Helicobacter pylori immunohistochemical staining result interpretation method and system based on artificial intelligence

Abstract

The invention provides a helicobacter pylori immunohistochemical staining result interpretation method and system based on artificial intelligence, and relates to the technical field of pathological diagnosis. The method sequentially comprises the steps of obtaining a full-view digital image, executing color standardization and denoising normalization preprocessing, detecting a thallus candidate region by using a first deep learning model and outputting a bounding box and spatial position information, performing color development condition classification by using a second deep learning model, and outputting a positive probability value. And finishing grading judgment based on a set threshold value, counting a positive area proportion to generate a slice-level interpretation conclusion, and generating a structured interpretation report in combination with spatial information. According to the method, the automatic interpretation efficiency and accuracy of the helicobacter pylori staining result can be effectively improved, the personal error is reduced, and the intelligence and standardization level of pathological diagnosis is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pathological diagnosis, in particular to a method and system for interpreting results of Helicobacter pylori immunohistochemical staining based on artificial intelligence. BACKGROUND

[0002] Detection of Helicobacter pylori (HP) in biopsy tissue of gastroscopy is still mainly based on manual microscope reading, often assisted by HE, Giemsa, methylene blue or immunohistochemical staining. Doctors judge whether there are brownish bacterial bodies on the stained section by naked eye and roughly estimate their quantity, thereby giving a grading result of "HP (-), HP (+), HP (++), HP (+++)" and the like.

[0003] With the maturity of digital pathology and deep learning technology, scanners can quickly digitize high-resolution section images; convolutional neural networks (CNN), visual Transformers and other models have been proven to achieve or even surpass expert level in tumor typing, lymph node metastasis detection and other scenarios. The industry is transitioning from "digital archiving" to "intelligent interpretation", promoting the informatization, automation and refinement of pathological diagnosis.

[0004] However, there are still many problems in the prior art, such as inconsistent identification of weak positive or bacterial body intensive area boundaries by different pathologists, resulting in poor result repeatability; large workload of manual interpretation of batch sections, which is prone to visual fatigue; low contrast, background coloring or artifacts will reduce the accuracy of manual recognition. And the existing deep learning scheme mainly focuses on morphological classification, and there is insufficient support for fine-grained bacterial counting and grading, and there is no integrated interpretation method deeply coupled with the pathological workflow. SUMMARY

[0005] In order to overcome the deficiencies of the prior art, the purpose of the present application is to provide a method and system for interpreting results of Helicobacter pylori immunohistochemical staining based on artificial intelligence, which effectively solves the technical bottlenecks of strong subjectivity, low efficiency and large error in traditional Helicobacter pylori staining interpretation.

[0006] To achieve the above-mentioned purpose, the present application provides the following solutions:

[0007] A method for interpreting results of Helicobacter pylori immunohistochemical staining based on artificial intelligence, comprising:

[0008] obtaining a full-field digital image of a Helicobacter pylori immunohistochemically stained tissue section;

[0009] performing color standardization, background denoising and scale normalization processing on the full-field digital image to obtain a preprocessed image;

[0010] In the pre-processed image, a first deep learning model is used to analyze the image region, output a bounding box set containing a plurality of candidate regions of bacteria and corresponding spatial position information;

[0011] The bounding box set is input into a second deep learning model, each bounding box set is classified according to the color development condition, and the positive probability value corresponding to each candidate region of bacteria is output;

[0012] According to a preset probability threshold rule, the positive probability value of each candidate region of bacteria is mapped to a corresponding hierarchical label, and the overall proportion of positive regions in the entire stained tissue section is counted to generate a section-level interpretation conclusion;

[0013] The section-level interpretation conclusion is fused with the spatial position information to generate a structured interpretation report.

[0014] Preferably, color standardization, background denoising and scale normalization processing are performed on the full-field digital image to obtain a pre-processed image, including:

[0015] Based on a preset reference color distribution model, a structured color transfer algorithm is used to perform color standardization on the full-field digital image to obtain a standardized image;

[0016] A background suppression algorithm based on Laplacian edge enhancement and region growing is used to remove background noise of non-tissue parts in the standardized image and enhance the color development condition boundary to obtain an enhanced image;

[0017] The enhanced image is resampled to a target image size under a unified resolution standard according to the pixel density value embedded in the scanning magnification information to obtain a pre-processed image.

[0018] Preferably, the expressions of the standardized image, the enhanced image and the pre-processed image are respectively:

[0019]

[0020]

[0021]

[0022] Wherein, is the original pixel vector of the full-field digital image; and are the color mean and covariance matrix of the full-field digital image, respectively; and are the mean and covariance matrix of the preset reference color distribution; is the pixel of the standardized image obtained by color transfer; The foreground mask output by the Laplacian edge enhancement-region growing combined algorithm has a value range of 0 or 1; The second-order spatial derivative of the Laplacian operator is applied to the preprocessed image to highlight the boundary of the color development case; The single enhancement factor is adaptively set by the local gradient variance; The per-channel multiplication is represented; The resampling normalization operator is performed on the corresponding pixel density according to the scanning magnification; The image pixel is enhanced; The final preprocessed image is obtained.

[0023] Preferably, the method for determining the foreground mask output by the Laplacian edge enhancement-region growing combined algorithm is as follows:

[0024] The normalized image is obtained by applying the Laplacian operator to the preprocessed image and taking the absolute value, and the gradient amplitude map at the pixel level is obtained

[0025] The global threshold is calculated from the gradient amplitude map based on the Otsu method , and the high threshold and the low threshold are set in proportion The pixel set satisfying is defined as the seed set

[0026] Starting from the seed set , the adjacent pixels are recursively absorbed according to the pixel color Euclidean distance and the gradient condition to obtain a plurality of connected domains ; wherein, The 0.5 times of the mean value of the color variance of the entire image is taken as an integer;

[0027] The connected domains with an area not less than the minimum effective area threshold are retained, and the pixels of the connected domains are assigned a value of 1, and the remaining pixels are assigned a value of 0 to generate a foreground mask .

[0028] Preferably, in the preprocessed image, a first deep learning model is used to analyze the image region, and a bounding box set containing a plurality of candidate regions of bacteria and corresponding spatial position information are output, including:

[0029] The preprocessed image is input into a convolution feature extractor with a residual network as the backbone to output a multi-scale feature map group;

[0030] ​​​​​Dense generate anchor box set on each scale feature map group according to preset aspect ratio and size parameters, and assign anchor box index;

[0031] Apply full convolution detection head on the anchor box set, respectively output the existence probability of each anchor box and the boundary box regression offset;

[0032] Decode the anchor box coordinates using the boundary box regression offset to obtain a preliminary candidate boundary box set, and remove redundant boxes based on the non-maximum suppression algorithm at a set intersection over union threshold;

[0033] Convert the remaining candidate boundary boxes to spatial position information in the full field of view coordinate system according to the resolution ratio relationship between the preprocessed image and the full field of view digital image, to obtain a boundary box set containing multiple candidate regions of bacteria and corresponding spatial position information.

[0034] Preferably, the boundary box set is input into a second deep learning model to classify each boundary box set according to the color development condition, and output a positive probability value corresponding to each candidate region of bacteria, including:

[0035] In the multi-scale feature map group of the preprocessed image, perform ROI Align operation according to the position of the boundary box set to obtain a uniform size feature tensor corresponding to each candidate region of bacteria;

[0036] Send the feature tensor into a double-layer self-attention module with residual connection to model the context interaction of local bacteria texture and surrounding tissue background information, and output context-enhanced features;

[0037] Apply a classification head containing a fully connected layer and a Soft-Max activation function to the context-enhanced features to obtain an initial probability vector of each color development condition; the color development conditions include negative, weak positive, moderate positive, and strong positive;

[0038] Use a temperature scaling factor pre-trained on a validation set to calibrate the initial probability vector to obtain a calibrated probability vector;

[0039] Extract the components corresponding to the positive color development categories from the calibrated probability vector as the positive probability value of the candidate region of bacteria and output them.

[0040] Preferably, the probability threshold rule at least includes:

[0041] Low threshold for distinguishing between negative and positive;

[0042] High threshold for distinguishing between moderate positive and high positive;

[0043] and satisfy 0 <1.

[0044] Preferably, the positive probability value of each bacterial candidate region is mapped to a corresponding hierarchical label according to a preset probability threshold rule, and the overall proportion of positive regions in the entire stained tissue section is counted to generate a section-level interpretation conclusion, including:

[0045] The positive probability value of each bacterial candidate region is compared with a preset threshold in turn and If , it is marked as negative, if , it is marked as moderately positive, if , it is marked as highly positive.

[0046] For the candidate region marked as positive, the pixel area of the candidate region in the preprocessed image is accumulated and the sum of the positive area is obtained , while the total area of the entire section tissue is calculated with the foreground mask ;

[0047] According to the formula , the positive region proportion is obtained ;

[0048] The positive proportion is compared with a preset proportion threshold set , when , HP(-) is output, when , HP(+) is output, when , HP(++) is output, when , HP(+++) is output, and the section-level interpretation conclusion is written into a structured interpretation report; the preset proportion threshold set satisfies .

[0049] Preferably, the section-level interpretation conclusion is fused with the spatial position information to generate a package structured interpretation report, including:

[0050] Based on the full-view digital image, each bacterial candidate region is given a corresponding color code according to its hierarchical label, and a bounding box and a number identification are drawn in the image to generate a bacterial candidate region visualization overlay;

[0051] The upper left and lower right pixel coordinates of each bacterial candidate region in the full-view image are recorded, and a coordinate index table is constructed in number order;

[0052] ​​The slice-level interpretation conclusion, the bacteria candidate region visual superposition graph, the coordinate index table and the generation time information are integrated and packaged into a structured interpretation report file conforming to the PDF / A-3 standard.

[0053] An artificial intelligence-based Helicobacter pylori immunohistochemical staining result interpretation system comprises:

[0054] A full-field image acquisition unit is configured to acquire a full-field digital image of a Helicobacter pylori immunohistochemically stained tissue section.

[0055] An image preprocessing unit is configured to perform color standardization, background denoising and scale normalization on the full-field digital image to obtain a preprocessed image.

[0056] A bacteria candidate detection unit is configured to analyze image regions in the preprocessed image using a first deep learning model to output a set of bounding boxes containing a plurality of bacteria candidate regions and corresponding spatial position information.

[0057] A color development classification unit is configured to input the set of bounding boxes into a second deep learning model to classify each set of bounding boxes according to color development conditions and output a positive probability value corresponding to each bacteria candidate region.

[0058] A hierarchical determination unit is configured to map the positive probability value of each bacteria candidate region to a corresponding hierarchical label according to a preset probability threshold rule, and to statistically determine the overall proportion of positive regions in the entire stained tissue section to generate a slice-level interpretation conclusion.

[0059] A report generation unit is configured to fuse the slice-level interpretation conclusion with the spatial position information to generate a structured interpretation report.

[0060] The present application discloses the following technical effects:

[0061] The present application proposes an integrated method that combines image preprocessing, candidate region detection, color development intelligent interpretation and structured report output to address the problems of strong subjectivity, low efficiency and large errors in the existing Helicobacter pylori immunohistochemical staining result interpretation process. BRIEF DESCRIPTION OF DRAWINGS

[0062] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description only aim to some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0063] Figure 1 The method flowchart provided for the embodiments of the present application is as shown in

[0064] Figure 2 The system structure diagram provided for the embodiments of the present application is as shown in DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments only aim to some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0066] The purpose of the present application is to provide a Helicobacter pylori immunohistochemical staining result interpretation method and system based on artificial intelligence, which can effectively solve the technical bottleneck of strong subjectivity, low efficiency and large error in traditional Helicobacter pylori staining interpretation.

[0067] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] Figure 1 The method flowchart provided for the embodiments of the present application is as shown in Figure 1 The present application provides a Helicobacter pylori immunohistochemical staining result interpretation method based on artificial intelligence, which comprises:

[0069] Step 100: obtaining a full-field digital image of a Helicobacter pylori immunohistochemical staining tissue section;

[0070] Step 200: performing color standardization, background denoising and scale normalization processing on the full-field digital image to obtain a preprocessed image;

[0071] Step 300: in the preprocessed image, a first deep learning model is used to analyze the image region, and a boundary box set containing multiple candidate bacterial regions and corresponding spatial position information are outputted;

[0072] Step 400: inputting the boundary box set into a second deep learning model, classifying the color development conditions of each boundary box set, and outputting a positive probability value corresponding to each candidate bacterial region;

[0073] Step 500: According to the preset probability threshold rule, the positive probability value of each bacterial candidate region is mapped to the corresponding hierarchical label, and the overall proportion of the positive region in the whole stained tissue section is counted to generate a section-level interpretation conclusion;

[0074] Step 600: The section-level interpretation conclusion is fused with the spatial position information to generate a package structured interpretation report.

[0075] Specifically, the process of step 200 of the embodiment is as follows:

[0076] In the embodiment, the full-field digital image of the Helicobacter pylori immunohistochemical stained tissue section is first acquired. To reduce the influence of staining differences between different experimental batches on the color development condition interpretation, color normalization processing is performed on the full-field digital image. Specifically, based on a preset reference color distribution model, a color transfer algorithm is used to map the image color distribution from the original mean μ_1, covariance _1 to the reference mean μ_R, covariance _R. The transformation relationship is as follows:

[0077]

[0078] wherein, represents the RGB vector value of the pixel point p in the original image, represents the pixel value after color normalization, μ and _ represent the mean and covariance matrix of the overall color distribution of the image. The image after normalization is used for subsequent tissue recognition and interpretation modeling.

[0079] Subsequently, the color normalized image is subjected to background noise removal and boundary enhancement processing. In the embodiment, the Laplacian operator is first used to extract the second-order gradient response of the image, and the edge saliency of each pixel is calculated:

[0080]

[0081] Based on the edge map L(p), the Otsu algorithm is used to obtain the global segmentation threshold τ, and the high threshold τH=1.2τ and the low threshold are further set, and the pixel points satisfying are extracted to form an initial seed set S. On this basis, the region growing algorithm is used to expand the connected region, and the growth conditions include that the color Euclidean distance satisfies:

[0082]

[0083] wherein, is a threshold constant determined based on overall color variance of the image, p is a pixel to be judged, and q is a neighboring pixel thereof. Finally, regions with a connected area less than a preset area threshold A_min are filtered out, and a main tissue region is retained to generate a foreground mask and edge features thereof are enhanced to obtain an enhanced image. Finally, to ensure consistency of slices from different sources in the input scale of the model, scale normalization processing is performed on the enhanced image. In this embodiment, pixel density s (unit: μm / pixel) is extracted from scanning magnification information of the image, and a unified target resolution r is set, and a required resampling ratio r is calculated and resampling operation is performed on the image based on the ratio. The operation generates an image of a unified size using a bicubic interpolation algorithm, that is, a preprocessed image in the present application which is used for subsequent detection of a bacterial candidate region and identification of a color development condition of a deep learning model.

[0084] In this embodiment, image color mean and covariance parameters required for color standardization are obtained by performing pixel-level statistical calculation on a reference set of dyeing slice images, wherein the reference set is composed of immunohistochemical slices with good dyeing quality and stable color distribution determined by laboratory experience, and the number of the slices is not less than 200 to ensure that the reference model is representative. Color statistical parameters of the original image are obtained by independently calculating all pixel channels of the current input image. An edge saliency threshold is automatically generated by an Otsu adaptive method, which determines a set of segmentation thresholds that can maximize the inter-class variance based on the overall gray scale distribution characteristics of the image; the high and low double thresholds are set to 120% and 80% of the threshold, respectively. The color similarity threshold during region growing is half of the average value of the color variance of all pixels in the standardized image to limit the growth range; the minimum effective area threshold is set to not less than 64 pixel points according to the actual observation scale of the salient region. The pixel density is directly read from the image scanner metadata, the target resolution adopts a standard value consistent with the model training data, and the ratio of the two is used to determine the resampling ratio to realize image scale normalization. All the above parameters are obtained by sample statistics, adaptive calculation or system embedded metadata, without manual intervention.

[0085] Optionally, in step 300 of the present embodiment, the preprocessed image is first input into a convolution feature extractor with a residual network as the backbone, and a feature pyramid structure is connected at the end of the backbone network, so that the features of different scales can simultaneously retain the morphological details and context information of Helicobacter pylori. For each scale feature layer, a set of anchor boxes is densely generated according to the aspect ratio and size parameters obtained by empirical statistics, and each anchor box is assigned a unique index to cover possible bacterial sizes and shapes.

[0086] A full convolutional detection head is applied on all anchor boxes, which is composed of a shared convolutional layer, a classification branch and a regression branch. The classification branch outputs the probability of the existence of H. pylori in each anchor box, and the regression branch predicts the horizontal and vertical displacement relative to the center point of the anchor box and the width-height offset. The detection head uses cross-entropy loss and smooth L1 loss to optimize in the training stage to improve the classification confidence and the boundary box positioning accuracy; in the inference stage, the probability of the existence of bacteria and the boundary box regression offset are obtained for each anchor box.

[0087] Finally, the anchor box coordinates are decoded by using the regression offset to obtain a preliminary candidate boundary box set; for the boundary boxes with an overlap degree higher than a set intersection ratio threshold for the same target, the non-maximum suppression algorithm is applied to retain the box with the highest probability and remove the redundancy. The filtered candidate boundary boxes are mapped to the full field coordinate system according to the resolution ratio relationship between the preprocessed image and the full field digital image, and finally a boundary box set containing multiple candidate regions of bacteria and the corresponding spatial position information are output.

[0088] In step 400 of the embodiment, first, for the boundary box set of the candidate region of bacteria, the ROI Align operation is performed in the multi-scale feature map group corresponding to the preprocessed image to extract the feature region in the boundary box with sub-pixel accuracy. According to the position of each boundary box, the operation maps it to the corresponding feature map scale and uniformly samples it into a fixed-size feature tensor through interpolation, so as to ensure that each candidate region can form a parallel processable input unit regardless of the original size.

[0089] Subsequently, the feature tensor is input into a double-layer self-attention module with a residual connection structure. The module is composed of two continuously stacked multi-head attention units and a feedforward neural network, which can model the fine-grained texture features of the bacteria in the candidate region and the color distribution context of the edge neighborhood, respectively. The residual connection enhances the stability of deep information propagation and helps the model to focus on both the significant color blocks of the positive bacteria and the interference suppression of the background tissue, thereby improving the classification discrimination ability.

[0090] Then, the context-enhanced feature tensor is input into a classification head containing a full connection layer and a Soft-Max activation function to perform multi-classification of the color development of each candidate region of bacteria, and an initial probability vector of four categories of negative, weak positive, moderate positive and strong positive is output. The classification result is trained based on the multi-class cross-entropy loss function, which can realize the discrimination and mapping of different positive degrees.

[0091] Finally, to improve the generalization ability of the model between different data sets, the temperature scaling calibration strategy is applied to the initial probability vector, and the temperature scaling factor is obtained by minimizing the negative log-likelihood loss on the independent validation set. The probability vector after temperature calibration is closer to the true distribution, and the component corresponding to the positive class is extracted as the positive probability value of the candidate region of the bacteria body, and the output is converged in all region dimensions to provide basic data support for subsequent grading determination.

[0092] In step 500 of the present embodiment, combined with the visual feature of positive brown staining of Helicobacter pylori immunohistochemical staining, the staining intensity grading information is used as the key discriminant of color development when training the second deep learning model, and the positive probability value of each bacteria body candidate region is output. To realize automatic grading interpretation, the system sets two probability thresholds, low threshold T1 and high threshold T2, which satisfy 0 < T1 < T2 < 1. Among them, T1 is used to distinguish between background or non-staining regions (i.e. non-brown) and positive initial regions, and T2 is used to distinguish between moderate brown and significant brown regions. The system compares the positive probability Pᵢ of each candidate region with T1 and T2. If Pᵢ < T1, it is considered as no obvious brown staining and marked as negative; if T1 ≤ Pᵢ < T2, it is considered as a small amount or mild brown color in the staining region, marked as moderate positive; if Pᵢ ≥ T2, the staining feature is obvious, and the brown region boundary is clear, marked as high positive.

[0093] After completing the regional grading, the system extracts all the bacteria body candidate regions marked as positive and counts their pixel areas in the preprocessed image. The bounding box of each region in the image is clear, and the system calculates the total number of pixels according to its horizontal and vertical coordinates, and then calculates the total area ΣAᵢ of all positive regions. At the same time, the foreground mask has accurately separated the tissue region from the background noise, and the system uses the mask to count the effective area A_tot of the tissue in the whole image to provide a basis for subsequent proportion calculation.

[0094] After obtaining the area of the positive region and the total area of the tissue, the proportion of the positive region in the whole slice R = ΣAᵢ / A tot Since the basis of regional grading is the brown color development, this proportion is equivalent to the "proportion of brown positive region in the tissue" in this embodiment. This proportion, as a comprehensive indicator, can quantitatively represent the Helicobacter pylori positive level, thereby replacing the manual estimation of "HP(+)", "HP(++)" and "HP(+++)" in traditional visual inspection.

[0095] The embodiment further sets a ratio threshold set {R1, R2, R3} satisfying 0 < R1 < R2 < R3 < 1. When the ratio R < R1, it indicates that there is no or very few brown positive areas in the image, and the output slice-level interpretation conclusion is "HP(-)"; when R1 ≤ R < R2, the brown positive areas are sporadically distributed, and "HP(+)" is output; when R2 ≤ R < R3, the staining area is more, and "HP(++)" is output; when R ≥ R3, the staining is dense and extensive, and "HP(+++) " is output. Finally, the slice-level interpretation conclusion is written into the structured interpretation report, and the position, classification and positive probability of each bacterial candidate area are saved together for archiving, reviewing or checking.

[0096] In step 600 of the embodiment, first, according to the slice-level interpretation conclusion and the classification label of each bacterial candidate area, the full-field digital image is subjected to visual superposition processing. The system sets the negative frame as transparent, the weak positive frame as light orange, the moderate positive frame as orange, and the strong positive frame as dark brown according to the pre-set color mapping rule; then a rectangular boundary frame is drawn for each candidate area in the original image coordinate system, and the number is marked beside the frame with Arabic numerals, generating a color superposition image containing all bacterial candidate areas, and a color-classification legend is attached to the lower right corner of the image, so that the reader can quickly identify the color development corresponding to each classification.

[0097] The embodiment synchronously records the upper left and lower right pixel coordinates of each bacterial candidate area, and sorts the number, classification label, positive probability value and coordinate information into a table in ascending order of number to generate a coordinate index table. The table contains fields: number, X1, Y1, X2, Y2, classification, probability, which can realize one-key jump to the corresponding area when viewing the report later, and facilitate the clinical doctors to review and compare the interpretation results.

[0098] Finally, the embodiment arranges the slice-level interpretation conclusion, the bacterial candidate area color superposition image, the coordinate index table and the generation timestamp according to the predetermined template, and embeds the patient information number and the scanner information in the metadata, and outputs the structured interpretation report file using the document specification conforming to the PDF / A-3 standard. The report file encapsulates the embedded PNG superposition image and CSV index table in the form of PDF container, ensuring long-term readability and data consistency in archiving, sharing and third-party system calling scenarios.

[0099] Corresponding to the above method, as shown in Figure 2 The embodiment also provides an artificial intelligence-based Helicobacter pylori immunohistochemical staining result interpretation system, which comprises:

[0100] A full-field image acquisition unit is configured to acquire a full-field digital image of a Helicobacter pylori immunohistochemically stained tissue section;

[0101] An image preprocessing unit is configured to perform color standardization, background denoising and scale normalization on the full field digital image to obtain a preprocessed image;

[0102] A candidate cell detection unit is configured to analyze image regions in the preprocessed image by using a first deep learning model, and output a bounding box set containing a plurality of candidate cell regions and corresponding spatial position information;

[0103] A color development classification unit is configured to input the bounding box set into a second deep learning model, classify each bounding box set according to color development conditions, and output a positive probability value corresponding to each candidate cell region;

[0104] A hierarchical determination unit is configured to map the positive probability value of each candidate cell region to a corresponding hierarchical label according to a preset probability threshold rule, and statistically determine the overall proportion of positive regions in the entire stained tissue section to generate a section-level interpretation conclusion;

[0105] A report generation unit is configured to fuse the section-level interpretation conclusion and the spatial position information to generate a structured interpretation report.

[0106] The beneficial effects of the present application are as follows:

[0107] (1) The present application combines color standardization with deep learning models to eliminate batch-to-batch staining differences and human subjective bias, enabling stable identification and automatic grading of H. pylori positive regions, and significantly improving the consistency and reliability of interpretation results.

[0108] (2) The present application uses a context enhancement mechanism and color development probability modeling strategy to effectively capture low-density, weak-brown cell distribution, improving the sensitivity of early infection or small cell region interpretation.

[0109] (3) The present application encapsulates the spatial coordinates of candidate regions, interpretation results and image overlay information into a standardized report format, supporting rapid positioning, review and data sharing, and improving the traceability and automation level of the diagnostic process.

[0110] (4) The present application introduces scale normalization and model lightweight design, making the method compatible with different types of scanners and resolution images, suitable for multi-institution, multi-platform clinical pathology assisted interpretation and batch quality control applications.

[0111] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts are described in the method part.

[0112] The principles and implementations of the present application are described in the specific examples, and the above examples are only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation and application range will be changed. In view of the above, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for interpreting Helicobacter pylori immunohistochemical staining results based on artificial intelligence, characterized in that: include: Acquire full-field digital images of tissue sections stained with Helicobacter pylori immunohistochemistry; Performing color standardization, background denoising and scale normalization processing on the full-field digital image to obtain a preprocessed image; In the preprocessed image, the image region is analyzed using a first deep learning model to output a set of bounding boxes containing multiple candidate bacterial regions and corresponding spatial position information; Input the bounding box set into the second deep learning model, classify the color of each bounding box set, and output the positive probability value corresponding to each bacterial candidate region; According to the preset probability threshold rule, the positive probability value of each bacterial candidate area is mapped to the corresponding graded label, and the overall proportion of positive areas in the entire stained tissue section is counted to generate a slice-level interpretation conclusion; The slice-level interpretation conclusion is integrated with the spatial location information to generate a package-structured interpretation report.

2. The method for interpreting Helicobacter pylori immunohistochemical staining results based on artificial intelligence according to claim 1, characterized in that: Performing color standardization, background denoising, and scale normalization on the full-field digital image to obtain a preprocessed image includes: Based on a preset reference color distribution model, a structured color transfer algorithm is used to perform color standardization on the full-field digital image to obtain a standardized image; Using a background suppression algorithm based on Laplace edge enhancement and region growing, background noise of non-tissue parts in the standardized image is eliminated and the color boundary is enhanced to obtain an enhanced image; The enhanced image is resampled to a target image size under a uniform resolution standard according to a pixel density value embedded in the scanning magnification information to obtain a preprocessed image.

3. The method for interpreting Helicobacter pylori immunohistochemical staining results based on artificial intelligence according to claim 2, characterized in that: The expressions of the standardized image, the enhanced image and the preprocessed image are respectively: in, is the original pixel vector of the full-field digital image; and are the color mean and covariance matrix of the full-field digital image respectively; and is the mean and covariance matrix of the preset reference color distribution; is the pixel of the standardized image obtained by color transfer; It is the foreground mask output by the Laplace edge enhancement-region growing joint algorithm, and its value range is 0 or 1; For The second-order spatial derivative of the Laplace operator is applied to highlight the boundary of the color display; for the reason A single enhancement factor with adaptive setting of local gradient variance; represents channel-by-channel multiplication; To scan by magnification The corresponding pixel density performs the resampling normalization operator; To enhance image pixels; The final pre-processed image.

4. The method for interpreting Helicobacter pylori immunohistochemical staining results based on artificial intelligence according to claim 3, characterized in that: The method for determining the foreground mask output by the Laplace edge enhancement-region growing joint algorithm is as follows: For standardized images Apply the Laplace operator and take the absolute value to obtain the pixel-level gradient magnitude map ; Based on the Otsu method, the gradient magnitude map Calculate the global threshold in , and set the high threshold proportionally With low threshold will satisfy The pixel set is defined as the seed set ; With the seed collection As the starting point, according to the Euclidean distance of pixel color and gradient conditions Recursively absorb adjacent pixels to obtain several connected domains ;in, Take 0.5 times the mean value of the color variance of the entire image; The reserved area is not less than the minimum effective area threshold The connected domain is assigned a value of 1 to the pixels in the connected domain and 0 to the remaining pixels to generate a foreground mask. .

5. The method for interpreting Helicobacter pylori immunohistochemical staining results based on artificial intelligence according to claim 1, characterized in that: In the preprocessed image, the first deep learning model is used to analyze the image region and output a set of bounding boxes containing multiple candidate bacterial regions and corresponding spatial position information, including: Inputting the preprocessed image into a convolutional feature extractor with a residual network as the backbone, and outputting a multi-scale feature map group; Based on the preset aspect ratio and size parameters on each scale feature map group, a dense anchor box set is generated and an anchor box index is assigned; Apply a fully convolutional detection head on the anchor box set, and output the bacterial presence probability and bounding box regression offset of each anchor box respectively; Decode the coordinates of the anchor boxes using the bounding box regression offset to obtain a preliminary set of candidate bounding boxes, and remove redundant boxes based on the non-maximum suppression algorithm under a set intersection-over-union ratio threshold; The retained candidate bounding boxes are converted into spatial position information in a full-view coordinate system according to the resolution ratio between the preprocessed image and the full-view digital image, thereby obtaining a bounding box set containing multiple bacterial candidate regions and corresponding spatial position information.

6. The method for interpreting Helicobacter pylori immunohistochemical staining results based on artificial intelligence according to claim 1, characterized in that: The bounding box set is input into the second deep learning model, the color display condition of each bounding box set is classified, and the positive probability value corresponding to each bacterial candidate region is output, including: In the multi-scale feature map group of the preprocessed image, a ROI Align operation is performed according to the position of the bounding box set to obtain a feature tensor of uniform size corresponding to each bacterial candidate region; The feature tensor is fed into a two-layer self-attention module with residual connections to perform contextual interaction modeling on the local bacterial texture and surrounding tissue background information, and output context-enhanced features; Applying a classification head including a fully connected layer and a soft-max activation function to the context-enhanced features to obtain an initial probability vector for each color condition; the color conditions include negative, weak positive, moderate positive, and strong positive; Performing distribution calibration on the initial probability vector using a temperature scaling factor pre-trained on a validation set to obtain a calibrated probability vector; The component corresponding to the positive color development category is extracted from the calibrated probability vector as the positive probability value of the bacterial candidate region and aggregated and outputted.

7. The method for interpreting Helicobacter pylori immunohistochemical staining results based on artificial intelligence according to claim 1, characterized in that: The probability threshold rule at least includes: Low threshold , used to distinguish between negative and positive; High Threshold , used to distinguish moderately positive from highly positive; And satisfy 0< < <1.

8. The method for interpreting Helicobacter pylori immunohistochemical staining results based on artificial intelligence according to claim 7, characterized in that: Based on the preset probability threshold rule, the positive probability value of each bacterial candidate area is mapped to a corresponding graded label, and the overall proportion of positive areas in the entire stained tissue section is counted to generate a slice-level interpretation conclusion, including: The positive probability value of each bacterial candidate region Sequentially with the preset threshold and Compare, if It is marked as negative if It is marked as moderately positive. It is marked as highly positive; For candidate regions marked as positive, the pixel area of ​​the candidate region in the preprocessed image is accumulated And get the sum of the positive areas , while using the foreground mask Calculate the total tissue area of ​​the entire section ; According to the formula Calculate the proportion of positive area ; The positive ratio Set with preset ratio threshold For comparison, when Output HP(-), when Output HP(+) when Output HP(++), when Output HP(+++) and write the slice-level interpretation conclusion into the structured interpretation report; the preset ratio threshold set meets .

9. The method for interpreting Helicobacter pylori immunohistochemical staining results based on artificial intelligence according to claim 1, characterized in that: The slice-level interpretation conclusion is integrated with the spatial location information to generate a structured interpretation report, including: Based on the full-field digital image, each bacterial candidate region is assigned a corresponding color code according to its classification label, and a bounding box and number mark are drawn in the image to generate a visual overlay map of the bacterial candidate region; For each candidate bacterial region, record its upper left corner and lower right corner pixel coordinates in the full field of view image, and construct a coordinate index table in numerical order; The slice-level interpretation conclusion, the bacterial candidate region visualization overlay map, the coordinate index table and the generation time information are integrated and packaged into a structured interpretation report file that complies with the PDF / A-3 standard.

10. An artificial intelligence-based Helicobacter pylori immunohistochemical staining result interpretation system, characterized in that: include: A full-field image acquisition unit, used for acquiring full-field digital images of Helicobacter pylori immunohistochemically stained tissue sections; An image preprocessing unit, configured to perform color standardization, background denoising, and scale normalization on the full-field digital image to obtain a preprocessed image; a bacterial candidate detection unit, configured to analyze an image region in the preprocessed image using a first deep learning model, and output a bounding box set containing a plurality of bacterial candidate regions and corresponding spatial position information; a color classification unit, configured to input the bounding box set into a second deep learning model, classify the color of each bounding box set, and output a positive probability value corresponding to each bacterial candidate region; A grading determination unit is used to map the positive probability value of each bacterial candidate region into a corresponding grading label according to a preset probability threshold rule, and to calculate the overall proportion of positive regions in the entire stained tissue section to generate a slice-level interpretation conclusion; A report generating unit is used to fuse the slice-level interpretation conclusion with the spatial position information to generate a structured interpretation report.

Images