregation enhanced images, effectively alleviating the problems of insufficient texture gradients and blurred edges in low-resolution images, and improving the clarity of tissue textures and edge structures.

[0015] Furthermore, in step S2, a directional aggregation enhancement image is generated, the specific process of which includes:

[0016] Based on low-resolution thoracic surgical pathology slide images, multiple offset positions are determined according to a preset set of angles and radii, with the current position as the center. The squared Euclidean distance between the feature vectors of the current position pixel and the offset position pixel in the channel dimension is calculated, and an exponential mapping is performed in combination with a similarity decay factor to construct a local directional lattice tensor.

[0017] In the set of angles and the set of radii, the double cosine of the angle is used as the angle adjustment factor, and exponential attenuation is performed by combining the radius and scale attenuation parameters. At the same time, the directional gain coefficient is introduced to adjust the amplitude, thus constructing the response adjustment factor.

[0018] Under all angle and radius combinations, the local directional lattice tensor is multiplied by the response adjustment factor term by term and accumulated at the corresponding pixel position to construct the directional response aggregation kernel tensor;

[0019] The difference between the current pixel and its horizontally adjacent pixels is used as the horizontal difference term, and the difference between the current pixel and its vertically adjacent pixels is used as the vertical difference term. The absolute values ​​of the differences are accumulated and normalized across the entire image to obtain the structure adaptive threshold.

[0020] At each pixel location, when the directional response aggregation kernel tensor is greater than the structure adaptive threshold, a logarithmic enhancement term constructed from the directional response aggregation kernel tensor is superimposed on the feature values ​​of the low-resolution thoracic surgical pathology slide image; when the directional response aggregation kernel tensor is less than or equal to the structure adaptive threshold, an exponential decay term constructed from the directional response aggregation kernel tensor is applied to the feature values ​​of the low-resolution thoracic surgical pathology slide image to obtain a nonlinear response reconstruction tensor.

[0021] Under all combinations of radii and angles, 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 aggregated response value of each pixel; the sum of the aggregation adjustment factors is used as a normalization factor to normalize the weighted aggregated response value to obtain the directional aggregated enhancement image.

[0022] Further, in step S2, during the generation of the directional aggregation enhanced image, firstly, within the set of angles and radii, the Euclidean distance between the feature vectors of the center pixel and each offset position in the channel dimension is calculated based on the low-resolution thoracic surgical pathology slide image. This distance is then combined with a similarity attenuation factor for exponential mapping to construct a local directional lattice tensor. This tensor accurately characterizes the local differences between pixels under different directions and distances, reflecting the directional sensitivity of the local neighborhood and providing a basic representation for subsequent directional responses. Subsequently, a response adjustment factor is formed by using the double cosine of the angle as the angle adjustment factor, combined with radius and scale attenuation parameters, and introducing a directional gain coefficient. This factor is used to adaptively weight local differences under different directions and scales, highlighting the contribution of the main direction to texture and edges while suppressing interference from non-main directions. Under all angle and radius combinations, the two factors are multiplied sequentially and accumulated at corresponding positions to obtain the directional response aggregation kernel tensor. This tensor is used to fuse local response results under multiple directions and scales to form a comprehensive directional sensitive response map. This provides global guidance for enhancing salient regions and texture structures. Then, a structure-adaptive threshold is constructed using horizontal and vertical differences. This threshold distinguishes between salient and non-salient responses based on the anisotropic distribution of the image structure, ensuring response enhancement in edge and textured regions while avoiding over-enhancement or false textures in smooth regions. At each pixel, when the aggregation kernel value is higher than the threshold, a logarithmic enhancement term constructed from the directional response is superimposed on the original feature; otherwise, exponential suppression is applied, generating a nonlinear response reconstruction tensor. This tensor effectively suppresses noise and reduces overshoot artifacts while highlighting salient edges and texture details, ensuring the stability and reliability of the reconstruction results. Finally, the nonlinear response reconstruction tensor is weighted and summed at all angle and radius offsets, and the weighted sum is normalized to obtain a directional aggregation-enhanced image. This enhanced image is clearer at the texture level, more continuous at the boundary level, and has strong noise robustness, serving as high-quality input data for subsequent recognition and analysis modules, improving the overall system reliability and accuracy.

[0023] Preferably, low-resolution thoracic surgical pathology slide images are prone to local detail loss in cellular structures and low-contrast areas. Traditional methods struggle to achieve a balance between smooth backgrounds and fine structures, resulting in insufficient tissue texture expression. Based on this characteristic, this invention proposes a local structure-guided mapping module. By statistically analyzing robust local benchmark values ​​and local texture energy within the neighborhood, and combining variance and contrast balance parameters to calculate structural strength weights, the difference between target pixels and the benchmark is differentially adjusted. Finally, residual-based and directional aggregation are used to enhance image fusion, resulting in a structure-guided enhanced image. This effectively alleviates the problem of detail weakening and improves the integrity and coherence of the overall tissue contour.

[0024] Furthermore, in step S3, a structure-guided enhancement image is generated, the specific process of which includes:

[0025] Based on the direction aggregation enhancement image, within the neighborhood range with the target location as a reference, the feature values ​​in the neighborhood are sorted, extreme value samples at both ends are removed according to a preset truncation ratio, and the remaining parts are accumulated and averaged to obtain the robust local benchmark value.

[0026] With reference to a robust local benchmark, the differences between each pixel value in the neighborhood of the direction-aggregated enhanced image and the benchmark are squared and accumulated, and then normalized according to the total number of neighborhood pixels to obtain the local texture energy.

[0027] The local texture energy and variance balance parameter are combined to form a suppression term, and then the absolute value of the difference between the enhanced image and its robust local reference value is aggregated by direction, and together with the contrast balance parameter, a regulation term is formed; the suppression term and the regulation term are multiplied to obtain the structural strength weight.

[0028] Scaling is performed on the difference between the directional aggregation enhanced image and the robust local benchmark value by applying a structural strength weight, resulting in a difference amplification term.

[0029] Based on the differential amplification term, a nonlinear gain coefficient is introduced to modulate its amplitude, and an activation function is used to perform nonlinear voltage control processing on the modulated differential amplification term. The nonlinearly limited differential amplification term is then fused with the directional aggregation enhancement image through residual connection to obtain the structure-guided enhancement image.

[0030] Further, in step S3, during the generation of the structure-guided enhancement image, firstly, within a neighborhood referenced to the target location, the pixel feature values ​​are sorted, and extreme value samples at both ends are removed according to a preset truncation ratio. Then, the remaining values ​​are accumulated and averaged to obtain a robust local benchmark value. This benchmark value provides a stable local reference, avoiding interference from abnormal pixels in subsequent calculations. Subsequently, the differences between each pixel value in the neighborhood and the robust local benchmark value are squared and accumulated, and normalized by combining this with the total number of neighborhood pixels to obtain local texture energy. This energy is used to measure the texture complexity of the local region and provides a basis for weight design. Next, the suppression term, composed of the local texture energy and variance balance parameters, is combined with the directional aggregation enhancement image. The structural strength weight is obtained by combining the absolute value of the difference between the robust local reference value and the contrast balance parameter, and multiplying the two terms. This weight is used to dynamically adjust the amplification of local differences to ensure the balance of detail enhancement. On this basis, the difference between the directional aggregation enhancement image and the robust local reference value is introduced into the structural strength weight for scaling to obtain the difference amplification term, which is used to highlight significant local structural differences. Finally, a nonlinear gain coefficient is introduced into the difference amplification term for modulation, and nonlinear voltage control is completed through the tanh activation function. The result is superimposed on the directional aggregation enhancement image in a residual connection manner to obtain the structure-guided enhancement image. This image can enhance details and edge information while maintaining overall stability, providing enhanced input features for subsequent processing.

[0031] Preferably, low-resolution thoracic surgical pathology slide images often exhibit contrast imbalance and over-enhancement risks in areas of tissue adhesion and overlap. Traditional methods are prone to generating artifacts during processing, affecting image realism. Based on these characteristics, this invention proposes a structural contrast mapping module. This module calculates the weighted mean and variance of the structure-guided enhanced image within a neighborhood range to generate a difference control factor. Combined with Gaussian convolution smoothing and nonlinear limiting processing, it achieves balanced control of local contrast differences, avoiding artifacts caused by over-enhancement while maintaining edge clarity. The final output is a structural contrast mapping image, which enhances the sense of layering and contrast stability of the tissue region.

[0032] Furthermore, in step S4, a structural contrast mapping image is generated, the specific process of which includes:

[0033] Using the current position as a reference, a neighborhood window is selected as the statistical range. The corresponding neighborhood weighting coefficient is determined based on the spatial distance between each pixel position in the neighborhood and the target pixel position. 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. The sum of the neighborhood weighting coefficients is then normalized to obtain the weighted local mean.

[0034] Within the neighborhood, the differences between each pixel value of the structure-guided enhancement image and the weighted local mean are calculated, and the differences are squared one by one and accumulated. Then, the results are normalized by combining the total number of neighborhood pixels to obtain the local change factor.

[0035] The local variation factor at the target location is obtained, and the local variance value is summed with the preset variance balance parameter to form the suppression factor; the absolute value of the difference between the pixel value at the current position of the structure-guided enhancement image and the weighted local mean is obtained, and the difference is summed with the contrast balance parameter to form the adjustment factor; the suppression factor and the adjustment factor are multiplied to obtain the structure strength weight.

[0036] The difference between the structure-guided enhancement image and the weighted local mean at the target location is obtained, and the absolute value of the difference is taken. The Euclidean norm of the difference is calculated in the channel dimension, and the contrast scale parameter is introduced to normalize the Euclidean norm. The result after taking the absolute value of the difference is added to the normalized Euclidean norm to obtain the comprehensive difference amplitude. The comprehensive difference amplitude is scaled using the structure strength weight as a modulation factor to obtain the difference regulation factor.

[0037] A smooth difference field is obtained by convolving the difference control factor with a Gaussian kernel function; a nonlinear activation function is applied to the smooth difference field for amplitude limiting and compression control in combination with nonlinear enhancement parameters; finally, the image is superimposed and fused with the structure-guided enhancement image in the form of residual connections to obtain a structure contrast mapping image.

[0038] Further, in step S4, during the generation of the structure contrast mapping image, firstly, at the target location, using a neighborhood window as the statistical range, the pixel values ​​of the structure-guided enhancement image within the neighborhood are weighted and accumulated according to a neighborhood weighting coefficient determined by the distance from the center pixel. This is then normalized to obtain a weighted local mean, which characterizes the local background intensity and serves as a reference for subsequent difference calculations. Subsequently, within the neighborhood, the differences between each pixel value of the structure-guided enhancement image and its corresponding weighted local mean are squared and accumulated, then normalized in conjunction with the total number of neighborhood pixels to obtain a local variation factor. This variance characterizes the magnitude of local texture changes and reflects the complexity of the region. Based on this, the absolute value of the difference between the local variation factor and the pixel value of the structure-guided enhancement image at the current location relative to the weighted local mean is used as two factors, adjusted respectively in conjunction with variance balance parameters and contrast balance parameters. Multiplying these two factors yields a structure intensity weight, which is used to suppress the enhancement effect in smooth regions and amplify structural differences in detailed regions, thereby achieving targeted enhancement. Further, the difference between the pixel value of the structure-guided enhancement image at the current location and the weighted local mean is calculated, and the absolute value is taken to obtain the initial difference amplitude. Simultaneously, the Euclidean norm of this difference is calculated along the channel dimension, and a contrast scaling parameter is introduced to normalize the Euclidean norm. This normalized value is then added to the initial difference amplitude. Finally, a structural strength weight is introduced as a modulation factor for scaling to obtain a difference modulation factor. This enhancement term is used to highlight significant differences and suppress non-critical responses. Subsequently, a Gaussian kernel is used to convolve and smooth the difference modulation factor to obtain a smoothed difference field. This smoothed field is used to eliminate local noise and maintain the overall structural continuity. Based on this, a nonlinear enhancement parameter is introduced to apply a nonlinear activation function to the smoothed difference field for amplitude limiting, so that high-amplitude responses are controlled to avoid artifacts or overshoot. Finally, the amplitude-limited smoothed difference field is fused with the structure-guided enhancement image using residual connections to obtain a structure contrast mapping image, which can effectively improve local contrast and enhance tissue boundaries and texture details, providing optimized input for subsequent resolution enhancement.

[0039] Preferably, in step S5, during the image reconstruction process, a convolution operation is performed on the structural contrast mapping image after feature extraction and enhancement to extract intermediate feature representations; then, based on a preset upsampling ratio, an upsampling structure composed of convolution and pixel rearrangement is constructed, and the image resolution is improved through feature channel expansion and spatial recombination; finally, the upsampled high-dimensional features are mapped to the image pixel domain using convolution operations to output a super-resolution thoracic surgical pathology slide image with rich texture details and excellent structure preservation.

[0040] Preferably, in step S6, a super-resolution reconstruction model for thoracic surgical pathology slide images is constructed, and discrete clustering tensor transform, local structure-guided mapping, structural contrast mapping, and image reconstruction operations are performed sequentially. First, the low-resolution thoracic surgical pathology slide image is input to the discrete clustering tensor transform module to achieve fusion of directional information and enhancement of texture details, resulting in a directional clustering enhanced image. Subsequently, the directional clustering enhanced image is input to the local structure-guided mapping module to differentially amplify significant detail regions and output a structure-guided enhanced image. Next, the structure-guided enhanced image is input to the structure contrast mapping module to balance local contrast and obtain a structure contrast mapping image. Finally, the structure contrast mapping image is input to the image reconstruction module composed of convolution and pixel rearrangement to complete spatial resolution enhancement and output a super-resolution thoracic surgical pathology slide image. By integrating the above stages end-to-end and using an iterative backpropagation training mechanism with minimizing L1 loss as the optimization objective, the clarity of tissue texture and the integrity of boundary structures are effectively enhanced, thereby achieving high-quality super-resolution reconstruction of fine-grained structures in thoracic surgical pathology slide images.

[0041] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0042] This invention utilizes a discrete-kernel tensor transformation module to take low-resolution thoracic surgical pathological slice images as input, constructs a local directional lattice tensor, introduces a response adjustment factor, and generates a directional response aggregation kernel tensor. Combined with a structure-adaptive threshold and a nonlinear response reconstruction mechanism, a directional aggregation enhanced image is obtained, thereby achieving simultaneous enhancement of fine-grained texture and boundaries in multiple directions. This mechanism can effectively alleviate the blurring and breakage phenomena in complex tissue boundary regions, maintain the continuity of local structures and the integrity of the overall contour, and significantly improve the clarity and stability of the reconstruction results at boundary details.

[0043] This invention introduces a local structure-guided mapping mechanism based on directional aggregation enhancement images. It performs statistical analysis on robust local benchmark values ​​and local texture energy within the neighborhood, and generates structural strength weights by combining variance balance parameters and contrast balance parameters. This amplifies the differences in significant detail regions and fuses them with the directional aggregation enhancement image using a residual connection method to obtain a structure-guided enhancement image. This mechanism can achieve balanced processing between smooth regions and fine structure regions, avoiding the problem of insufficient detail depiction in low-contrast regions in traditional methods, thereby enhancing local texture details and improving the hierarchical expressiveness of tissue structures.

[0044] This invention proposes a structural contrast mapping mechanism, which involves inputting a structure-guided enhanced image, calculating a weighted mean and a local change factor within the neighborhood, combining a difference control factor with Gaussian convolution smoothing, and limiting the amplitude through a nonlinear activation function to output a structural contrast mapping image. This mechanism can effectively avoid contrast imbalance and over-enhancement in tissue adhesion and overlapping areas, suppress artifact generation, and ensure the natural transition of boundary area structures, thereby significantly improving the stability of the reconstructed image in terms of contrast and the realism of boundary details. Attached Figure Description

[0045] Figure 1 This is a flowchart of the super-resolution reconstruction method for thoracic surgical pathology slide images provided by the present invention.

[0046] Figure 2 This is a structural diagram of the discrete nuclear tensor transformation module provided by the present invention.

[0047] Figure 3 This is a structural diagram of the local structure guidance mapping module provided by the present invention.

[0048] Figure 4 This is a structural diagram of the structural comparison mapping module provided by the present invention.

[0049] Figure 5 This is a comparison image of a thoracic surgical pathology slide image reconstructed by super-resolution and a low-resolution thoracic surgical pathology slide image provided by the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Please see the appendix Figure 1 To be continued Figure 5 This invention provides a method for super-resolution reconstruction of thoracic surgical pathological slide images.

[0052] Please see Figure 1 As shown in the embodiment of this application, a method for super-resolution reconstruction of thoracic surgical pathology slide images is described, and the specific steps are as follows.

[0053] S1. Acquire high-resolution thoracic surgical pathology slide images, generate corresponding low-resolution thoracic surgical pathology slide images, and construct a thoracic surgical pathology slide image dataset.

[0054] Further, in step S1, a dataset of thoracic surgical pathology slide images is constructed, specifically as follows: First, the raw data of thoracic surgical pathology slide images obtained by scanning at 40× optical magnification are acquired, and invalid image samples with abnormal staining distribution, focal plane shift, and blurred tissue structures are excluded; then, the raw data is subjected to standardized block processing, cropping the entire slide image into 512×512 pixel image blocks. During the cropping process, a sliding window method with a stride of 128 pixels is used to enhance sample diversity, obtaining high-resolution thoracic surgical pathology slide images, which are uniformly saved as .png format; then, a bicubic method is used... The interpolation algorithm downsamples the high-resolution image by a factor of 2 to generate corresponding low-resolution thoracic surgical pathology slide image samples, and pairs the low-resolution images with the corresponding high-resolution images one by one. Next, pixel normalization is performed on the low-resolution and high-resolution images, normalizing the RGB three-channel pixel values ​​to the [0,255] range, and uniform color standardization is performed to maintain the consistency of image staining features. Finally, the above image samples are manually divided, with the training set containing 650 pairs of samples and the test set containing 150 pairs of samples, to construct a dataset for training and evaluating the super-resolution model of thoracic surgical pathology slide images.

[0055] Furthermore, in step S2, a discrete kernel tensor transformation module is constructed. The specific construction steps of the module are shown in the appendix. Figure 2 As shown, the specific implementation of the module includes the following steps.

[0056] S2. Calculate the local directional lattice tensor and response regulation factor based on the low-resolution thoracic surgical pathological section image, construct the directional response aggregation kernel tensor, and process it by combining the structure adaptive threshold and nonlinear response reconstruction mechanism to obtain the directional aggregation enhanced image.

[0057] Based on low-resolution thoracic surgical pathology slide images, multiple offset positions are determined according to a preset set of angles and radii, with the current position as the center. The squared Euclidean distance between the feature vectors of the current position pixel and the offset position pixel in the channel dimension is calculated, and an exponential mapping is performed in combination with a similarity decay factor to construct a local directional lattice tensor.

[0058] Furthermore, during implementation, based on low-resolution thoracic surgical pathology slide images, under a preset set of radii and angles, for any radius and angle, the corresponding horizontal and vertical offsets are first calculated using cosine and sine functions, and then rounded to obtain the pixel value at the offset position. The squared Euclidean distance between the feature vectors of the current pixel and the offset pixel in the channel dimension is calculated, and an exponential mapping is performed using a similarity decay factor to construct a local directional lattice tensor. The mathematical model of the local directional lattice tensor is as follows:

[0059] ;

[0060] in, For position The local directional lattice tensor at that point For radius, For angle, For low-resolution thoracic surgical pathology slide images in Pixel value at that location, Low-resolution thoracic surgical pathology slide images at offset positions Pixel value at that location, This is the vertical offset. This is the lateral offset. As a similarity decay factor, It represents exponential operations with the natural constant e as the base;

[0061] In this embodiment, radius Used to control the neighborhood range , The radius set controls the neighborhood search range, and is set as follows: ,angle Used to set the sampling direction , This is an angle set used to define the angle distribution of neighborhood sampling, ensuring the capture of anisotropic tissue structure features in multi-angle space, achieving comprehensive representation of edges and contours. , and Determine the position of neighboring pixels relative to the target pixel to locate the position of the corresponding neighboring pixels at a certain angle and radius. , ;

[0062] Similarity decay factor The influence of spatial differences on local response is controlled; the larger the value, the more pronounced the amplification of neighborhood differences, making it easier to highlight fibrous boundaries and lesion areas in pathological tissue sections. This embodiment will... Limited to the range [0.3, 1.0], and set... By designing the parameters described above, a local directional lattice distribution function is constructed. It can describe the similarity between the current pixel and the offset pixel at a certain angle and radius, ensuring the accurate expression of the directional features of image details and effectively enhancing the texture continuity of pathological sections.

[0063] In the set of angles and the set of radii, the double cosine of the angle is used as the angle adjustment factor, and exponential attenuation is performed by combining the radius and scale attenuation parameters. At the same time, the directional gain coefficient is introduced to adjust the amplitude, thus constructing the response adjustment factor.

[0064] Furthermore, during implementation, a direction adjustment term is constructed from the double-angle cosine of the angles within the angle and radius sets, and this is combined with the radius and scale attenuation parameters. To form an exponential decay term, multiply the directional adjustment term by the exponential decay term, and then apply the result through the directional gain coefficient. By scaling up or down, the response adjustment factor is obtained. The mathematical model of the response adjustment factor is as follows:

[0065] ;

[0066] in, In response to the adjustment factor, This is the directional gain coefficient. For scale attenuation parameters;

[0067] In this embodiment, the directional gain coefficient This is used to adjust the contribution in different directions, ensuring that the edge structure in important directions is more prominent, and setting... =1.0, scale decay parameter This is used to control the weight decay of distant neighbor pixels, suppress noise interference, and highlight the structural features of nearby neighbors. In this embodiment, Limited to the interval [2,4], and set =3, and through the above parameter design, a response adjustment factor is constructed. This enables the highlighting of important angles and neighboring information, thereby enhancing the salience of local structures.

[0068] Under all angle and radius combinations, the local directional lattice tensor is multiplied by the response adjustment factor term by term and accumulated at the corresponding pixel position to construct the directional response aggregation kernel tensor;

[0069] Furthermore, during implementation, at each pixel location, using all radii and angles as indices, the product of the local directional lattice response and the response adjustment factor is calculated item by item, and these products are summed on the radius-angle plane to obtain the directional response aggregation kernel tensor. The mathematical model of the directional response aggregation kernel tensor is as follows:

[0070] ;

[0071] in, The directional response aggregation kernel tensor integrates response features across multiple directions and scales, enabling it to characterize the overall structural strength of pixel locations and providing a unified basis for subsequent thresholding and nonlinear mapping.

[0072] The difference between the current pixel and its horizontally adjacent pixels is used as the horizontal difference term, and the difference between the current pixel and its vertically adjacent pixels is used as the vertical difference term. The absolute values ​​of the differences are accumulated and normalized across the entire image to obtain the structure adaptive threshold.

[0073] Furthermore, during implementation, the difference terms in the horizontal and vertical directions are calculated for low-resolution thoracic surgical pathology slide images, and the absolute value of the difference is taken, summed and normalized over the entire image range to obtain the structural adaptive threshold. The mathematical model of the structural adaptive threshold is as follows:

[0074] ;

[0075] in, This is a structure-adaptive threshold used to distinguish between regions with significant structural features and homogeneous regions. and These represent the height and width of a low-resolution thoracic surgical pathology slide image, respectively. =64, =64, The difference term is in the horizontal direction. These are the difference terms in the vertical direction;

[0076] In this embodiment, the difference term in the horizontal direction Difference terms in the vertical direction The calculation formula is:

[0077] ;

[0078] ;

[0079] in, For low-resolution thoracic surgical pathology slide images in Pixel value at that location, For low-resolution thoracic surgical pathology slide images in The pixel value at that location.

[0080] At each pixel location, when the directional response aggregation kernel tensor is greater than the structure adaptive threshold, a logarithmic enhancement term constructed from the directional response aggregation kernel tensor is superimposed on the feature values ​​of the low-resolution thoracic surgical pathology slide image; when the directional response aggregation kernel tensor is less than or equal to the structure adaptive threshold, an exponential decay term constructed from the directional response aggregation kernel tensor is applied to the feature values ​​of the low-resolution thoracic surgical pathology slide image to obtain a nonlinear response reconstruction tensor.

[0081] Furthermore, during implementation, when the directional response aggregation kernel tensor is greater than the structure adaptive threshold, a logarithmic enhancement term is superimposed on the feature values ​​of the low-resolution thoracic surgical pathology slide image to perform nonlinear enhancement on the significant structural regions; when the directional response aggregation kernel tensor is less than or equal to the structure adaptive threshold, an exponential decay term is subtracted from the feature values ​​of the low-resolution thoracic surgical pathology slide image to perform nonlinear suppression on the homogeneous regions, thereby obtaining a piecewise nonlinear response reconstruction tensor.

[0082] ;

[0083] in, By reconstructing the tensor for nonlinear responses, salient structural regions can be enhanced while homogeneous regions are suppressed, thereby improving the structural visibility of pathological images. This is a logarithmic enhancement term used for nonlinear enhancement of salient structural regions. It is an exponential decay term used to suppress exponential decay in homogeneous regions, smoothing the background region and avoiding artifacts caused by over-enhancement.

[0084] Under all combinations of radii and angles, 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 aggregated response value of each pixel; the sum of the aggregation adjustment factors is used as a normalization factor to normalize the weighted aggregated response value to obtain the directional aggregated enhancement image.

[0085] Furthermore, during implementation, for all combinations of radii and angles, the pixel values ​​of the nonlinear response reconstruction tensor at the offset positions are taken and accumulated according to the aggregation adjustment factor to obtain a weighted aggregation response value; the aggregation adjustment factor gradually decreases with increasing radius; to avoid distortion of the results as the scale range expands, all aggregation adjustment factors are summed over the set of radii and orientation angles to obtain a normalization factor; finally, the weighted aggregation response value is divided by the normalization factor to obtain the directional aggregation and normalized directional aggregation enhanced image. The specific mathematical model is as follows:

[0086] ;

[0087] in, To enhance the image by directional aggregation, As the normalization factor, Reconstruct the tensor at the offset position for the nonlinear response Pixel value at that location, It is a polymerization regulator;

[0088] In this embodiment, the aggregation regulation factor and normalization factor The calculation formula is:

[0089] ;

[0090] ;

[0091] In this embodiment, the aggregation regulation factor This is used to ensure that nearest neighbors contribute more to the final result, thereby prioritizing the enhancement of nearest neighbor responses and suppressing distant neighbor responses, preserving local details while suppressing irrelevant interference, and a normalization factor. To ensure a stable output value range, the above design results in directional aggregation enhancement images that significantly improve cell edges, gland contours, and other complex tissue structures. At the same time, artifacts and over-enhancement are avoided in homogeneous areas, thereby improving the interpretability of pathological sections.

[0092] Furthermore, in step S3, a local structure boot mapping module is constructed. The specific construction steps of the module are attached. Figure 3 As shown, the specific implementation of the module includes the following steps.

[0093] S3. Aggregate and enhance the image according to the direction, calculate the robust local reference value in the neighborhood, construct the local texture energy, calculate the structure strength weight by combining the variance balance parameter and the contrast balance parameter, generate the difference amplification term, and fuse it with the enhanced image in a residual manner through nonlinear modulation to obtain the structure-guided enhanced image.

[0094] Based on the direction aggregation enhancement image, within the neighborhood range with the target location as a reference, the feature values ​​in the neighborhood are sorted, extreme value samples at both ends are removed according to a preset truncation ratio, and the remaining parts are accumulated and averaged to obtain the robust local benchmark value.

[0095] Furthermore, during implementation, the enhanced images are aggregated according to direction to target location. For reference, the feature values ​​within its neighborhood are sorted in ascending order; the number of extreme value samples to be removed at both ends is determined according to the truncation ratio; after removing the first and last samples of the sequence, the average value of the remaining part is selected to obtain the robust local benchmark. The formula for calculating the robust local benchmark value is as follows:

[0096] ;

[0097] in, For robust local reference values, The total number of neighboring pixels. The number of extreme value samples at both ends. To calculate the number of neighboring pixels remaining after removing extreme samples at both the high and low ends, The first one after sorting the neighborhood One intensity sample;

[0098] In this embodiment, the total number of neighboring pixels The number of pixels in a 5×5 neighborhood, i.e., 25, and the number of extreme samples at both ends. , To achieve the desired truncation ratio, this embodiment will... Limited to the range [0.1, 0.2], and set... By designing the parameters described above, a robust local baseline value is constructed, which can suppress the interference of staining particles, fragmented tissue, and outliers across the margin on the statistics.

[0099] With reference to a robust local benchmark, the differences between each pixel value in the neighborhood of the direction-aggregated enhanced image and the benchmark are squared and accumulated, and then normalized according to the total number of neighborhood pixels to obtain the local texture energy.

[0100] Furthermore, during implementation, with reference to a robust local benchmark, the differences between the pixel values ​​in the neighborhood of the direction-aggregated enhanced image and the robust local benchmark are accumulated by square and normalized according to the total number of neighborhood pixels to obtain the local texture energy. The formula for calculating the local texture energy is as follows:

[0101] ;

[0102] in, This represents local texture energy, used to describe the complexity of local structures. It has a larger value in regions with significant structures and a smaller value in homogeneous regions, thus prioritizing responses to important structures. The neighborhood is 5×5. To enhance the image in the neighborhood by directional aggregation Pixel value at;

[0103] The local texture energy and variance balance parameter are combined to form a suppression term, and then the absolute value of the difference between the enhanced image and its robust local reference value is aggregated by direction, and together with the contrast balance parameter, a regulation term is formed; the suppression term and the regulation term are multiplied to obtain the structural strength weight.

[0104] Furthermore, in the implementation process, the absolute value of the difference between the local texture energy and the directional aggregation enhanced image relative to the robust local benchmark is jointly introduced into the weight design. First, a suppression term is constructed based on the local texture energy and the variance balance parameter, so that the weight is effectively reduced in smooth regions where the local texture energy approaches zero, avoiding over-enhancement of homogeneous regions. Then, an adjustment term is constructed based on the absolute value of the difference between the directional aggregation enhanced image and the robust local benchmark and the contrast balance parameter. The product of the two terms yields the structural strength weight, the mathematical model of which is:

[0105] ;

[0106] in, The structural strength weight increases with increasing edge saliency. For variance balance parameters, This refers to the contrast balance parameter;

[0107] In this embodiment, the variance balance parameter This embodiment will be used to suppress false enhancement in homogeneous regions. Limited to the range [0.01, 0.05], and set... Contrast balance parameters To limit excessive enhancement at high-contrast single points, this embodiment will Limited to the range [0.05, 0.10], and set... By designing the above parameters, structural strength weights are constructed. It can enhance the recognition in areas with significant structural features such as edges and textures, while maintaining suppression in homogeneous areas, thus ensuring the sensitivity of recognition.

[0108] Scaling is performed on the difference between the directional aggregation enhanced image and the robust local benchmark value by applying a structural strength weight, resulting in a difference amplification term.

[0109] Furthermore, during implementation, the difference between the directional aggregation enhanced image and the robust local reference value is introduced into the structural strength weight for modulation, thereby generating a difference amplification term. The mathematical model of the difference amplification term is as follows:

[0110] ;

[0111] in, This is a difference amplification term, used to achieve selective enhancement of key structures and adaptive suppression of the background.

[0112] Based on the differential amplification term, a nonlinear gain coefficient is introduced to modulate its amplitude, and an activation function is used to perform nonlinear voltage control processing on the modulated differential amplification term; the nonlinear voltage-controlled differential amplification term is fused with the orientation aggregation enhancement image through residual connection to obtain the structure-guided enhancement image;

[0113] Furthermore, during implementation, a nonlinear gain coefficient is introduced to adjust the amplitude of the differential amplification term, achieved by weighting the differential amplification term point by point. Subsequently, the adjusted differential amplification term is input to the tanh activation function, and nonlinear voltage control is applied to the value at each pixel position, causing it to gradually saturate in the high-amplitude region while maintaining a linear response in the low-amplitude region, thus achieving overall amplitude limiting. Next, through residual connections, the nonlinearly voltage-controlled differential amplification term is superimposed pixel by pixel onto the directional aggregation enhancement image to obtain the structure-guided enhancement image. The mathematical model of the structure-guided enhancement image is as follows:

[0114] ;

[0115] in, Enhance images to guide structure. The tanh activation function is used to limit high-amplitude inputs and reduce the risk of overshoot and artifacts. This is the nonlinear gain coefficient, used to adjust the amplitude of the nonlinear response to ensure that the enhancement level adapts to different slice samples;

[0116] In this embodiment, the nonlinear gain coefficient The calculation formula is:

[0117] ;

[0118] in, This represents the 75th quantile of the absolute error between the directional aggregation-enhanced image and the robust local reference value. To prevent division by zero by small constants and to ensure numerical stability of the calculation, the value is set to 0.000001.

[0119] Furthermore, in step S4, a structure comparison mapping module is constructed. The specific construction steps of the module are shown in the appendix. Figure 4 As shown, the specific implementation of the module includes the following steps.

[0120] S4. Within the neighborhood, the weighted mean and variance of the structure-guided enhancement image are calculated. The variance balance parameter and contrast balance parameter are introduced to calculate the structure strength weight. A difference control factor is constructed. After convolution smoothing and nonlinear limiting processing, it is fused with the structure-guided enhancement image to obtain a structure contrast mapping image.

[0121] Using the current position as a reference, a neighborhood window is selected as the statistical range. The corresponding neighborhood weighting coefficient is determined based on the spatial distance between each pixel position in the neighborhood and the target pixel position. 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. The sum of the neighborhood weighting coefficients is then normalized to obtain the weighted local mean.

[0122] Furthermore, in the implementation process, a neighborhood window is first set as the statistical range at each pixel location, and the corresponding neighborhood weighting coefficient is calculated based on the distance between each pixel in the neighborhood and the target pixel to determine the contribution of the pixels in the neighborhood to the target pixel. The pixel values ​​of the structure-guided enhancement image in the neighborhood are multiplied one by one by the corresponding neighborhood weighting coefficients, and all products are accumulated. The accumulated result is normalized according to the sum of the neighborhood weighting coefficients to obtain the weighted local mean. The mathematical model of the weighted local mean is as follows:

[0123] ;

[0124] in, For weighted local mean, These are neighborhood weighting coefficients, used to control the contribution of neighboring pixels to the mean. Enhance the image in the neighborhood location for structure guidance Pixel value at;

[0125] In this embodiment, the neighborhood weighting coefficient The calculation formula is:

[0126] ;

[0127] in, The Euclidean distance between neighboring locations and the center location describes the spatial relationship between pixels. Spatial scale parameter;

[0128] In this embodiment, the Euclidean distance between the neighborhood location and the center location is... The calculation formula is:

[0129] ;

[0130] In this embodiment, the neighborhood weighting coefficient This is used to control the contribution of neighboring pixels to the mean. Distance sensitivity is modulated by constructing an exponential decay function and introducing a spatial scale parameter. The decay rate of the weights is controlled as the distance between neighboring pixels and the center position changes. To ensure that the neighborhood weighting coefficients effectively reflect the spatial positional relationships within the neighborhood, this embodiment will... Limited to [1.0, 1.5], and set Through the above parameter design, the weighted local mean is obtained. It can effectively reflect the spatial correlation of pixels within a local range.

[0131] Within the neighborhood, the differences between each pixel value of the structure-guided enhancement image and the weighted local mean are calculated, and the differences are squared one by one and accumulated. Then, the results are normalized by combining the total number of neighborhood pixels to obtain the local change factor.

[0132] Furthermore, during implementation, within the neighborhood, the difference between each pixel value of the structure-guided enhancement image and its corresponding weighted local mean is squared, and the results are accumulated. Then, the results are normalized by combining this with the total number of neighborhood pixels to obtain a local change factor. The formula for calculating the local change factor is as follows:

[0133] ;

[0134] in, This is a local variation factor used to indicate the magnitude of change in a local area.

[0135] The local variation factor at the target location is obtained, and the local variance value is summed with the preset variance balance parameter to form the suppression factor; the absolute value of the difference between the pixel value at the current position of the structure-guided enhancement image and the weighted local mean is obtained, and the difference is summed with the contrast balance parameter to form the adjustment factor; the suppression factor and the adjustment factor are multiplied to obtain the structure strength weight.

[0136] Furthermore, during implementation, a preset variance balance parameter is introduced into the denominator of the local change factor at the target location. The local variance value is summed with the variance balance parameter to obtain a suppression factor used to balance the amplitude of local differences. The pixel value of the structure-guided enhancement image at the current location is extracted and its difference is calculated with the weighted local mean at the corresponding location. The absolute value of the difference is taken, and then a contrast balance parameter is introduced into the denominator for summing to obtain an adjustment factor used to regulate the sensitivity of local differences. The suppression factor and the adjustment factor are combined through a product operation to obtain the structural strength weight at the target location. The mathematical model of the structural strength weight is as follows:

[0137] ;

[0138] in, For structural strength weight, For variance balance parameters, This refers to the contrast balance parameter;

[0139] In this embodiment, the variance balance parameter To avoid excessive enhancement of the smooth region, this embodiment will Limited to [0.01, 0.05], and set Contrast balance parameters To suppress excessive contrast at a single point, this embodiment will Limited to [0.05, 0.1], and set By setting the above parameters, the constructed structural strength weight can take higher values ​​at the cell nucleus and gland boundary, and lower values ​​in homogeneous regions.

[0140] The difference between the structure-guided enhancement image and the weighted local mean at the target location is obtained, and the absolute value of the difference is taken. The Euclidean norm of the difference is calculated in the channel dimension, and the contrast scale parameter is introduced to normalize the Euclidean norm. The result after taking the absolute value of the difference is added to the normalized Euclidean norm to obtain the comprehensive difference amplitude. The comprehensive difference amplitude is scaled using the structure strength weight as a modulation factor to obtain the difference regulation factor.

[0141] Furthermore, during implementation, at the target location, the difference between the pixel value of the structure-guided enhancement image and the corresponding weighted local mean is first obtained, and the absolute value of the difference is taken to obtain the initial amplitude. Euclidean norm operation is performed on this difference along the channel dimension to synthesize the overall amplitude of the differences across channels. A contrast scaling parameter is introduced to normalize the Euclidean norm, limiting its numerical range and achieving scale consistency. The initial amplitude is added to the normalized Euclidean norm to obtain the comprehensive difference amplitude, which is used to more completely characterize the local difference features. Finally, the comprehensive difference amplitude is multiplied by the structural strength weight to scale the comprehensive difference amplitude, generating a difference control factor. The mathematical model of the difference control factor is as follows:

[0142] ;

[0143] in, As a differential regulatory factor, This is the contrast scale parameter;

[0144] In this embodiment, the contrast scale parameter This embodiment uses channel differences to normalize and enhance stability. Limited to the range [0.1, 0.3], and set... By setting the parameters as described above, the constructed differential regulatory factor can be effectively amplified in salient structural regions while maintaining an inhibitory effect in smooth regions.

[0145] A smooth difference field is obtained by convolving the difference control factor with a Gaussian kernel function; a nonlinear activation function is applied to the smooth difference field for amplitude limiting and compression control in combination with nonlinear enhancement parameters; finally, the image is superimposed and fused with the structure-guided enhancement image in the form of residual connections to obtain a structure contrast mapping image.

[0146] Furthermore, during implementation, a Gaussian kernel is introduced into the difference control factor for convolution operation to smooth the difference within the local neighborhood, resulting in a smoothed difference field. Subsequently, a tanh activation function is applied to the smoothed difference field using nonlinear enhancement parameters for amplitude limiting, ensuring numerical stability in high-amplitude regions and avoiding artifacts caused by over-enhancement. Finally, the amplitude-limited smoothed difference field is superimposed onto the structure-guided enhancement image via residual connection to generate a structure contrast mapping image. The mathematical model of the structure contrast mapping image is as follows:

[0147] ;

[0148] in, For structural contrast mapping images, For nonlinear enhancement parameters, For A Gaussian kernel with standard deviation;

[0149] In this embodiment, the nonlinear enhancement parameter To control the amplitude limiting intensity, this embodiment will Limited to the range [0.6, 1.0], and set... Standard deviation This embodiment will be used to adjust the range and intensity of convolution smoothing. Limited to [0.8, 1.2], and set By setting the parameters mentioned above, the obtained structural contrast mapping image can significantly improve the details of the structural edges while maintaining smoothness and consistency in a homogeneous background.

[0150] S5. Input the structural contrast mapping image into the image reconstruction module to improve the resolution.

[0151] Furthermore, in the specific implementation process, the structural contrast mapping image is input to the image reconstruction module for resolution enhancement processing; the image reconstruction module specifically includes: firstly, performing convolution operation on the structural contrast mapping image to achieve channel dimension expansion, and then completing spatial dimension upsampling through pixel rearrangement operation, thereby achieving 2x super-resolution reconstruction.

[0152] S6. Construct a super-resolution reconstruction model for thoracic surgical pathology slide images, train the model until convergence using high-resolution images as supervision signals, and realize super-resolution reconstruction of thoracic surgical pathology slide images.

[0153] Furthermore, a super-resolution reconstruction model for thoracic surgical pathology slide images was constructed, and discrete clustering tensor transformation, local structure-guided mapping, structural contrast mapping, and image reconstruction operations were performed sequentially. First, the low-resolution thoracic surgical pathology slide image was input into the discrete clustering tensor transformation module. A directional response clustering kernel tensor was generated by constructing a local directional lattice tensor and a response adjustment factor. This process performed multi-directional feature fusion and detail enhancement on the low-resolution thoracic surgical pathology slide image, highlighting texture and edge structures while suppressing noise interference, resulting in a directional clustering enhanced image. Subsequently, the directional clustering enhanced image was input into the local structure-guided mapping module. Within the neighborhood of the target location, statistics were performed based on robust local benchmark values ​​and local texture energy. Structural strength weights were calculated using variance balance parameters and contrast balance parameters, and the target pixel and benchmark difference were then compared. The image is amplified and superimposed onto the directional aggregation enhancement image using residual connections to achieve fine depiction of subtle textures and low-contrast regions, outputting a structure-guided enhancement image. Next, the structure-guided enhancement image is input to the structure contrast mapping module, where a weighted mean and variance are calculated within the neighborhood to construct a locally differentiable structural strength weight. The difference between the target pixel and its neighborhood is scaled and adjusted to obtain a difference control factor, which is then suppressed and constrained using a tanh activation function to ensure contrast balance and stability, outputting a structure contrast mapping image. Finally, the structure contrast mapping image is input to the image reconstruction module, where convolution and pixel rearrangement achieve a 2x resolution boost, resulting in a high-resolution thoracic surgical pathology slide image with clear texture and continuous structure. By integrating these stages end-to-end and employing an iterative backpropagation training mechanism that minimizes L1 loss, high-precision restoration and clear reconstruction of fine-grained structures in thoracic surgical pathology slide images are achieved.

[0154] Furthermore, the super-resolution reconstruction model for thoracic surgical pathology slide images proposed in this invention is implemented using the Python programming language. Based on the Linux operating system environment, the model is trained on an NVIDIA 3090 GPU platform using the PyTorch deep learning framework. The experimental dataset contains 650 pairs of training samples and 150 pairs of test samples. Random rotation and flipping operations are introduced during training to enhance data diversity and improve the model's generalization ability. The optimizer used is AdamW, with momentum factors β1 and β2 set to 0.9 and 0.999 respectively, weight decay coefficient λ set to 0.0001, and initial learning rate of 0.0002. The batch size used for model training is 32, with a total of 500 training epochs, and the L1 loss function is selected. Peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are used as performance evaluation metrics to comprehensively measure the visual quality and structural fidelity of the reconstructed thoracic surgical pathology slide images.

[0155] Furthermore, the dataset of thoracic surgical pathology slide images was input into the constructed super-resolution reconstruction model for processing, and the experimental results are as follows: Figure 5 As shown; Figure 5 The super-resolution results of thoracic surgical pathology slide images reconstructed by the model proposed in this invention are presented. It can be observed that while improving image resolution, the model can effectively preserve the detailed features of tissue texture and structural boundaries in the pathology slides, avoiding the blurring and structural distortion problems common in traditional interpolation methods. Compared with the original low-resolution image, the reconstruction results are significantly improved in terms of texture clarity, boundary sharpness, and structural continuity, especially showing higher restoration accuracy in lesion areas and fine blood vessels and cell boundaries, verifying the accuracy and robustness of this invention in the super-resolution reconstruction task of thoracic surgical pathology slide images.

[0156] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

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 the 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; taking the current position as a reference, selecting a neighborhood window as a statistical range, determining the corresponding neighborhood weighting coefficients based on the spatial distance between each pixel position in the neighborhood and the target pixel position, weighting and accumulating the pixel values of the structure-guided enhancement images in the neighborhood according to the neighborhood weighting coefficients determined by the distance from the center pixel, normalizing through the sum of the neighborhood weighting coefficients to obtain the weighted local mean; the current position is each pixel position of the structure-guided enhancement image; within the neighborhood range, calculate the difference between each pixel value of the structure-guided enhancement image and the weighted local mean, square each difference, and then accumulate the squared differences to obtain the local change factor; within the neighborhood range, calculate the weighted mean and variance of the structure-guided enhancement image, introduce the variance balance parameter and the contrast balance parameter to calculate the structure intensity weight, construct the difference regulation factor, perform convolution smoothing and nonlinear limiting processing, and then fuse it with the structure-guided enhancement image to obtain the structure contrast mapping image; input the structure contrast mapping image into the image reconstruction module for resolution enhancement; construct a thoracic surgery pathology section image super-resolution reconstruction model, train the model to convergence with high-resolution images as supervision signals, and realize the 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 exclude image samples with abnormal staining distribution, focal plane deviation, or blurred tissue structure; perform standardized patch 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 the corresponding high-resolution label images; perform pixel normalization 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 image, a plurality of offset positions are determined in a preset angle set and radius set with the current position as the center, the Euclidean distance square of the feature vectors of the current position pixels and the offset position pixels 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 the radius and a scale decay parameter, while a direction gain coefficient is introduced for amplitude adjustment to construct a response adjustment factor; Under all angle and radius combinations, the local directional point array tensor and the response adjustment factor are multiplied item by item, and are accumulated at the corresponding pixel positions to construct a direction 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; At each pixel position, when the direction response aggregation kernel tensor is greater than the structure adaptive threshold, a logarithmic enhancement item constructed by the direction response aggregation kernel tensor is superimposed on the eigenvalue of the low-resolution thoracic surgery pathology section image; When the direction response aggregation kernel tensor is less than or equal to the structure adaptive threshold, an exponential decay item constructed by the direction response aggregation kernel tensor is applied to the eigenvalue of the low-resolution thoracic surgery pathology section image to obtain a nonlinear response reconstruction tensor; Under all radius and angle combinations, the values of the nonlinear response reconstruction tensor at the corresponding offset positions are weighted and accumulated according to an aggregation adjustment factor to obtain a 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 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 high and low 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 the pixel values 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 a variance balance parameter are combined to form a suppression item, and the absolute value of the difference between the direction aggregation enhancement image and its robust local reference value is taken together with a contrast balance parameter to form an adjustment item; the suppression item and the adjustment item are multiplied to obtain a structure strength weight; The difference between the direction aggregation enhancement image and the robust local reference value is scaled by the structure strength weight to obtain a difference amplification item; On the basis of the difference amplification item, a nonlinear gain coefficient is introduced to modulate its amplitude, and a nonlinear voltage control process is performed on the modulated difference amplification item by combining an activation function; The direction aggregation enhancement image and the difference amplification item 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 1, 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; an absolute value of a difference between a pixel value of the structure-guided enhancement 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; 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 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; 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