A lesion pathology image sample recognition enhancement system

Abstract

The application discloses a kind of pathological image sample identification enhancement systems, it is related to pathological image processing technical field.The system includes sample acquisition module, dyeing deviation correction module, pathological dynamic focusing module and sample identification enhancement module;The present application, by collecting the original pathological image of multiple pathological tissues, and combining pseudo-spectrum matching and high-frequency residual retention strategy, the accurate correction of dye color deviation is realized;Further by saliency pre-positioning, multi-scale region sorting and local texture sensitive enhancement network, the focusing and detail enhancement of pathological region are completed;Finally, the present application improves the structural definition and texture expression ability of lesion area, enhances the identification robustness of different staining conditions and pathological types, and is suitable for pathological auxiliary diagnosis image enhancement and the like scene.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and pathological auxiliary recognition technology, specifically a disease pathological image sample recognition enhancement system. Background Technology

[0002] The Pathological Image Recognition and Enhancement System is a medical aid tool based on artificial intelligence and image processing technology. It aims to improve doctors' accuracy and efficiency in identifying lesions through intelligent analysis of pathological images. The system automatically detects, labels, and enhances abnormal lesion features (such as tumors and inflammation) in images using algorithms, while reducing noise interference and optimizing image contrast. This assists doctors in locating lesions and determining their nature more quickly and accurately. Its core function is to compensate for the subjectivity and fatigue errors of human interpretation, providing objective digital support for early disease screening, graded diagnosis, and treatment plan development. It represents a key technological application in smart healthcare.

[0003] However, in the existing process of identifying pathological images of lesions, there are technical problems such as unstable accuracy of existing pathological image identification, image feature distortion caused by inconsistent staining, and blurred boundaries of lesion areas. In the existing process of staining deviation correction, there are technical problems such as insufficient modeling of color drift between staining batches, components and equipment, and traditional color normalization methods ignoring structural details and easily causing image blurring or over-correction of colors. In the existing process of dynamic focusing of lesions, there are technical problems such as low accuracy of candidate region positioning, limited edge blurring enhancement effect, and poor interpretability of focusing strategy relying on depth model. Summary of the Invention

[0004] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a pathological image sample recognition and enhancement system. Addressing the technical problems of unstable recognition accuracy, image feature distortion caused by inconsistent staining, and blurred lesion region boundaries in existing pathological image sample recognition processes, this solution creatively employs a comprehensive pathological image enhancement method combining staining deviation correction and dynamic lesion focusing. This achieves end-to-end consistency optimization of pathological images from source data correction to regional focusing enhancement, improving the clarity, texture expression, and classification accuracy of lesion regions, and enhancing the system's robustness to differences in multi-center staining conditions. Furthermore, it addresses the shortcomings of existing staining deviation correction processes, such as insufficient modeling of color drift between staining batches, components, and equipment, and the tendency of traditional color normalization methods to ignore structural details and easily cause image blurring. To address the technical challenges of color overcorrection, this solution creatively employs a pseudo-spectral matching color gamut enhancement method combined with a high-frequency residual preservation strategy to correct staining color shifts. This achieves the preservation of tissue structure details and boundary contours while unifying the color space distribution, improving image color gamut consistency, structural clarity, and the stability of downstream recognition models. Furthermore, addressing the technical issues of low candidate region localization accuracy, limited edge blurring enhancement effects, and poor interpretability of focusing strategies relying on deep models in existing lesion dynamic focusing processes, this solution creatively adopts a multi-scale region sorting local lesion texture-sensitive enhancement network method for dynamic focusing of lesion regions. This achieves synergistic optimization among structure preservation, texture enhancement, and contrast regulation, effectively improving the lesion region recognition and general image enhancement capabilities, making it suitable for diverse needs across different lesion types and image resolutions.

[0005] The technical solution adopted by the present invention is as follows: The present invention provides a pathological image sample recognition and enhancement system, including a sample acquisition module, a staining deviation correction module, a lesion dynamic focusing module and a sample recognition and enhancement module;

[0006] The sample acquisition module is used for acquiring pathological image samples of lesions. Through the acquisition of pathological image samples of lesions, the original pathological image data is obtained, and the original pathological image data is sent to the staining deviation correction module.

[0007] The staining deviation correction module is used to correct the staining color deviation. Through staining color deviation correction, a color-standardized image is obtained, and the color-standardized image is sent to the lesion dynamic focusing module.

[0008] The lesion dynamic focusing module is used for dynamic focusing of the lesion area. By dynamically focusing the lesion area, a lesion focused image is obtained, and the lesion focused image is sent to the sample recognition enhancement module.

[0009] The sample recognition enhancement module is used to enhance pathological images of lesions, and through the enhancement of pathological images of lesions, an enhanced recognition feature map is obtained.

[0010] Furthermore, the pathological image sample acquisition is used to obtain images of the lesion area to be analyzed, specifically by acquiring raw data and performing basic preprocessing operations to obtain raw pathological image data.

[0011] The original pathological image data specifically includes the original image to be corrected and the standard stained reference image;

[0012] The original image to be corrected is used to perform dye color shift correction; the standard dyeing reference image is used as the source image for dyeing standard sample correction of the dye color shift correction.

[0013] Furthermore, the staining color shift correction is used to correct the color difference in the image caused by differences in staining batches or components through color space mapping. Specifically, based on the original pathological image data, a pseudo-spectral matching color gamut enhancement method combined with a high-frequency residual preservation strategy is used to perform staining color shift correction to obtain a color-normalized image. This includes the following steps: pseudo-spectral vector generation, standard color gamut model construction, adaptive color gamut matching, pseudo-spectral inverse transformation, high-frequency residual preservation structure enhancement, and color shift correction.

[0014] The pseudo-spectral vector generation is used to extract the color shift pattern in the spectral space. Specifically, it uses principal component analysis to convert the RGB vectors of the original image pixels into pseudo-spectral vector data based on the original image to be corrected in the original pathological image data, thereby obtaining spectral projection matrix data.

[0015] The standard color gamut model is constructed to build a stable color distribution benchmark across images. Specifically, based on the standard stained reference image in the original pathological image data, the standard spectral projection matrix data is obtained by generating pseudo-spectral vectors from the standard stained reference image, and the standard color gamut model data is obtained by statistically analyzing the Gaussian distribution parameters of the standard pseudo-spectrum based on the standard spectral projection matrix data.

[0016] The adaptive color gamut matching is used to register the color gamut distribution of the source image to the standard distribution. Specifically, based on the standard color gamut model data, a linear kernel matching strategy is adopted and an improved matching matrix is ​​constructed to perform color gamut registration of the original image to be corrected on the spectral projection matrix data, so as to obtain color gamut matching correction data.

[0017] The pseudo-spectral inverse transform is used to map the corrected pseudo-spectral image to the RGB image space. Specifically, it involves performing a pseudo-spectral inverse transform operation on the color gamut matching correction data to obtain inverse-corrected image data.

[0018] The high-frequency residual preservation structure enhancement is used to improve the ability of cell boundaries and texture details to be preserved during the color correction process. Specifically, by introducing a high-frequency residual preservation strategy, cell boundaries and texture details are enhanced in the inverse correction image data to obtain detail-enhanced color correction image data.

[0019] The color shift correction specifically involves applying color shift correction to the original image to be corrected using the obtained detail-enhanced color-corrected image data to obtain a color-normalized image.

[0020] Furthermore, the dynamic focusing of the lesion area is used to dynamically clarify and focus on key lesion areas. Specifically, based on the color-normalized image, a local lesion texture-sensitive enhancement network method with multi-scale region sorting is used to dynamically focus the lesion area to obtain a focused lesion image. This includes the following steps: pre-positioning of salient areas, extraction and segmentation of candidate areas, multi-scale region structure sorting, dynamic focus enhancement, and dynamic focusing of the lesion area.

[0021] The salient region pre-location is used to predict suspected lesion areas in an image and narrow down the search space. Specifically, it involves converting the color-normalized image into a standard grayscale image to obtain grayscale image data, and then extracting edges using the Sobel operator to obtain edge intensity map data. The edge intensity map data and the color-normalized image data are then normalized and fused to obtain salient region map data. Finally, a salient region pre-location mask is generated by setting a salientity threshold parameter to obtain binary mask map data for the pre-located lesion area.

[0022] The candidate region extraction and segmentation is used to filter out suspected lesion local image patches. Specifically, based on the binary mask image data of the pre-located lesion region, all pixel image regions with a binary pixel value of 1 are extracted as candidate image patches, and connected component analysis is performed on the candidate image patches to extract the minimum bounding rectangle of each candidate image patch to obtain suspected lesion local image patch data.

[0023] The multi-scale regional structure sorting is used to select the most informative focused lesion regions from the local patches of suspected lesions. Specifically, based on the local patch data of suspected lesions, a texture scoring weighting function is constructed to perform weighted calculation of regional structure information to obtain local lesion information content scoring parameters.

[0024] The dynamic focus enhancement is used to enhance the boundary texture features of the focused lesion area. Specifically, it selects the top three areas with the highest scores in the local lesion information content scoring parameters as the focus enhancement object image, and constructs a lesion texture sensitive enhancement network to enhance the structural details of the focus enhancement object image to obtain a lesion boundary texture focus enhancement image.

[0025] The lesion texture-sensitive enhancement network includes a guided edge enhancement subnetwork, a multi-scale texture enhancement subnetwork, and a local brightness contrast modulation subnetwork;

[0026] The guided edge enhancement subnetwork is used to explicitly capture the edge contour of the lesion area before enhancement, providing structural guidance. Specifically, it is constructed by a three-layer convolutional network that embeds the standard Sobel operator and attention channel weights, with the number of channels set to 16, 32 and 64 respectively.

[0027] The multi-scale texture enhancement subnetwork is used to extract and enhance information at different texture scales. Specifically, it is constructed by building three 3×3 convolutional paths and then concatenating and fusing the outputs of each convolutional path after feature pooling.

[0028] The local brightness contrast control subnet is used to adaptively adjust the contrast according to the changes in the brightness of the image patch. Specifically, it calculates the local brightness and standard deviation and constructs a dynamic adjustment factor to perform adaptive contrast adjustment enhancement based on brightness fluctuations.

[0029] The dynamic focusing of the lesion area is used to output a focused image. Specifically, based on the focused enhancement image of the lesion boundary texture, the lesion area is dynamically focused on the color-normalized image to obtain a focused image of the lesion.

[0030] Furthermore, the pathological image enhancement is used to optimize the boundary and texture features of the lesion structure. Specifically, based on the focused image of the lesion, the focused image of the lesion is fused with the original pathological image data by pixel weighting to obtain an enhanced recognition feature map.

[0031] The beneficial effects achieved by adopting the above solution are as follows:

[0032] (1) In the process of identifying existing pathological images, there are technical problems such as unstable identification accuracy of existing pathological images, image feature distortion caused by inconsistent staining, and blurred boundaries of lesion area identification. This solution creatively adopts a comprehensive pathological image enhancement method that combines staining deviation correction and dynamic focusing of lesions. It realizes end-to-end consistency optimization of pathological images from source data correction to regional focusing enhancement, improves the clarity, texture expression ability and classification recognition accuracy of lesion areas, and enhances the robustness of the system to differences in multi-center staining conditions.

[0033] (2) In the existing process of color deviation correction, there are technical problems such as insufficient modeling of color drift between dyeing batches, components and equipment, and traditional color normalization methods ignoring structural details and easily causing image blurring or color overcorrection. This solution creatively adopts a pseudo-spectral matching color gamut enhancement method combined with high-frequency residual preservation strategy to correct the color shift of dyes. This achieves the preservation of tissue structure details and boundary contours while unifying the color space distribution, and improves the consistency of image color gamut, structural clarity and the stability of downstream recognition models.

[0034] (3) In view of the technical problems in the existing dynamic focusing process of lesions, such as low candidate region localization accuracy, limited edge blur enhancement effect, and poor interpretability of focusing strategy relying on deep model, this solution creatively adopts the local lesion texture sensitive enhancement network method of multi-scale region sorting to perform dynamic focusing of lesion region, realizing synergistic optimization between structure preservation, texture enhancement and contrast regulation, effectively improving the recognition of lesion region and the general image enhancement capability, and is suitable for the diverse needs of different lesion types and image resolutions. Attached Figure Description

[0035] Figure 1 A schematic diagram of the structure of a pathological image sample recognition and enhancement system provided by the present invention;

[0036] Figure 2 A flowchart illustrating the steps performed by the system provided for this invention;

[0037] Figure 3 This is a flowchart illustrating the steps performed in the staining deviation correction module.

[0038] Figure 4 This is a flowchart illustrating the steps performed by the lesion dynamic focusing module.

[0039] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] Example 1, see Figure 1The technical solution adopted by the present invention is as follows: The present invention provides a pathological image sample recognition and enhancement system, including a sample acquisition module, a staining deviation correction module, a lesion dynamic focusing module and a sample recognition and enhancement module;

[0042] The sample acquisition module is used for acquiring pathological image samples of lesions. Through the acquisition of pathological image samples of lesions, the original pathological image data is obtained, and the original pathological image data is sent to the staining deviation correction module.

[0043] The staining deviation correction module is used to correct the staining color deviation. Through staining color deviation correction, a color-standardized image is obtained, and the color-standardized image is sent to the lesion dynamic focusing module.

[0044] The lesion dynamic focusing module is used for dynamic focusing of the lesion area. By dynamically focusing the lesion area, a lesion focused image is obtained, and the lesion focused image is sent to the sample recognition enhancement module.

[0045] The sample recognition enhancement module is used to enhance pathological images of lesions, and through the enhancement of pathological images of lesions, an enhanced recognition feature map is obtained.

[0046] By performing the above operations, this solution addresses the technical problems in existing pathological image sample recognition processes, such as unstable recognition accuracy, image feature distortion caused by inconsistent staining, and blurred lesion region boundaries. It creatively employs a comprehensive pathological image enhancement method combining staining deviation correction and dynamic lesion focusing. This achieves end-to-end consistency optimization of pathological images from source data correction to regional focusing enhancement, improving the clarity, texture representation, and classification accuracy of lesion regions, and enhancing the system's robustness to differences in staining conditions across multiple centers. Specifically, in traditional digital pathology recognition, inconsistent staining conditions between different hospitals cause significant differences in the representation of the same type of lesion in the image color space, leading to the failure of pre-trained model transfer. Furthermore, whole-image processing strategies often neglect the spatial focusing characteristics of lesion regions, resulting in a significant decrease in recognition rate, especially in breast cancer scenarios with small lesions and multiple backgrounds.

[0047] Example 2, this example is based on the above example, see below. Figure 1 , Figure 2 The pathological image sample acquisition is used to obtain images of the lesion area to be analyzed. Specifically, it involves obtaining raw pathological image data through raw data acquisition and basic preprocessing operations.

[0048] The raw data acquisition specifically involves selecting pathological slide images of tumor tissue sections as raw data samples and obtaining the raw image set through routine HE staining.

[0049] The basic preprocessing operation specifically refers to performing image format normalization, size cropping normalization, and preliminary manual screening operations sequentially on the original image set to obtain an optimized image set;

[0050] Preferably, by performing color clustering on the images in the optimized image set, calculating the color concentration score and contrast index score, and using the image with the highest color concentration score and the lowest contrast index score as the standard staining reference image, the original pathological image data is obtained by selecting the standard staining reference image for each lesion type from the optimized image set.

[0051] The specific types of lesions include lung adenocarcinoma lung tissue lesions, invasive breast cancer breast tissue lesions, hepatocellular carcinoma liver tissue lesions, and gastric mucosal glandular atrophy tissue lesions.

[0052] The original pathological image data specifically includes the original image to be corrected and the standard stained reference image;

[0053] The original image to be corrected is used to perform dye color shift correction; the standard dyeing reference image is used as the standard sample correction reference source image for the dye color shift correction.

[0054] The data content of the original pathological image data specifically includes the image RGB matrix, image resolution parameters, image size parameters, pathological lesion label information, and image acquisition metadata.

[0055] The image acquisition metadata specifically includes image number, acquisition date, staining method, sample number, and tissue location.

[0056] Example 3, this example is based on the above examples, see reference. Figure 1 , Figure 2 and Figure 3 The staining color shift correction is used to correct the color difference in the image caused by differences in staining batches or components by color space mapping. Specifically, based on the original pathological image data, a pseudo-spectral matching color gamut enhancement method combined with a high-frequency residual preservation strategy is used to correct the staining color shift and obtain a color-standardized image. The steps include: pseudo-spectral vector generation, standard color gamut model construction, adaptive color gamut matching, pseudo-spectral inverse transformation, high-frequency residual preservation structure enhancement, and color shift correction.

[0057] The pseudo-spectral vector generation is used to extract the color shift pattern in the spectral space. Specifically, it uses principal component analysis to convert the RGB vectors of the original image pixels into pseudo-spectral vector data based on the original image to be corrected in the original pathological image data, thereby obtaining spectral projection matrix data.

[0058] Preferably, the principal component analysis method introduces a principal component feature weighting factor for pseudo-spectral vector transformation, and the calculation formula is as follows:

[0059] ;

[0060] In the formula, S x,y It is spectral projection matrix data. It is the transpose of the principal component projection matrix, I rgb This is a three-channel vector of the original image, where x is the horizontal pixel index and y is the vertical pixel index. It is the pixel mean vector, w c It is the principal component feature weighting factor;

[0061] The standard color gamut model is constructed to build a stable color distribution benchmark across images. Specifically, based on the standard stained reference image in the original pathological image data, the standard spectral projection matrix data is obtained by generating pseudo-spectral vectors from the standard stained reference image, and the standard color gamut model data is obtained by statistically analyzing the Gaussian distribution parameters of the standard pseudo-spectrum based on the standard spectral projection matrix data.

[0062] The formula for calculating the Gaussian distribution parameters of the standard pseudospectrum is as follows:

[0063] ;

[0064] In the formula, It is the standard pseudospectral mean vector. It is the standard pseudospectral covariance matrix, s std It is standard spectral projection matrix data. It is a Gaussian distribution statistical function;

[0065] The adaptive color gamut matching is used to register the color gamut distribution of the source image to the standard distribution. Specifically, based on the standard color gamut model data, a linear kernel matching strategy is adopted and an improved matching matrix is ​​constructed to perform color gamut registration of the original image to be corrected on the spectral projection matrix data, so as to obtain color gamut matching correction data.

[0066] The calculation formula for the linear kernel matching strategy is as follows:

[0067] ;

[0068] In the formula, This is color gamut matching correction data, and A is the improved matching matrix. It is the pseudo-spectral mean vector of the source image that needs to be registered in color gamut. It is the mean vector of the standard pseudospectrum;

[0069] The formula for calculating the improved matching matrix is ​​as follows:

[0070] ;

[0071] In the formula, It is the pseudospectral covariance matrix of the source image that needs to be registered in color gamut;

[0072] The pseudo-spectral inverse transform is used to map the corrected pseudo-spectral image to the RGB image space. Specifically, it involves performing a pseudo-spectral inverse transform operation on the color gamut matching correction data to obtain inverse-corrected image data.

[0073] The calculation formula for the inverse correction graphic data is:

[0074] ;

[0075] In the formula, It is inversely modified image data. It is the principal component projection matrix;

[0076] The high-frequency residual preservation structure enhancement is used to improve the ability of cell boundaries and texture details to be preserved during the color correction process. Specifically, by introducing a high-frequency residual preservation strategy, cell boundaries and texture details are enhanced in the inverse correction image data to obtain detail-enhanced color correction image data.

[0077] The calculation formula for the detail-enhanced color-corrected image data is as follows:

[0078] ;

[0079] In the formula, I enh It is image data with enhanced detail and color correction. is the high-frequency residual enhancement weighting factor, and GB(·) is the Gaussian blur operation function;

[0080] Preferably, the high-frequency residual enhancement weighting factor is used to control the enhancement intensity, and the default value range is [0.3, 0.5].

[0081] The color shift correction specifically involves applying color shift correction to the original image to be corrected using the obtained detail-enhanced color-corrected image data to obtain a color-normalized image.

[0082] By performing the above operations, this solution addresses the technical problems in existing staining deviation correction processes, such as insufficient modeling of color drift between staining batches, components, and equipment, and the tendency of traditional color normalization methods to ignore structural details and cause image blurring or over-correction. It creatively employs a pseudo-spectral matching color gamut enhancement method combined with a high-frequency residual preservation strategy to correct staining color shifts. This achieves the preservation of tissue structural details and boundary contours while unifying the color space distribution, improving image color gamut consistency, structural clarity, and the stability of downstream recognition models. Specifically, while commonly used color normalization methods such as the Macenko or Reinhard methods can normalize color distribution in the RGB space, they cannot simultaneously ensure color reproduction and edge detail preservation. This is particularly problematic in tissues requiring high staining fineness, such as lung adenocarcinoma, where edge blurring or nuclear pseudo-color drift often occurs, thus affecting the performance of segmentation or detection models.

[0083] Example 4, this example is based on the above examples, see below. Figure 1 and Figure 4 The dynamic focusing of the lesion area is used to dynamically clarify and focus on key lesion areas. Specifically, based on the color-normalized image, a local lesion texture-sensitive enhancement network method with multi-scale region sorting is used to dynamically focus the lesion area to obtain a focused lesion image. The method includes the following steps: pre-positioning of salient areas, extraction and segmentation of candidate areas, sorting of multi-scale region structures, dynamic focusing enhancement and dynamic focusing of lesion areas.

[0084] The salient region pre-location is used to predict suspected lesion areas in an image and narrow down the search space. Specifically, it involves converting the color-normalized image into a standard grayscale image to obtain grayscale image data, and then extracting edges using the Sobel operator to obtain edge intensity map data. The edge intensity map data and the color-normalized image data are then normalized and fused to obtain salient region map data. Finally, a salient region pre-location mask is generated by setting a salientity threshold parameter to obtain binary mask map data for the pre-located lesion area.

[0085] The calculation formula for the normalized fusion is as follows:

[0086] ;

[0087] In the formula, S(x,y) represents the saliency region map data, x is the horizontal pixel index, and y is the vertical pixel index. The normalized fusion weight is set to 0.4 by default. E(x,y) is the edge intensity map data, and C(x,y) is the color-normalized image data.

[0088] The saliency threshold parameter has a default value of 0.6; the calculation formula for generating the saliency region pre-positioning mask is as follows:

[0089] ;

[0090] In the formula, M(x,y) is the binary mask image data of the pre-localized lesion region. It is a significance threshold parameter;

[0091] The candidate region extraction and segmentation is used to filter out suspected lesion local image patches. Specifically, based on the binary mask image data of the pre-located lesion region, all pixel image regions with a binary pixel value of 1 are extracted as candidate image patches, and connected component analysis is performed on the candidate image patches to extract the minimum bounding rectangle of each candidate image patch to obtain suspected lesion local image patch data.

[0092] The multi-scale regional structure sorting is used to select the most informative focused lesion regions from the local patches of suspected lesions. Specifically, based on the local patch data of suspected lesions, a texture scoring weighting function is constructed to perform weighted calculation of regional structure information to obtain local lesion information content scoring parameters.

[0093] The texture scoring weighting function specifically calculates the local lesion information content scoring parameters by sequentially weighting the edge density parameter, grayscale standard deviation parameter, structural complexity parameter, and structural compactness parameter of the suspected lesion local patch data.

[0094] The edge density parameter is specifically extracted from the edge intensity map data; the grayscale standard deviation parameter is specifically extracted from the grayscale image data; the structural complexity parameter is specifically obtained by calculating the information entropy parameter; and the structural compactness parameter is specifically obtained by calculating the quotient of the region pixel area divided by the area of ​​the circumscribed rectangle through connected component analysis.

[0095] Preferably, the formula for calculating the edge density parameter is:

[0096] ;

[0097] In the formula, Edge Density is the edge density parameter;

[0098] The formula for calculating the grayscale standard deviation parameter is as follows:

[0099] ;

[0100] In the formula, Grayscale Contrast is the grayscale standard deviation parameter, N is the total number of pixels, i is the pixel index, and I... i It is the grayscale value of the i-th pixel. It is the average grayscale value of all pixels;

[0101] The formula for calculating the structural complexity parameter is as follows:

[0102] ;

[0103] In the formula, Entropy is the structural complexity parameter, and L is the grayscale value. It is a grayscale index. Is the grayscale value The probability of a pixel appearing in an image;

[0104] The formula for calculating the structural compactness parameter is as follows:

[0105] ;

[0106] In the formula, Extent is the structural compactness parameter;

[0107] The dynamic focus enhancement is used to enhance the boundary texture features of the focused lesion area. Specifically, it selects the top three areas with the highest scores in the local lesion information content scoring parameters as the focus enhancement object image, and constructs a lesion texture sensitive enhancement network to enhance the structural details of the focus enhancement object image to obtain a lesion boundary texture focus enhancement image.

[0108] The lesion texture-sensitive enhancement network includes a guided edge enhancement subnetwork, a multi-scale texture enhancement subnetwork, and a local brightness contrast modulation subnetwork;

[0109] The guided edge enhancement subnetwork is used to explicitly capture the edge contour of the lesion area before enhancement, providing structural guidance. Specifically, it is constructed by a three-layer convolutional network that embeds the standard Sobel operator and attention channel weights, with the number of channels set to 16, 32 and 64 respectively.

[0110] The multi-scale texture enhancement subnetwork is used to extract and enhance information at different texture scales. Specifically, it is constructed by building three 3×3 convolutional paths and then concatenating and fusing the outputs of each convolutional path after feature pooling.

[0111] The local brightness contrast control subnet is used to adaptively adjust the contrast according to the changes in the brightness of the image patch. Specifically, it calculates the local brightness and standard deviation and constructs a dynamic adjustment factor to perform adaptive contrast adjustment enhancement based on brightness fluctuations.

[0112] The dynamic focusing of the lesion area is used to output a focused image. Specifically, based on the focused enhancement image of the lesion boundary texture, the lesion area is dynamically focused on the color-normalized image to obtain a focused image of the lesion.

[0113] Example 5, this example is based on the above examples, see below. Figure 1 and Figure 2 The pathological image enhancement is used to optimize the boundary and texture features of the lesion structure. Specifically, based on the focused image of the lesion, the focused image of the lesion is fused with the original pathological image data by pixel weighting to obtain an enhanced recognition feature map.

[0114] By performing the above operations, this solution addresses the technical problems of low candidate region localization accuracy, limited edge blurring enhancement effect, and poor interpretability of focusing strategies relying on deep models in existing dynamic lesion focusing processes. It creatively employs a multi-scale region sorting local lesion texture-sensitive enhancement network method for dynamic focusing of lesion regions, achieving synergistic optimization between structure preservation, texture enhancement, and contrast regulation. This effectively improves the recognition accuracy of lesion regions and the general image enhancement capability, making it suitable for diverse needs of different lesion types and image resolutions. Specifically, while existing attention mechanisms or saliency models can guide focusing in deep networks, they lack fine-grained texture level modeling. Especially in hepatocellular carcinoma with complex lesion boundaries and uneven density, this can easily lead to region expansion and texture fading, affecting the classification boundary accuracy of subsequent recognition modules. This solution, through multi-scale screening and edge-guided enhancement fusion, better matches the actual pathological manifestations.

[0115] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process or method.

[0116] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0117] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A system for identifying and enhancing pathological images of lesions, characterized in that: It includes a sample acquisition module, a staining deviation correction module, a lesion dynamic focusing module, and a sample recognition enhancement module; The sample acquisition module is used for acquiring pathological image samples of lesions. Through the acquisition of pathological image samples of lesions, the original pathological image data is obtained, and the original pathological image data is sent to the staining deviation correction module. The staining deviation correction module is used to correct the staining color deviation. Through staining color deviation correction, a color-normalized image is obtained, and the color-normalized image is sent to the lesion dynamic focusing module. The module includes the following steps: pseudo-spectral vector generation, standard color gamut model construction, adaptive color gamut matching, pseudo-spectral inverse transformation, high-frequency residual preservation structure enhancement, and color deviation correction. The pseudo-spectral vector generation adopts the principal component analysis method, based on the original image to be corrected in the original pathological image data, converts the RGB vector of the original image pixels into pseudo-spectral vector data, and obtains spectral projection matrix data. The standard color gamut model is constructed by generating pseudo-spectral vectors from a standard stained reference image to obtain standard spectral projection matrix data, and then performing Gaussian distribution parameter statistics on the standard pseudo-spectrum based on the standard spectral projection matrix data to obtain standard color gamut model data. The adaptive color gamut matching employs a linear kernel matching strategy and constructs an improved matching matrix to perform color gamut registration on the spectral projection matrix data of the original image to be corrected, thereby obtaining color gamut matching correction data. The pseudo-spectral inverse transform is performed on the color gamut matching correction data to obtain inverse-corrected image data; The high-frequency residual preservation structure enhancement introduces a high-frequency residual preservation strategy to enhance cell boundaries and texture structure in the inverse correction image data, resulting in detail-enhanced color-corrected image data. The color shift correction yields a color-normalized image; The lesion dynamic focusing module is used for dynamic focusing of the lesion area. By dynamically focusing the lesion area, a lesion focused image is obtained, and the lesion focused image is sent to the sample recognition enhancement module. The module includes the following steps: pre-positioning of salient regions, extraction and segmentation of candidate regions, sorting of multi-scale region structures, dynamic focusing enhancement, and dynamic focusing of the lesion area. The candidate region is extracted and segmented to obtain local patch data of suspected lesions; The multi-scale regional structure sorting is used to construct a texture scoring weighting function, and the regional structure information content is weighted to obtain the local lesion information content scoring parameters. The dynamic focus enhancement selects the top three regions with the highest scores in the local lesion information content scoring parameters as the focus enhancement object image, and enhances the structural details of the focus enhancement object image by constructing a lesion texture sensitive enhancement network to obtain a lesion boundary texture focus enhancement image; The lesion texture-sensitive enhancement network includes a guided edge enhancement subnetwork, a multi-scale texture enhancement subnetwork, and a local brightness contrast modulation subnetwork; The lesion area is dynamically focused by performing dynamic focusing on the lesion area on the color-normalized image to obtain a focused lesion image. The sample recognition enhancement module performs pixel-weighted fusion of the focused image of the lesion and the original pathological image data to obtain an enhanced recognition feature map.

2. The pathological image sample recognition and enhancement system according to claim 1, characterized in that: The acquisition of pathological image samples is used to obtain images of the lesion area to be analyzed. Specifically, it involves obtaining raw pathological image data through raw data acquisition and basic preprocessing operations. The original pathological image data specifically includes the original image to be corrected and the standard stained reference image; The original image to be corrected is used to perform dye color shift correction; the standard dyeing reference image is used as the source image for dyeing standard sample correction of the dye color shift correction.

3. The pathological image sample recognition and enhancement system according to claim 2, characterized in that: The dynamic focusing of the lesion area is used to dynamically clarify and focus on key lesion areas. Specifically, based on the color-normalized image, a local lesion texture-sensitive enhancement network method with multi-scale region sorting is used to dynamically focus the lesion area and obtain a focused lesion image. The method includes the following steps: pre-location of salient regions, extraction and segmentation of candidate regions, sorting of multi-scale region structures, dynamic focusing enhancement, and dynamic focusing of the lesion area.

4. The pathological image sample recognition and enhancement system according to claim 3, characterized in that: The salient region pre-location is used to predict suspected lesion areas in an image and narrow down the search space. Specifically, it involves converting the color-normalized image into a standard grayscale image to obtain grayscale image data, and then extracting edges using the Sobel operator to obtain edge intensity map data. The edge intensity map data and the color-normalized image data are then normalized and fused to obtain salient region map data. Finally, a salient region pre-location mask is generated by setting a salientity threshold parameter to obtain binary mask map data for the pre-located lesion area. The candidate region extraction and segmentation is used to filter out local image patches of suspected lesions. Specifically, based on the binary mask image data of the pre-located lesion region, all pixel image regions with a binary pixel value of 1 are extracted as candidate image patches, and connected component analysis is performed on the candidate image patches to extract the minimum bounding rectangle of each candidate image patch, thereby obtaining local image patch data of suspected lesions.

5. The pathological image sample recognition and enhancement system according to claim 4, characterized in that: The guided edge enhancement subnetwork is used to explicitly capture the edge contour of the lesion area before enhancement, providing structural guidance. Specifically, it is constructed by a three-layer convolutional network that embeds the standard Sobel operator and attention channel weights, with the number of channels set to 16, 32 and 64 respectively. The multi-scale texture enhancement subnetwork is used to extract and enhance information at different texture scales. Specifically, it is constructed by building three 3×3 convolutional paths and then concatenating and fusing the outputs of each convolutional path after feature pooling. The local brightness contrast control subnet is used to adaptively adjust the contrast according to the changes in the brightness of the image patches. Specifically, it calculates the local brightness and standard deviation and constructs a dynamic adjustment factor to perform adaptive contrast adjustment enhancement based on brightness fluctuations.

Images