A method and system for intelligent diagnostic analysis of pathological tissue samples

Abstract

This invention discloses an intelligent diagnostic analysis method and system for pathological tissue samples, relating to the field of pathological image processing technology. It constructs three independent feature maps—surface scratches, middle paraffin-debris mixture, and bottom global color variations—through RGB three-channel pixel separation and multi-scale feature extraction. Impurities are removed layer by layer from the surface inwards. The surface layer uses linear morphology adaptive filtering, the middle layer relies on binary discrimination rules to accurately distinguish impurities from lesions, and the bottom layer uses global color balance correction. After impurity removal, bilinear interpolation and texture synthesis interpolation algorithms are used to reconstruct the purified image with fidelity based on pixels of healthy tissue surrounding the lesion. Finally, a standardized pathological feature reference library is retrieved to automatically compare the purified image with the standard map and mark suspected lesion areas, achieving integrated processing of layered impurity removal, lesion fidelity preservation, and auxiliary judgment.

Description

Technical Field

[0001] This invention relates to the field of pathological image processing technology, specifically to an intelligent diagnostic analysis method and system for pathological tissue samples. Background Technology

[0002] Paraffin-embedded pathology sections are crucial for the clinical diagnosis of lesions and are widely used in digital pathological diagnostic scenarios across multiple departments, including gastroenterology, breast surgery, and gynecology. However, during the preparation of sections, such as dewaxing, mounting, staining, and transport, multiple layers of interfering impurities can easily accumulate: linear scratches on coverslips due to friction during mounting, solid paraffin particles remaining from incomplete dewaxing, small free tissue debris scattered during the trimming process, and diffuse background discoloration caused by unbalanced staining. These four types of impurities are layered and superimposed according to the imaging depth, forming a complex mixed background that seriously interferes with pathologists' accurate interpretation of tissue morphology and lesion characteristics.

[0003] Existing commercial image processing solutions and publicly available patented noise reduction technologies generally employ global mean filtering, median filtering, or single fixed threshold noise reduction, which can only eliminate random single-point salt-and-pepper noise. For the multi-layered complex impurity scenario addressed by this invention, there are three inherent drawbacks: First, it cannot differentiate between spatially distributed multi-layered impurities; global filtering either removes impurities incompletely or smooths large areas of image details, leading to the loss of tissue texture information. Second, the filtering operation simultaneously erases subtle gray-scale abrupt changes and fine textures caused by early micro-lesions, resulting in missed detection of in-situ micro-lesions. Third, small debris and punctate lesions have similar characteristics, easily leading to misidentification of debris as lesions and lesions as impurities, severely reducing the accuracy of pathological slide reading from low-quality slides prepared at the basal level. Currently, there is a lack of integrated pathological preprocessing and atlas comparison-assisted slide reading solutions that combine layered impurity removal with lesion fidelity preservation. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an intelligent diagnostic analysis method and system for pathological tissue samples. It constructs three independent feature maps—surface scratches, middle paraffin-debris mixture, and bottom global color variations—through RGB three-channel pixel separation and multi-scale feature extraction. Impurities are removed layer by layer from the surface inwards. The surface layer uses linear morphology adaptive filtering, the middle layer relies on binary discrimination rules to accurately distinguish impurities from lesions, and the bottom layer uses global color equalization correction. After impurity removal, bilinear interpolation and texture synthesis interpolation algorithms are used to reconstruct the purified image with fidelity based on pixels of healthy tissue surrounding the lesion. Finally, a standardized pathological feature reference library is retrieved to automatically compare the purified image with the standard map and mark suspected lesion areas, achieving integrated processing of layered impurity removal, lesion fidelity preservation, and auxiliary judgment.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In one aspect, an intelligent diagnostic analysis method for pathological tissue samples, the method comprising: S1: A digital pathology scanner is used to perform continuous full-field scanning of paraffin-embedded pathology sections to acquire original panoramic pathology images. S2: Perform RGB three-channel pixel separation on the original image, extract the single-channel grayscale information of R, G and B respectively, and based on three types of feature parameters such as pixel grayscale gradient, texture scale and edge curvature, divide the impurity into surface scratch impurity layer, middle paraffin-debris mixed impurity layer and bottom global background noise layer according to the depth position of the impurity imaging. Construct independent feature maps for each impurity layer. S3: Call the multi-layer background impurity intelligent layer stripping algorithm to remove impurities layer by layer. Each layer is configured with an adaptive filtering operator. The impurity removal process divides pixel attributes according to the preset impurity-lesion binary discrimination rule. Weak lesion pixels that meet the lesion feature threshold are locked and retained, and impurity pixels are marked and removed. S4: After peeling off the layered impurities layer by layer, for the blank pixel areas formed by impurity removal, the pixel set of adjacent healthy tissue around the lesion is selected as the interpolation reference data source, and the neighborhood texture interpolation algorithm is used to fill the missing pixels to generate a purified pathological tissue image without background interference. S5: Input the purified pathological tissue image into the pre-trained convolutional neural network diagnostic model, automatically complete the lesion coordinate localization, lesion classification and grading, and output a standardized electronic diagnostic report.

[0006] Furthermore, in S1, a digital pathology scanner is used to perform line-by-line continuous raster scanning on the entire paraffin section, and the images are stitched together to generate a full-view panoramic original pathological image. The scanner output images are uniformly in standard RGB bitmap format, which completely preserves tissue texture, lesion details, and information on four types of impurities: scratches, paraffin particles, tissue debris, and global color variations. The acquired images are cached in real time for later use, avoiding secondary compression of images and loss of details.

[0007] Furthermore, in S2, the original RGB image is split into three independent monochrome channel grayscale images: R, G, and B. Within each channel, three feature parameters are extracted pixel by pixel: grayscale gradient, local texture scale, and edge contour curvature. According to the depth of impurity imaging, it is divided into a surface scratch layer, a middle paraffin-debris mixed layer, and a bottom global staining impurity layer. A multi-scale feature pyramid is constructed using a multi-scale Gaussian difference operator, with three types of convolution scales: small, medium, and large. The small-scale operator extracts fine debris edge features, the medium-scale operator captures linear scratch contours, and the large-scale operator extracts continuous diffuse impurity regions, generating three independent impurity feature maps layer by layer.

[0008] Furthermore, in S3, the layer peeling sequence is as follows: Surface cover glass scratch removal: A linear morphology adaptive filtering algorithm is used to remove scratches; The separation of middle-layer paraffin particles from free tissue debris: A dynamic adaptive threshold filtering algorithm is used to generate pixel distinction thresholds through local statistical features, and then impurity removal and lesion preservation are completed according to binary discrimination rules. Bottom-level global diffuse color stripping: The global color balance correction algorithm is used to correct the background color deviation problem, correct the abnormal color background to the standard color distribution, and preserve the inherent color information of tissue and lesion.

[0009] Furthermore, in step S3, a linear morphology adaptive filtering algorithm is used to remove scratches. The algorithm formula is as follows: ,in, Coordinates within the surface scratch feature map The original pixel grayscale value; Output pixel grayscale values ​​at the same coordinate positions after surface scratch removal; These are the filter control coefficients used on the surface layer, adaptively tuned based on the linear length, width, and curvature characteristics of the scratches; The second-order Laplacian differential operator for the grayscale of surface pixels; For the full surface feature map Figure 2 Maximum absolute value of the first derivative It is a surface-specific binary morphological mask matrix, generated based on the linear contour features of the scratch. The scratch area is assigned a valid identifier, while the rest of the area is set to zero.

[0010] Furthermore, in step S3, a dynamic adaptive threshold filtering algorithm is employed to generate a pixel distinction threshold through local statistical features. This dynamic adaptive threshold filtering algorithm includes a threshold calculation formula and a pixel determination formula. The threshold calculation formula is as follows: The pixel determination formula is: ,in, For The average grayscale value of the pixels in the central neighborhood; The standard deviation of the grayscale of the corresponding neighboring pixels; The correction coefficient is based on the adaptive change of the morphology of intermediate impurities; This is a locally adaptive segmentation threshold; Threshold for determining lesion characteristics; The threshold for impurity detection; , These are the input and output pixel grayscale values ​​of the mid-layer feature map, respectively.

[0011] Furthermore, the binary discrimination rule is as follows: the gray value range of impurities and the gray value range of lesions are predefined to form a binary judgment benchmark. Through the single-point threshold calculated in real time, the gray value range of each pixel is compared. Paraffin particles and free debris pixels that fall into the value range corresponding to impurities are directly cleared and removed. Early micro lesion pixels whose gray value index falls into the value range of lesions retain their original pixel values. Normal tissue pixels whose gray value is in the intermediate transition range are retained as is.

[0012] Furthermore, the algorithm for correcting background color cast using a global color balance correction method is as follows: ,in: Represents any color channel of the RGB spectrum; Original single channel Location pixel value; Output pixel values ​​after color correction; , These represent the mean and standard deviation of the background values ​​for the corresponding channels in the bottom noise map; , These represent the mean and standard deviation of the corresponding channel in the standard noise-free reference image.

[0013] Furthermore, in S4, bilinear interpolation and texture synthesis interpolation algorithms are used to fill the missing pixels. The interpolation process is limited to selecting healthy pixels in the vicinity of the lesion as the interpolation data source. The color distribution, grayscale changes and texture direction features of the neighboring pixels are statistically analyzed simultaneously to complete the filling of missing pixels step by step.

[0014] For the blank pixel areas formed after impurity removal, coordinate positioning and area statistics are first performed to classify the blank areas into two categories: small blank areas (area ≤ 100 pixels) and large blank areas (area > 100 pixels). For each blank area, the set of healthy tissue pixels within a radius of ≤ 20 pixels around it is limited to the interpolation reference data source, and the use of any background noise pixels across the lesion boundary is prohibited.

[0015] For small gaps, a bilinear interpolation algorithm is used to fill them; for larger gaps, a block-matching-based texture synthesis interpolation algorithm is used to improve the texture reproduction effect in large areas. After filling, the edges of all gap areas are smoothed with a 3×3 Gaussian smoothing process to eliminate splicing marks.

[0016] The bilinear interpolation first performs single-dimensional grayscale interpolation along the horizontal pixel direction, and then performs secondary interpolation along the vertical pixel direction based on the horizontal interpolation result. Throughout the interpolation calculation, background noise pixels are prohibited from being taken across the lesion boundary. The original texture details of the missing position are restored according to the texture change rules of the surrounding healthy tissue.

[0017] On the other hand, a pathological tissue sample intelligent diagnostic analysis system includes an image acquisition unit, a layered feature parsing unit, a layered impurity stripping calculation unit, an image reconstruction unit, and an AI intelligent diagnostic unit. The image acquisition unit is used to acquire panoramic raw images of pathological slides and perform format standardization preprocessing. The hierarchical feature parsing unit is used to realize RGB channel splitting, multi-scale feature extraction, three-layer impurity layer division and pixel-level feature map construction; The layered impurity stripping calculation unit is embedded with a multi-layer background impurity intelligent layered stripping algorithm and a binary discrimination rule calculation program, which automatically removes impurities layer by layer, identifies and retains lesion pixels; The image reconstruction unit is equipped with a neighborhood texture interpolation and completion program, which automatically repairs and removes missing pixels and outputs purified pathological images. The AI-powered intelligent diagnostic unit is used to load the trained pathological diagnostic model to achieve lesion localization, classification and grading, and report generation. Beneficial effects Compared with existing technologies, this intelligent diagnostic analysis method and system for pathological tissue samples has the following advantages: This invention uses RGB three-channel separation and multi-scale feature extraction to divide scratches, paraffin particles and tissue debris, and global staining blemishes on paraffin sections into three independent feature maps according to the imaging depth. Differentiated filtering operators are configured layer by layer to perform targeted removal, avoiding the problems of incomplete removal or smoothing of details caused by global filtering. A binary discrimination rule is introduced in the middle layer to predefine the gray-level ranges of impurities and lesions, so that the pixels of small lesions and debris are strictly distinguished at the judgment level. This solves the problem that traditional noise reduction will simultaneously erase early weak lesions or misjudge debris as lesions. The empty areas after impurity removal are filled with bilinear interpolation and texture synthesis interpolation based on the pixels of healthy tissue around the lesion. Finally, combined with a standardized pathological feature reference library, the purified image is automatically compared with the standard map, and the suspected lesion area is automatically annotated, providing pathologists with intuitive auxiliary judgment basis.

[0018] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 A flowchart of an intelligent diagnostic analysis method for pathological tissue samples; Figure 2 This is a structural block diagram of an intelligent diagnostic analysis system for pathological tissue samples. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] This embodiment proposes an intelligent diagnostic analysis method and system for pathological tissue samples. It employs a data processing mechanism that utilizes RGB three-channel layered feature extraction, intelligent layered stripping of multi-layered background impurities, and lesion-fidelity reconstruction. This method overcomes the shortcomings of existing global filtering for incomplete impurity removal, easy omission of small lesions, and easy confusion between impurities and lesions. It achieves accurate separation of multi-layered composite impurities from pathological lesions and performs high-precision and robust intelligent diagnostic analysis on paraffin-embedded pathological tissue samples with different slide quality. It is particularly suitable for the clinical diagnostic needs of low-quality slides in primary hospitals.

[0023] like Figure 1 As shown, the intelligent diagnostic analysis method for pathological tissue samples in this embodiment specifically includes: S1: A digital pathology scanner is used to perform continuous full-field scanning of paraffin-embedded pathology sections to acquire original panoramic pathology images. S2: Perform RGB three-channel pixel separation on the original image, extract the single-channel grayscale information of R, G and B respectively, and based on three types of feature parameters such as pixel grayscale gradient, texture scale and edge curvature, divide the impurity into surface scratch impurity layer, middle paraffin-debris mixed impurity layer and bottom global background noise layer according to the depth position of the impurity imaging. Construct independent feature maps for each impurity layer. S3: Call the multi-layer background impurity intelligent layer stripping algorithm to remove impurities layer by layer. Each layer is configured with an adaptive filtering operator. The impurity removal process divides pixel attributes according to the preset impurity-lesion binary discrimination rule. Weak lesion pixels that meet the lesion feature threshold are locked and retained, and impurity pixels are marked and removed. S4: After peeling off the layered impurities layer by layer, for the blank pixel areas formed by impurity removal, the pixel set of adjacent healthy tissue around the lesion is selected as the interpolation reference data source, and the neighborhood texture interpolation algorithm is used to fill the missing pixels to generate a purified pathological tissue image without background interference. S5: Input purified pathological tissue images into a pathology-specific AI diagnostic model that has completed pre-training with massive amounts of samples, automatically complete lesion coordinate localization, lesion classification and grading, and output standardized electronic diagnostic reports.

[0024] Specifically, this invention addresses the distribution characteristics of four types of impurities generated during paraffin section preparation based on their imaging depth. It decomposes the complex mixed interference background into three independent layers: surface, middle, and bottom. Differential modeling and precise capture of the morphology, grayscale, and texture characteristics of impurities at each layer are performed. Surface features are modeled based on the high aspect ratio and low curvature characteristics of linear scratches; middle-layer features are modeled based on the grayscale value differences between granular impurities and punctate lesions; and bottom-layer features are modeled based on the global color distribution characteristics of diffused impurities. By configuring a dedicated adaptive filtering operator for each layer of impurities, corresponding impurity components are generated and precisely extracted from the original image. Simultaneously, according to a preset impurity-lesion binary discrimination rule, weak feature pixels of early micro-lesions are locked and preserved in real time during the impurity removal process. Finally, a hierarchical interpolation reconstruction algorithm based on healthy tissue texture is used to repair pixel gaps caused by impurity removal, resulting in a purified pathological image free of background interference. This image is then automated for diagnosis using a multi-scale pathology-specific AI model, achieving accurate identification and grading of different types and levels of pathological lesions.

[0025] Specifically, paraffin-embedded pathological sections underwent standardized preprocessing, with the section thickness controlled at 4μm±0.5μm. The process included dewaxing, HE staining, and mounting, ensuring uniform staining and the absence of obvious air bubbles. Before scanning, a dust-free air blower was used to remove surface dust from the sections to avoid introducing additional impurities. A fully automated digital pathology scanner equipped with a 40x flat-field apochromatic objective and a 0.75NA numerical aperture was used to perform continuous raster scanning of the entire section. A sub-pixel-level image stitching algorithm was used to generate a full-view panoramic original pathological image. The scanner output images were uniformly formatted as 24-bit uncompressed standard RGB bitmaps, with a sampling resolution set to 0.25μm / pixel, fully preserving tissue texture, lesion details, and information on four types of impurities: coverslip scratches, paraffin residue particles, free tissue debris, and global staining discoloration. Acquired images were immediately written to a high-speed RAID5 array cache, and any form of lossy compression was strictly prohibited throughout the process to prevent the loss of tissue and lesion details.

[0026] For example, scanning a standard breast pathology slide with a size of 25mm×75mm generates an original panoramic image with a pixel size of approximately 100000×300000, a file size of approximately 8.6GB, a scanning time of approximately 115s, and an array read / write speed that is stable at over 600MB / s, which can support real-time parallel computing in subsequent image processing workflows.

[0027] Specifically, the original RGB panoramic image was split into three independent single-channel grayscale images: R, G, and B. Based on the optical characteristics of HE-stained pathological images, the G channel with the highest contrast was used as the main feature extraction channel, while the R and B channels were used as auxiliary verification channels. Within each channel, three core feature parameters were calculated pixel by pixel: grayscale gradient, local texture scale, and edge contour curvature. According to the order of impurity imaging depth from the surface to the inside, all interfering impurities were divided into the surface cover glass scratch layer, the middle layer of paraffin-free tissue debris mixture layer, and the bottom layer of global staining blemish layer.

[0028] A three-level multi-scale feature pyramid was constructed using a multi-scale Gaussian difference operator, with downsampling ratios of 1:1, 1:4, and 1:16, corresponding to three different sizes of convolution kernels: 3×3, 15×15, and 64×64. The 3×3 small-scale convolution kernel was used to extract the edge features of fine free tissue debris with a diameter ≤5μm, the 15×15 medium-scale convolution kernel was used to capture the outline features of linear coverslip scratches with an aspect ratio ≥10:1, and the 64×64 large-scale convolution kernel was used to extract the features of contiguous diffuse background noise regions with an area ≥10000 pixels. Based on the above extraction results, three independent feature maps of surface, middle, and bottom impurities with a feature overlap of ≤3% were generated.

[0029] For example, feature extraction was performed on a gastric cancer pathological slide image containing four typical types of impurities. The generated surface feature map retained only the outlines of all linear scratches, the middle feature map retained only the dot-like features of paraffin particles and tissue debris, and the bottom feature map retained only the yellow color distribution information of the overall background. The three feature maps showed no obvious cross-interference, laying the foundation for subsequent layered impurity removal.

[0030] Specifically, for the surface feature map, a linear morphological adaptive filtering algorithm is used to remove scratches. First, all linear structures in the image are detected by Hough transform. Then, normal linear structures such as blood vessels and ducts in the tissue are excluded by the screening conditions of aspect ratio ≥10:1 and curvature ≤0.1, generating a surface-specific binary morphological mask matrix. Then, the linear edge features of the scratches are enhanced by the Laplacian second-order differential operator. Combined with the filter adjustment coefficients adaptively tuned according to the scratch morphology, only the scratch areas marked by the mask are grayscale corrected, while the original pixel values ​​of non-scratched areas remain unchanged.

[0031] The calculation process can be represented as follows: ,in, Coordinates within the surface scratch feature map The original pixel grayscale value; Output pixel grayscale values ​​at the same coordinate positions after surface scratch removal; These are the filter control coefficients used on the surface layer, adaptively tuned based on the linear length, width, and curvature characteristics of the scratches; The second-order Laplacian differential operator for the grayscale of surface pixels; For the full surface feature map Figure 2 Maximum absolute value of the first derivative It is a surface-specific binary morphological mask matrix, generated based on the linear contour features of the scratch. The scratch area is assigned a valid identifier, while the rest of the area is set to zero.

[0032] For example, a typical cover glass scratch with a length of 200 pixels, a width of 2 pixels, and an aspect ratio of 100:1 is processed and adaptively adjusted. =1.05, and after generating the corresponding binary mask matrix, a filtering operation is performed. After processing, the scratch completely disappears, and the edges of the normal gastric glandular epithelial tissue around the scratch are not blurred or lost in detail. The overall removal rate of surface scratches is ≥96%, and the misprocessing rate of normal tissue is ≤0.5%.

[0033] Specifically, for the mid-layer feature map, a dynamic adaptive threshold filtering algorithm combined with an impurity-lesion binary discrimination rule is used to remove impurities and retain lesions. The mean and standard deviation of the grayscale value of an 11×11 neighborhood centered on each pixel are calculated to generate a pixel-by-pixel local adaptive segmentation threshold. Then, based on the pre-defined and batch-calibrated grayscale value ranges of impurities and lesions, the attributes of each pixel are determined. Impurity pixels are cleared and removed, lesion pixels are locked and retained, and normal tissue pixels remain unchanged. Before each batch of scanning, the threshold is calibrated using standard HE-stained slides certified by the National Pathology Quality Control Center to ensure adaptation to color differences between different staining batches.

[0034] It includes a threshold calculation formula and a pixel determination formula, wherein the threshold calculation formula is: The pixel determination formula is: ,in, For The average grayscale value of the 11×11 neighboring pixels centered at the center; The standard deviation of the grayscale of the corresponding neighboring pixels; The correction coefficient is based on the adaptive change of the morphology of intermediate impurities, and is adaptively varied according to the impurity particle diameter in the range of 0.6 to 0.9. This is a locally adaptive segmentation threshold; The threshold for determining lesion characteristics is preset to an 8-bit grayscale value of 120 after calibration with standard slices. The impurity detection threshold is preset to an 8-bit grayscale value of 60 after calibration. , These are the input and output pixel grayscale values ​​of the mid-layer feature map, respectively.

[0035] The binary discrimination rule is as follows: the gray value range of impurities [0,60] and the gray value range of lesions [120,255] are predefined to form a binary discrimination benchmark. The gray value range is compared pixel by pixel. Paraffin particles and free debris pixels falling into the [0,60] range are directly cleared and removed. Early micro lesion pixels falling into the [120,255] range retain their original values ​​and are locked and retained. Normal tissue pixels with gray values ​​in the (60,120) transition range are retained as is.

[0036] For example, a mid-slice image containing a 3μm diameter early gastric cancer lesion and 4μm diameter paraffin particles is processed, and local neighborhood calculations are performed to obtain... =90, =15, adaptive tuning k=0.7, calculated local threshold T=100.5, the grayscale value of the pixel of the small lesion is 135, which is greater than... =120, locked and stored; paraffin particle pixel grayscale value is 45, less than... =60, which was cleared and removed. According to statistics, the retention rate of early micro lesions with a diameter ≤5μm in this step is ≥98.5%, and the probability of paraffin particles and tissue debris being misjudged as lesions is ≤1.8%.

[0037] Specifically, for the underlying feature map, a global color balance correction algorithm is used to correct the background color cast problem. First, the Otsu threshold segmentation algorithm is used to automatically extract the unorganized pure background region in the image, and the mean and standard deviation of the background in each channel of the region are calculated. Then, combined with the corresponding channel statistical parameters of the standard no-noise HE staining reference image, the color of each pixel in the original image is corrected channel by channel, correcting the abnormal noisy background to the standard color distribution, while completely preserving the inherent color information of the tissue and lesions.

[0038] The calculation process can be represented as follows: ,in, Represents any color channel of the RGB spectrum; Original single channel Location pixel value; Output pixel values ​​after color correction; , These represent the mean and standard deviation of the background values ​​for the corresponding channels in the bottom noise map; , These represent the mean and standard deviation of the corresponding channel in the standard noise-free reference image.

[0039] For example, a standard HE-stained slide certified by the National Pathology Quality Control Center was selected as a reference image, and its R channel was measured. =180、 =25, G channel =120、 =20, Channel B =100、 =15, correcting the underlying image of a colorectal pathology slide with obvious yellow tint. Before correction, the mean value of the background R channel was 225, and after correction, the mean value of the R channel was 181. The deviation from the standard value was ≤0.6%, the overall background color difference was ≤1.8, the color contrast between the tissue cell nucleus and cytoplasm was improved by more than 35%, and the color fidelity of the lesion was ≥96%.

[0040] Specifically, coordinate positioning and area statistics are performed on all blank pixel regions formed after layer-by-layer cleanup. The blank regions are divided into two categories: small blanks (area ≤ 100 pixels) and large blanks (area > 100 pixels). For each blank region, the set of healthy tissue pixels within a radius of ≤ 20 pixels around it is limited to the interpolation reference data source. It is prohibited to take any background noise pixels across the lesion boundary. For small blanks, a bilinear texture interpolation algorithm is used for filling. For large blanks, a block matching-based texture synthesis interpolation algorithm is used to improve the texture restoration effect of large areas. After filling, the edges of all blank regions are smoothed by 3×3 Gaussian to eliminate splicing marks.

[0041] For example, a circular empty area with a diameter of 15 pixels (approximately 177 pixels in area) is filled. 24 healthy colorectal acinar tissue pixels within a radius of 10 pixels around the empty area are selected as the reference data source. A texture synthesis interpolation algorithm is used for filling. After filling, the texture similarity between the empty area and the surrounding healthy tissue is ≥93%, with no obvious splicing traces, no artificial repair artifacts, and the overall integrity of the image reaches 100%.

[0042] Specifically, a pathology-specific convolutional neural network diagnostic model, pre-trained with over 150,000 labeled pathology samples, was loaded. This model uses ResNet-50 as its backbone network and combines it with the FPN feature pyramid to achieve multi-scale lesion detection. The classification head employs two fully connected layers to output the pathological classification and TNM grading of the lesions. The training data covers common pathological types from 12 clinical departments, including gastroenterology, breast cancer, gynecology, and respiratory medicine, and includes over 30% low-quality slides to ensure the model's generalization ability. Purified pathological tissue images are input into the model. First, the model uses a target detection network to accurately locate the lesion coordinates, generating lesion bounding boxes with an accuracy of ≤1 pixel. Then, a classification network completes the pathological classification and grading of the lesions. Finally, it automatically extracts key information such as lesion location, size, number, classification, and grading, generating a standardized electronic diagnostic report. The report can include clinically practical content such as immunohistochemical recommendations and follow-up suggestions.

[0043] For example, the purified colorectal pathological image after the above processing is input into the model. The model completes lesion localization within 2.8 seconds, accurately identifies an adenocarcinoma lesion with a diameter of 2.5 mm, with a subtyping accuracy of 97% and a grading accuracy of 94% for grade I. It also automatically generates a standardized electronic report containing lesion image, location annotation, pathological feature description, diagnostic conclusion and follow-up recommendations. The total report generation time is ≤4.5 seconds.

[0044] Optionally, the intelligent diagnostic analysis method for pathological tissue samples further includes a slide quality pre-assessment module, which acquires the original panoramic pathological image after scanning; automatically assesses three core quality indicators of the slide: impurity content, staining uniformity, and tissue integrity, and generates a slide quality score from 0 to 100. For severely substandard slides with a score <60, the module automatically prompts for re-slide preparation; for general quality slides with a score 60 ≤ score < 80, the module automatically adjusts the parameter thresholds of the subsequent impurity removal algorithm to improve the impurity removal effect.

[0045] On the other hand, such as Figure 2 As shown, the present invention also provides an intelligent diagnostic analysis system for pathological tissue samples, which is used to perform the above-described method. The system includes: Image acquisition unit: Used to control the digital pathology scanner, acquire panoramic raw images of pathology slides, and perform format standardization and cache preprocessing.

[0046] Hierarchical Feature Parsing Unit: Used to realize RGB channel splitting, multi-scale DoG feature extraction, and to divide mixed impurities into three independent feature maps: surface, middle and bottom layers.

[0047] Layered impurity stripping operation unit: Embedded multi-layer background impurity intelligent layered stripping algorithm and binary discrimination rule operation program, automatically removes impurities layer by layer in a directional manner, and accurately identifies and retains early micro lesion pixels.

[0048] Image reconstruction unit: Equipped with bilinear interpolation and texture synthesis interpolation completion program based on neighboring healthy tissue, it automatically repairs the missing pixel areas formed after impurity removal and outputs high-fidelity purified pathological images.

[0049] AI Intelligent Diagnostic Unit: Used to load the trained convolutional neural network diagnostic model, perform inference on the input purified pathological images, and automatically complete lesion localization, classification and grading, and standardized electronic report generation.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent diagnostic analysis of pathological tissue samples, characterized in that, The method includes: S1: A digital pathology scanner is used to perform continuous full-field scanning of paraffin-embedded pathology sections to acquire original panoramic pathology images. S2: Perform RGB three-channel pixel separation on the original image, extract the single-channel grayscale information of R, G and B respectively, and based on three types of feature parameters such as pixel grayscale gradient, texture scale and edge curvature, divide the impurity into surface scratch impurity layer, middle paraffin-debris mixed impurity layer and bottom global background noise layer according to the depth position of the impurity imaging. Construct independent feature maps for each impurity layer. S3: Call the multi-layer background impurity intelligent layer stripping algorithm to remove impurities layer by layer. Each layer is configured with an adaptive filtering operator. The impurity removal process divides pixel attributes according to the preset impurity-lesion binary discrimination rule. Weak lesion pixels that meet the lesion feature threshold are locked and retained, and impurity pixels are marked and removed. S4: After peeling off the layered impurities layer by layer, for the blank pixel areas formed by impurity removal, the pixel set of adjacent healthy tissue around the lesion is selected as the interpolation reference data source, and the neighborhood texture interpolation algorithm is used to fill the missing pixels to generate a purified pathological tissue image without background interference. S5: Input the purified pathological tissue image into the pre-trained convolutional neural network diagnostic model, automatically complete the lesion coordinate localization, lesion classification and grading, and output a standardized electronic diagnostic report.

2. The intelligent diagnostic analysis method and system for pathological tissue samples according to claim 1, characterized in that, The original pathological panoramic image was superimposed with four types of interfering impurities: linear scratches on the coverslip, paraffin residue particles, free tissue debris, and global staining blemishes.

3. The intelligent diagnostic analysis method for pathological tissue samples according to claim 1, characterized in that, In S2, when constructing the hierarchical impurity feature map, a multi-scale feature pyramid is constructed using a multi-scale Gaussian difference operator. Edge features of fine debris, linear scratches, and large-area continuous color impurities are extracted by convolution kernels of different scales, thereby achieving differentiated feature segmentation of the three types of hierarchical impurities.

4. The intelligent diagnostic analysis method for pathological tissue samples according to claim 1, characterized in that, In step S3, the layer peeling sequence is as follows: Surface cover glass scratch removal: A linear morphology adaptive filtering algorithm is used to remove scratches; The separation of middle-layer paraffin particles from free tissue debris: A dynamic adaptive threshold filtering algorithm is used to generate pixel distinction thresholds through local statistical features, and then impurity removal and lesion preservation are completed according to binary discrimination rules. Bottom-level global diffuse color stripping: The global color balance correction algorithm is used to correct the background color deviation problem, correct the abnormal color background to the standard color distribution, and preserve the inherent color information of tissue and lesion.

5. The intelligent diagnostic analysis method for pathological tissue samples according to claim 4, characterized in that, In step S3, a linear morphology adaptive filtering algorithm is used to remove scratches. The algorithm formula is as follows: ,in, Coordinates within the surface scratch feature map Original pixel grayscale value; Output pixel grayscale values ​​at the same coordinate positions after surface scratch removal; These are the filter control coefficients used on the surface layer, adaptively tuned based on the linear length, width, and curvature characteristics of the scratches; The second-order Laplacian differential operator for the grayscale of surface pixels; The maximum absolute value of the second-order differential of the entire surface feature map. It is a surface-specific binary morphological mask matrix, generated based on the linear contour features of the scratch. The scratch area is assigned a valid identifier, while the rest of the area is set to zero.

6. The intelligent diagnostic analysis method for pathological tissue samples according to claim 4, characterized in that, In step S3, a dynamic adaptive threshold filtering algorithm is used to generate a pixel distinction threshold through local statistical features. The dynamic adaptive threshold filtering algorithm includes a threshold calculation formula and a pixel determination formula. The threshold calculation formula is as follows: The pixel determination formula is: ,in, For The average grayscale value of the pixels in the central neighborhood; The standard deviation of the grayscale of the corresponding neighboring pixels; The correction coefficient is based on the adaptive change of the morphology of intermediate impurities; This is a locally adaptive segmentation threshold; Threshold for determining lesion characteristics; The threshold for impurity detection; , These are the input and output pixel grayscale values ​​of the mid-layer feature map, respectively.

7. The intelligent diagnostic analysis method for pathological tissue samples according to claim 1, characterized in that, The binary discrimination rule is as follows: the gray value range of impurities and the gray value range of lesions are predefined to form a binary judgment benchmark. Through real-time calculation of single-point threshold, the gray value range of each pixel is compared. Pixels of paraffin particles and free debris that fall into the value range corresponding to impurities are directly cleared and removed. Pixels of early micro lesions whose gray value index falls into the value range of lesions retain their original pixel values. Pixels of normal tissue whose gray value is in the intermediate transition range are retained as is.

8. The intelligent diagnostic analysis method for pathological tissue samples according to claim 1, characterized in that, The global color balance correction algorithm used to correct background color cast is as follows: ,in, Represents any color channel of the RGB spectrum; Original single channel Location pixel value; Output pixel values ​​after color correction; , These represent the mean and standard deviation of the background values ​​for the corresponding channels in the bottom noise map; , These represent the mean and standard deviation of the corresponding channel in the standard noise-free reference image.

9. A pathological tissue sample intelligent diagnostic analysis system, applicable to the pathological tissue sample intelligent diagnostic analysis method according to any one of claims 1-8, characterized in that, The system includes an image acquisition unit, a hierarchical feature parsing unit, a hierarchical impurity stripping calculation unit, an image reconstruction unit, and an AI intelligent diagnosis unit; The image acquisition unit is used to acquire panoramic raw images of pathological slides and perform format standardization preprocessing. The hierarchical feature parsing unit is used to realize RGB channel splitting, multi-scale feature extraction, three-layer impurity layer division and pixel-level feature map construction; The layered impurity stripping calculation unit is embedded with a multi-layer background impurity intelligent layered stripping algorithm and a binary discrimination rule calculation program, which automatically removes impurities layer by layer, identifies and retains lesion pixels; The image reconstruction unit is equipped with a neighborhood texture interpolation and completion program, which automatically repairs and removes missing pixels and outputs purified pathological images. The AI-powered intelligent diagnostic unit is used to load the trained pathological diagnostic model to achieve lesion localization, classification and grading, and report generation.

Images