A method and system for automatically interpreting the results of tissue chip immunohistochemistry

Abstract

This invention provides an automated method and system for interpreting tissue microarray immunohistochemistry (IHC) results. The method includes acquiring and automatically cropping a full slide image, segmenting it into image blocks containing individual pathological samples, labeling target regions within the cropped image blocks to generate labeled data, which includes category information for positive and negative staining areas. After data processing, an optimized SOLOv2 instance segmentation model is trained using Dice Loss as the contour regression loss function and Focal Loss as the classification loss function to obtain a trained interpretation model. The trained interpretation model is then used to predict the image blocks at various points within the tissue microarray, outputting the immunohistochemical staining intensity grade and the percentage of positive cells for each image block, and automatically matching this with clinical information to generate structured results. This invention achieves standardized and objective interpretation of tissue microarray IHC slides, improving the consistency and efficiency of tissue microarray IHC interpretation.

Description

Technical Field

[0001] This invention relates to the field of medical computing technology, and in particular to a method and system for automatically interpreting tissue microarray immunohistochemistry results. Background Technology

[0002] In the interpretation of tissue microarray immunohistochemistry (IHC) results, a semi-quantitative assessment method is currently widely used in the research field. Typically, antibody staining intensity is first assigned a score of 0, 1, 2, or 3 based on the absence / weak / medium / strong staining, and then scored from 0 to 4 based on the percentage of positive cells (0-100%) in increments of 25%. The expression levels are then converted into a binary classification (low expression / high expression) by summing or multiplying these scores and combining them with thresholds set by the literature or the interpreter's experience. This is then used for statistical analysis in conjunction with clinicopathological features.

[0003] Because most researchers lack clinical pathology experience, their interpretations often require assistance from pathologists or require them to learn the techniques on the fly and make independent judgments. The former can easily lead to delays in the work process, while the latter can easily result in interpretation biases or even errors due to insufficient professional knowledge. Furthermore, even experienced pathologists can exhibit inconsistencies in results between different observers and even within the same observer at different times, affecting data stability and comparability.

[0004] In recent years, the widespread application of tissue microarrays has driven the development of immunohistochemical interpretation towards automation, high throughput, and quantification. Digital pathology image analysis systems can, to some extent, overcome the subjectivity and efficiency bottlenecks of traditional manual interpretation, enabling standardized quantitative analysis of batch samples and improving data reproducibility and accuracy. The multi-sample integration characteristics of tissue microarrays also help to unify interpretation standards, providing technical support for large-scale clinical research and biomarker validation. However, existing automation solutions still have shortcomings in the automatic interpretation of intensity and positive percentage of tissue microarray immunohistochemical images, particularly lacking a complete workflow that can directly output intensity and percentage scores and generate standardized results.

[0005] Therefore, there is an urgent need for a standardized and objective method and system for automatically interpreting tissue microarray immunohistochemistry results to reduce the influence of subjective factors and improve the consistency of interpretation. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for automatic interpretation of tissue microarray immunohistochemistry results, aiming to solve the technical problems of inaccurate interpretation, low efficiency and inconsistent results of immunohistochemistry by non-pathologists, so as to realize automatic quantitative interpretation of tissue microarray IHC.

[0007] To achieve the above objectives, in a first aspect, the present invention provides a method for automatically interpreting tissue microarray immunohistochemistry results, the steps of which include:

[0008] S1. Obtain whole slide imaging (WSI) of the tissue chip. Preferably, the whole slide imaging is generated by scanning the entire tissue chip using a digital pathology slide scanner.

[0009] S2. The whole slide image is automatically cropped to segment image blocks including individual pathological samples;

[0010] S3. Mark the target region in the cropped image block to generate annotation data, which includes the positive staining intensity level and the category information of the negative region.

[0011] S4. Process the labeled image blocks and convert the labeled data into COCO format;

[0012] S5. The processed image patch is trained based on the optimized SOLOv2 instance segmentation model. During the training process, Dice Loss is used as the loss function for contour regression and Focal Loss is used as the classification loss function. The trained interpretation model is output.

[0013] S6. Use the trained interpretation model to automatically predict the image blocks at each point in the tissue chip, and output the immunohistochemical staining intensity grade and the percentage of positive cells for each image block.

[0014] S7. Automatically match the predicted staining intensity grade and percentage of positive cells with the corresponding clinical information to generate structured interpretation results.

[0015] As a further improvement to the above solution, in step S2, the automatic cutting step includes:

[0016] First, the original WSI image is reduced to a preset size and then binarized using the Otsu threshold segmentation algorithm. Noise is removed by combining morphological operations. Then, the contour and geometric center coordinates of the pathological sample are extracted using an edge detection algorithm and mapped back to the original high-resolution image for precise cropping.

[0017] As a further improvement to the above scheme, the SOLOv2 instance segmentation model uses ResNet50 as the backbone and FPN as the neck, and combines transfer learning and dynamic convolution techniques to generate cell binary masks.

[0018] Preferably, a matrix nonmaximum suppression algorithm is introduced into the SOLOv2 instance segmentation model to improve the detection speed.

[0019] As a further improvement to the above scheme, in step S5, when training the model, an enhancement method including at least flipping, rotating, mirroring, translating, scaling and filtering is used to optimize the SOLOv2 instance segmentation model.

[0020] As a further improvement to the above scheme, the method for automatically matching the predicted staining intensity grade and percentage of positive cells with the corresponding clinical information in step S7 is as follows:

[0021] S71. Determine the row and column coordinates of chip points in the tissue chip through cluster analysis;

[0022] S72. Name each chip point, establish an image data identification system according to preset rules, associate it with pathology number, and realize a one-to-one mapping between immunohistochemical interpretation results and clinical pathological features.

[0023] Preferably, the preset rule is: chip name_antibody name_horizontal index_vertical index_coordinate X_coordinate Y_width_height.

[0024] As a further improvement to the above scheme, the steps for determining the row and column coordinates of chip points in the tissue chip through cluster analysis are as follows:

[0025] S711. First, cluster analysis is performed on the x-axis and y-axis coordinates of each chip point to determine the spatial distribution structure of the chip and obtain a two-dimensional array consisting of n rows and m columns.

[0026] S712. Then, based on the geometric center point of each pathological sample, calculate the minimum distance between it and the center of the already clustered rows and columns, thereby determining the row and column coordinates of the corresponding pathological sample.

[0027] As a further improvement to the above scheme, the structured interpretation results include the coordinate information, staining intensity grading, percentage of positive cells, and corresponding clinicopathological feature data for each image block.

[0028] As a further improvement to the above scheme, in step S5, when training the model, multi-center tissue chip data of at least laryngeal cancer, esophageal cancer and nasopharyngeal cancer are used for training, and the SGD optimizer is used to improve the model's adaptability to samples from different sources.

[0029] As a further improvement to the above scheme, a training set for an automated immunohistochemistry interpretation model is constructed. During data annotation, at least two pathologists annotate the same target area, which is a positive cell cluster or stained area. The outline and geometric center coordinates of each annotation are recorded.

[0030] Calculate the Euclidean distance between the center points of the two annotations and the intersection-union ratio (IoU) of the two annotation regions;

[0031] A qualified threshold is set. If the Euclidean distance is less than a preset value and the IoU is greater than a preset value, the annotations are considered to be basically consistent, and the annotation data of experienced doctors are selected into the training set. If the Euclidean distance is too large or the IoU is too small, the annotation consistency is considered insufficient. A third senior pathologist or the original annotator will review and correct the annotations until the consistency requirements are met before they can be used.

[0032] Secondly, the present invention also provides an automatic interpretation system for tissue microarray immunohistochemistry results, comprising:

[0033] The image acquisition module is used to acquire whole slide imaging (WSI) of tissue microarrays.

[0034] An automatic cropping module is used to segment the whole slide imaging and extract image blocks including individual pathological samples;

[0035] The annotation module is used to annotate target regions in image patches and generate annotation data;

[0036] The data processing module is used to perform format conversion, image preprocessing, data augmentation, and data normalization on the labeled image blocks, and to convert the labeled data into COCO format;

[0037] The model training module is used to train the processed image patches based on the SOLOv2 instance segmentation model and output the trained interpretation model.

[0038] The prediction module is used to automatically predict image patches at various points in the tissue chip using a trained interpretation model, and output the staining intensity grade and the percentage of positive cells.

[0039] The results matching module is used to automatically match the predicted results with the corresponding clinical information to generate structured interpretation results.

[0040] As a further improvement to the above solution, the automatic cropping module includes an image scaling unit, a binarization processing unit, a noise removal unit, a contour extraction unit, and a precise cropping unit, which are used to sequentially perform WSI image reduction, binarization segmentation, noise removal, sample contour extraction, and precise cropping of high-resolution images.

[0041] As a further improvement to the above scheme, the SOLOv2 instance segmentation model in the model training module uses ResNet50 as the backbone and FPN as the neck. During training, Dice Loss is used as the loss function for contour regression and Focal Loss is used as the classification loss function. The model performance is optimized by data augmentation methods such as flipping, rotating, mirroring, translating, scaling, and filtering.

[0042] As a further improvement to the above scheme, the result matching module determines the row and column positions through spatial coordinate clustering analysis of chip points, matches the row and column coordinates by combining the coordinates of the geometric center point, and finally associates clinical data with pathology numbers to generate structured data including staining intensity grading, percentage of positive cells, and clinicopathological features.

[0043] Because the present invention adopts the above technical solutions, the beneficial effects of this application are as follows:

[0044] This invention provides an automated method for interpreting tissue microarray immunohistochemistry results. Through an automated, standardized, and quantitative interpretation process, it overcomes the subjectivity, inefficiency, and inconsistency of traditional manual interpretation. At the same time, it is adapted to the high-throughput characteristics of tissue microarrays and the complexity of multi-center data. It provides researchers and clinicians with an efficient, accurate, and reproducible immunohistochemistry analysis tool, which significantly promotes the progress of biomarker screening and precision medicine research.

[0045] Specifically, firstly, an optimized SOLOv2 instance segmentation model (combining ResNet50Backbone and FPNNeck architectures, trained using Dice Loss and Focal Loss dual loss functions, and incorporating matrix nonmaximum suppression to improve detection speed) is used to automatically segment and interpret immunohistochemical images, replacing the traditional subjective judgment relying on human experience. The model accurately identifies positively stained regions and quantifies staining intensity (0-3 points) and the percentage of positive cells (0-4 points) through deep learning, avoiding interpretation biases caused by experience differences or fatigue among different observers (including pathologists and non-professionals), as well as fluctuations in conclusions from the same observer at different times, thus significantly improving the consistency and objectivity of the interpretation results.

[0046] Secondly, through fully automated processing (including automatic cropping of WSI images, automatic model prediction, and automatic matching of intensity / percentage data with clinical information), the system eliminates the need for pathologists to manually verify each image point or for non-professionals to learn interpretation on the fly. It can quickly generate standardized interpretation results (such as staining intensity grading and positive percentage) for dozens to hundreds of chip points. Compared to the inefficient process of traditional manual interpretation, which requires comparing each slice point by point, this invention supports batch automated processing (such as simultaneous analysis of multiple WSI chips), significantly shortening the time researchers spend acquiring key data and solving the problem of work delays caused by reliance on pathologists. It is particularly suitable for the high-throughput screening needs of large-scale clinical samples.

[0047] Furthermore, considering the integrated nature of multi-chip points in tissue microarrays, a dedicated preprocessing workflow was designed for automatic cropping. For example, high-resolution WSI images were scaled down by a factor of 0.004 to reduce computational load. Otsu thresholding and morphological erosion / dilation were used to separate tissue from the background. After edge detection and mapping to the geometric center, chip points were precisely cropped. The image blocks were further divided into 1024×1024 pixel patches using an overlapping cropping strategy with a stride of 256. This approach ensured efficient segmentation of multiple chip points while preserving key features of edge cells through overlapping cropping, avoiding the loss of edge information caused by traditional coarse segmentation. Combined with data augmentation and image preprocessing during model training, the segmentation accuracy of complex samples was further improved, providing a more reliable raw data foundation for subsequent quantitative interpretation.

[0048] Furthermore, the model automatically outputs the staining intensity grade and percentage of positive cells for each chip point, and determines the row and column coordinates of the chip points through x / y axis clustering analysis. Combined with a pre-defined identifier system, it associates the chip points with pathology numbers, ultimately generating structured interpretation results that map one-to-one with clinicopathological features. This data can be directly used for statistical analysis (such as exploring the correlation between antibody expression levels and patient prognosis and treatment response), avoiding the tedious process of manually recording and organizing data required in traditional manual interpretation. Moreover, the continuous numerical variables (rather than binary classification results) provide richer dimensions for in-depth exploration of data correlations, significantly improving the reproducibility and analytical depth of research data, and contributing to the accurate discovery and validation of biomarkers.

[0049] In some preferred embodiments, the present invention further enhances the generalization ability of the interpretation model by incorporating multi-center tissue microarray data from laryngeal cancer, esophageal cancer, nasopharyngeal carcinoma, and other sources (using the SGD optimizer to train the model). This allows the interpretation model to adapt to complex situations such as staining differences and tissue morphology variations in samples from different sources. Combined with standardized data processing procedures, the comparability of cross-platform and cross-institutional data is further ensured, resolving the model bias problem caused by training on a single dataset and providing reliable technical support for large-scale multi-center studies. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0051] Figure 1 This is a flowchart illustrating an automatic interpretation method for tissue microarray immunohistochemistry results disclosed in this invention.

[0052] Figure 2 This is a schematic diagram of the SOLOv2 instance segmentation model structure disclosed in this invention;

[0053] Figure 3 This is a schematic diagram illustrating the WSI image of the tissue chip obtained according to the present invention.

[0054] Figure 4 This is a schematic diagram illustrating the use of Labelme software to annotate WSI images of tissue microarrays, as disclosed in this invention.

[0055] Figure 5 This shows the changing trends of the three types of losses during the training process of the SOLOv2 instance segmentation model disclosed in this invention.

[0056] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] It should be noted that the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0059] Example 1

[0060] See Figures 1-5 This invention provides a method for automatically interpreting tissue microarray immunohistochemistry results, the steps of which include:

[0061] S1. Obtain full-slide imaging of the tissue chip:

[0062] A digital pathology slide scanner was used to perform full-slide scanning of the tissue microarray, generating high-resolution whole-slide imaging (WSI). In this embodiment, tissue microarrays of laryngeal and esophageal cancer from the Second Xiangya Hospital of Central South University were selected, and the entire slides were scanned using a digital pathology slide scanner from Jiangfeng Biotechnology Co., Ltd. The scanned images were initially viewed and their quality confirmed using K-Viewer software to ensure clarity, completeness, and the absence of slide preparation defects. High-resolution WSI provides a high-quality raw data foundation for subsequent accurate segmentation and interpretation, avoiding misinterpretations due to image blurring or slide preparation issues, and ensuring the reliability of the interpretation results.

[0063] S2. Automatically crop the WSI image to segment individual pathological sample image blocks:

[0064] Because WSI images have extremely high resolution, the amount of data is huge, and direct processing is computationally complex. In this embodiment, the original image is first reduced to 0.004 times by a preset ratio to reduce the load. Then, the Otsu threshold segmentation algorithm is used to binarize the scaled image to separate the tissue region from the background. The threshold is preferably automatically determined by the grayscale histogram.

[0065] Further morphological manipulation (erosion and expansion) is used to remove noise and smooth tissue boundaries;

[0066] Then, the contour information of each pathological sample is accurately extracted by edge detection algorithms (such as the Canny operator), and its geometric center coordinates and bounding box size are calculated;

[0067] Finally, the aforementioned feature parameters are mapped back to the original high-resolution image, and precise cropping of individual pathological sample image blocks is completed at the corresponding locations.

[0068] The above process, taking into account the characteristics of multi-chip point integration in tissue microarrays, uses a combination strategy of downsampling, binarization, morphological optimization, and edge detection to efficiently extract high-quality individual pathological image blocks. This reduces computational complexity while ensuring cropping accuracy (geometric center mapping back to the original image avoids offset), providing clean data focused on the target region for subsequent model training and reducing the impact of background interference on interpretation.

[0069] S3. Pathologists annotate target areas and generate annotation data:

[0070] The cropped image blocks were annotated by at least two experienced pathologists. The outline of the target area was delineated using Labelme software. Annotation categories included positive staining intensity grades, such as P-, P+, P++, and Adjacent. For details, see [link to documentation]. Figure 4The areas enclosed by red lines correspond to P+ classification, green lines to Adjacent classification, yellow lines to P- classification, and dark blue lines to P++ classification. After annotation, the image patches are cut into 1024×1024 pixel patches and converted to the COCO format required for model training. The COCO format includes image information, target category, and pixel-level mask coordinates.

[0071] Doctor annotation combined with an overlapping cut strategy ensured the accuracy and completeness of the labeled data; the COCO format provided standardized input for model training, while the clear distinction between positive and negative categories laid the foundation for the subsequent model to learn the quantitative features of staining intensity and percentage, avoiding model confusion caused by non-standardized annotation.

[0072] S4. Perform data processing on the labeled image blocks:

[0073] The labeled image blocks undergo format conversion (unified to the model input format), image preprocessing (including top-bottom-hat transformation to enhance contrast and gamma transformation to adjust brightness distribution), data augmentation (including at least flipping, rotating, mirroring, translating, scaling, and filtering to simulate sample diversity), and data normalization (scaling pixel values ​​to a fixed range). At the same time, the labeled data is parsed and converted into COCO format (clearly defining the target category and mask coordinates).

[0074] Preprocessing and augmentation operations improve the model's adaptability to complex samples; normalization and COCO format conversion standardize the distribution of input data, ensuring the stability and generalization of model training and reducing interpretation bias caused by data differences.

[0075] S5. Training of the optimized SOLOv2 instance segmentation model:

[0076] The SOLOv2 instance segmentation model employed uses ResNet50 as the backbone and FPN as the neck. ResNet50 extracts multi-level convolutional features, while FPN effectively fuses features at different scales, thus balancing the representation of large tissue regions with fine cellular structures. Building upon this, transfer learning is introduced, initializing model parameters using pre-trained weights from existing immunohistochemical or other pathological image segmentation tasks. This accelerates convergence and reduces dependence on large-scale labeled data. Furthermore, dynamic convolution technology is combined, adaptively generating convolution kernel parameters in certain convolutional layers based on the input content. This enhances the model's feature capture capabilities when dealing with staining differences and tissue deformation, resulting in more accurate cell binary masks.

[0077] At the loss function level, considering the characteristics of the SOLOv2 instance segmentation task, the model's two-branch loss is defined as follows:

[0078] loss_cls corresponds to classification loss: Focal Loss is used to solve class imbalance (large difference in the proportion of positive / negative samples) and improve the classification accuracy of small targets / rare categories;

[0079] loss_mask corresponds to the segmentation loss: Dice Loss is used to optimize the region segmentation quality. If the task requires strong boundary constraints, Dice Loss can be supplemented with the gain logic for contour accuracy. However, SOLOv2 is originally based on mask IOU, so the classic configuration is preferred here.

[0080] The total loss is the weighted sum of loss_cls and loss_mask. The weights are optimized according to task requirements, and are either equal by default or dynamically allocated according to the magnitude of the loss.

[0081] For training configuration, the optimizer chosen was SGD (initial learning rate 0.01, momentum 0.9, weight decay = 0.0001, batch size = 16), and training was performed for 50 epochs (data was divided into training set 60%, validation set 20%, and test set 20% using the hold-out method to ensure independence between training, validation, and testing). In model convergence evaluation, besides the final mAP, the training set loss curve was the core observation indicator, helping to determine whether the amount of data supported effective learning: if the loss decreased rapidly and steadily, it indicated sufficient data; if the loss oscillated or did not decrease, the data needed to be expanded or the model adjusted.

[0082] Figure 5 This shows the changing trends of the three types of loss during the training of the SOLOv2 instance segmentation model. The horizontal axis represents the number of iterations (Iter), and the vertical axis represents the loss value.

[0083] The blue curve 'loss' represents the total loss, the green curve 'loss_mask' represents the region segmentation loss, and the orange curve 'loss_cls' represents the region classification loss.

[0084] As the iteration progresses, all three types of losses show a monotonically decreasing trend and tend to stabilize: the total loss decreases from the initial ~4.0 to ~0.2, and the segmentation loss_mask and classification loss_cls also converge from ~3.5 and ~0.5 to ~0.2 and ~0.1, respectively. The rapid decrease and stable convergence of the loss curves indicate that the model has efficiently learned the feature representations of classification and segmentation during training, without the risk of gradient vanishing or severe overfitting, and can be well adapted to the current instance segmentation task.

[0085] The SOLOv2 instance segmentation model generates target binary masks directly without bounding box detection, accurately delineating positively stained regions. Combining ResNet50 and FPN structures with dynamic convolution techniques, it achieves precise segmentation at the cell level. The combination of Dice Loss and Focal Loss optimizes segmentation boundaries and addresses class imbalance issues, while matrix NMS improves detection speed. Transfer learning utilizes existing data to initialize the model, reducing reliance on large-scale labeled data. This model training process ultimately outputs a high-precision interpretation model, providing a reliable tool for subsequent automatic prediction.

[0086] S6. Automatically predict staining intensity and percentage using an interpretation model:

[0087] The image blocks of each point on the tissue chip to be interpreted are input into the trained interpretation model. The interpretation model automatically identifies the positive staining areas and outputs the immunohistochemical staining intensity grade (0-3 points, corresponding to none / weak / medium / strong) and the percentage of positive cells for each chip point (0-4 points, 0% is recorded as 0, and the remaining 25% increments to 100% are recorded as 4 points).

[0088] The interpretation model directly outputs quantitative intensity and percentage data, replacing manual subjective grading and avoiding intensity / percentage misjudgments caused by experience differences in traditional semi-quantitative interpretation. The automated prediction process significantly improves efficiency, and the segmentation accuracy ensures the accuracy of intensity and percentage calculations, providing a reliable numerical basis for subsequent statistical analysis.

[0089] Specifically, the prediction results obtained through the aforementioned interpretation model are shown in Table 1. Table 1 displays the area distribution prediction results of the model at four classification levels: Adj, P-, P+, and P++. In Table 1, Label represents the area proportion of each category in the manually labeled (true value), Pred represents the corresponding area proportion predicted by the model, and Diff = |Pred-Label| is used to measure the area difference between the prediction result and the true label. The smaller the Diff value, the more consistent the model prediction is with the label; a larger difference indicates that the model has a bias in its discrimination of that category, where a positive bias can be understood as the model overestimating the area of ​​that category, and a negative bias indicates underestimation. Each table contains four columns representing the area proportion of different categories.

[0090] Table 1. Prediction Results of the Interpretation Model

[0091] ;

[0092] From the perspective of inter-sample consistency, the predictive performance varies among different samples. The model performs well on some samples but exhibits errors on others, suggesting that the prediction results may be affected by the complexity of tissue structure, differences in signal distribution, or sample heterogeneity. Overall, the model has good overall discriminative ability, but there is still room for further optimization in terms of extreme categories and inter-sample stability.

[0093] S7. Prediction results are automatically matched with clinical information to generate structured results:

[0094] The predicted staining intensity grade and percentage of positive cells are automatically matched with the corresponding clinical information to generate structured interpretation results.

[0095] Specifically, cluster analysis is first used to determine the row and column coordinates of the chip points in the tissue chip;

[0096] Furthermore, a unique identifier is generated based on the chip point coordinates, such as "chip name_antibody name_horizontal index_vertical index_coordinate X_coordinate Y_width_height". The original clinical data, such as patient age, gender, pathological diagnosis, and lymph node metastasis, is linked through the pathology number to finally generate a structured interpretation result that includes staining intensity grading, percentage of positive cells, and clinicopathological features.

[0097] Based on the obtained row and column coordinates, each chip point is uniquely named, and an image data identification system is established according to preset rules, so that the immunohistochemical interpretation results are mapped one-to-one with the clinical pathological features.

[0098] The preferred preset rule is:

[0099] Chip Name_Antibody Name_Horizontal Index_Vertical Index_Coordinate X_Coordinate Y_Width_Height

[0100] in:

[0101] Chip name: Identifies the origin and serial number of the chip;

[0102] Antibody name: Indicates the immunohistochemical marker being detected;

[0103] Horizontal / Vertical Index: Derived from the row and column numbers obtained from cluster analysis;

[0104] X / Y coordinates: These are the pixel coordinates of the geometric center of the clipping block in the original WSI.

[0105] Width / Height: The pixel dimensions of the cropped block in the original image.

[0106] This naming convention binds image data to the spatial location, detection indicators, and morphological features of chip points as a unique identifier, and can establish association fields with the patient's pathology number and clinical information (such as diagnosis, stage, treatment plan, follow-up results, etc.) in the database to achieve automatic matching.

[0107] The precise positioning of row and column coordinates ensures a strict correspondence between the interpretation result of each chip point and its original position. The association between the identification system and the pathology number realizes a data closed loop. The generated structured results can be directly imported into statistical software for correlation analysis, such as the correlation between antibody expression intensity and patient prognosis. This avoids the problems of error-prone and difficult-to-trace traditional manual recording, and significantly improves the reproducibility and analysis efficiency of scientific research data.

[0108] This invention addresses the core issues of inaccurate interpretation by non-pathologists (the model replaces subjective human judgment), low efficiency (full-process automation), and inconsistent results (standardized segmentation and quantitative output) through a tissue microarray-specific cropping process, physician annotation and overlapping cutting strategies, optimized SOLOv2 model training, quantitative prediction, and data loop matching. Furthermore, the processing flow designed for the multi-chip point integration characteristics of tissue microarrays, the dual loss function and dynamic convolution technology in model training, and the structured data generation scheme further ensure the accuracy, generalization, and research applicability of the interpretations, making it suitable for high-throughput screening of large-scale clinical samples and biomarker discovery.

[0109] In a preferred embodiment, in the automatic cropping process of step S2, the original WSI is reduced in size by a preset ratio, which significantly reduces the image data volume and facilitates real-time processing. In this embodiment, the Otsu thresholding algorithm is used to binarize the reduced image to automatically separate the tissue region from the background region. The specific implementation process is as follows:

[0110] 1. Grayscale conversion: The reduced color image is first converted into a grayscale image, and the RGB three-channel information is fused into a single-channel grayscale value to obtain the grayscale intensity range of each pixel.

[0111] 2. Calculate the grayscale histogram: Count the number of pixels at each grayscale level in the grayscale image to form a grayscale histogram. This histogram reflects the distribution of different grayscale values ​​in the image. Typically, the tissue region and the background region exhibit an approximately bimodal shape in their grayscale distribution: one peak corresponds to the background (lower or higher grayscale values, depending on staining and scanning conditions), and the other peak corresponds to the tissue region.

[0112] 3. Traversing all possible thresholds: The core of the Otsu algorithm is to find an optimal grayscale threshold T to divide pixels into two categories: foreground (organism) and background. The algorithm traverses all possible thresholds t (0~255) and calculates the intra-class variance of pixels for each category. For each t:

[0113] Pixels with gray values ​​≤ t are classified as background, and those > t are classified as foreground (or vice versa, depending on the actual histogram peak position);

[0114] Calculate the proportion of background pixels to total pixels Average gray level and the proportion of foreground pixels Average gray level ; Calculate the between-class variance .

[0115] 4. Select the optimal threshold: to minimize the inter-class variance. The maximum grayscale value T is the optimal threshold automatically determined by Otsu. This threshold maximizes the difference between the tissue and background classes, thus effectively separating them during binarization.

[0116] 5. Generate a binary image: Use the optimal threshold to binarize the grayscale image: pixels with grayscale values ​​greater than T are set as foreground (e.g., 1, representing tissue), and pixels with grayscale values ​​less than or equal to T are set as background (e.g., 0, representing blank or impurities). This results in a binary image containing only tissue and background regions.

[0117] The Otsu thresholding algorithm is used for binarization of downsized images without requiring manual threshold setting. It automatically selects the optimal segmentation point based on the grayscale distribution of the downsized image, adapting to different batches and staining conditions of WSI, reducing inconsistencies caused by human intervention. Global thresholding on the downsized image quickly yields clear tissue / background boundaries, providing a reliable baseline for subsequent morphological denoising and edge detection, reducing false positives and false negatives. Compared to directly calculating the Otsu threshold on the high-resolution original image, downsizing significantly reduces the number of pixels, drastically shortening the computation time for grayscale statistics and thresholding, meeting the real-time preprocessing requirements of large-size WSI.

[0118] In a preferred embodiment, during the automatic cropping process in step S2, after obtaining the binarized image through Otsu thresholding, isolated white dots or small black holes often appear in the image. These artifacts can affect the accuracy of subsequent contour extraction. Therefore, morphological operations are used to denoise and smooth the binary image, as follows:

[0119] 1. Erosion Operation: Select a structuring element of appropriate size, such as a 3×3 or 5×5 pixel circular or square kernel, and slide it pixel by pixel on the binary image. If there is a background pixel (value 0) within the coverage area of ​​the structuring element, the center pixel is set to background (0). This process causes the edges of the white target area (tissue) to shrink inward, which can remove isolated white noise with an area smaller than the structuring element and break the small connections formed by noise.

[0120] 2. Dilation Operation: Using a structuring element that is the same as or slightly larger than that used for erosion, dilation is performed on the eroded image. If there is a foreground pixel within the structuring element's coverage area, the center pixel is designated as the foreground. Dilation restores most of the original size of the shrunken tissue area and fills in small voids caused by erosion, smoothing jagged edges of the target.

[0121] 3. Opening Operation - Erosion followed by Dilation: In practical denoising, erosion and dilation are usually combined sequentially to form the opening operation. Its function is to effectively remove isolated noise points smaller than the structuring element and smooth the contour of the target while maintaining the basic shape and position of the main target. For binarized tissue chip images, this step can significantly reduce false targets caused by uneven staining or scanning noise, making the contours obtained by subsequent edge detection more continuous and closer to the real tissue boundaries.

[0122] As a preferred embodiment, in the model training and subsequent interpretation stage of step S5, the SOLOv2 instance segmentation model used in this invention can perform pixel-level target segmentation on immunohistochemistry images. The processing flow, combined with the structure shown in the attached figures, can be further described as follows:

[0123] Feature Extraction: The SOLOv2 instance segmentation model uses ResNet50 as the backbone and FPN as the neck to extract multi-scale features. The input image is first processed by a fully convolutional network (FCN) to extract features, resulting in a feature map I with uniform resolution. This feature map preserves the semantic and detailed information of the input image under different receptive fields, providing basic features for subsequent segmentation.

[0124] Grid partitioning: The feature map I is divided into S×S regular grids in spatial dimension (e.g., S=16 or 32, depending on the image resolution and chip dot size). Each grid is responsible for predicting the instance objects falling within it, thereby achieving location-aware instance segmentation.

[0125] Dual-branch structure and dynamic convolution: The Kernel Branch selects the corresponding convolutional kernel parameters based on the number and location of instances predicted from the current feature map. This branch learns to encode the category and location information of different instances into specific convolutional kernels for subsequent mask generation.

[0126] The Feature Branch further processes the feature map I to obtain feature sub-maps that correspond one-to-one with the convolution kernels output by the Kernel Branch.

[0127] By combining the two approaches, dynamic convolution calculations are performed at the corresponding grid positions: that is, the convolution kernel generated by the Kernel Branch for this instance is used to perform a convolution operation on the feature sub-map output by the Feature Branch, directly generating the binary mask for this instance. This dynamic matching method of "position-kernel-feature" enables the model to adaptively generate masks for different instances, avoiding the segmentation rigidity caused by fixed convolution kernels.

[0128] Matrix Nonmaximum Suppression Optimizes Detection Speed: In the post-processing stage, the Matrix Nonmaximum Suppression (MatrixNMS) algorithm is introduced to sort overlapping prediction masks by confidence and iteratively suppress redundant regions. This can significantly reduce computational overhead while maintaining segmentation accuracy and improve the response speed during batch interpretation.

[0129] The rich features extracted by FCN, combined with the positional constraints of the S×S grid, enable the model to accurately locate each stained region in immunohistochemical images. The dynamic convolution mechanism of the Kernel Branch and Feature Branch can adaptively generate masks based on the morphology and staining features of instances, significantly improving the segmentation accuracy of cellular and even subcellular structures. This network structure naturally supports the segmentation of instances of different sizes and staining intensities. With the addition of transfer learning and data augmentation strategies, it can stably process chip images from different batches and centers, reducing the segmentation performance degradation caused by staining differences or fluctuations in scanning conditions. The introduction of matrix NMS accelerates inference speed while ensuring mask quality, enabling the entire system to quickly output staining intensity grades and the percentage of positive cells in high-throughput scenarios. This avoids the inefficiency and subjective bias of manual point-by-point interpretation, effectively solving the problems of inaccurate interpretation, low efficiency, and inconsistent results by non-pathologists, and providing a reliable segmentation basis for subsequent automatic matching with clinical information.

[0130] In a preferred embodiment, in step S5, the training process of the model employs image enhancement methods including at least flipping, rotating, mirroring, translating, scaling, and filtering to expand the training samples.

[0131] Specifically, by using flip / mirror / rotation operations, the spatial orientation of the image can be changed, improving the model's robustness to changes in orientation;

[0132] By employing translation / scaling operations, the position and size changes of samples in the field of view can be simulated, avoiding the model's over-reliance on a fixed spatial distribution;

[0133] Filtering is employed, such as Gaussian filtering or mean filtering, to introduce slight blurring or noise, thereby improving the model's tolerance to uneven staining and fluctuations in scan quality.

[0134] These enhancements enable the model to learn richer morphological and textural variations during training, reducing performance degradation caused by differences in sample source, staining batch, or shooting conditions.

[0135] As a preferred embodiment, to achieve a strict correspondence between the immunohistochemical interpretation results of each chip point in the tissue chip and its original spatial location, and to provide a reliable positioning basis for subsequent automatic matching with clinical information, the method for determining the row and column coordinates of the chip points is further improved in step S7. The specific steps are as follows:

[0136] S711, Row and Column Clustering Analysis:

[0137] First, cluster analysis (using algorithms such as K-means or DBSCAN) is performed on the x-axis and y-axis coordinates of each chip point on the tissue chip. Based on the distribution density and spacing of the coordinate values, the spatial row and column arrangement pattern of the chip points is identified, determining that it is a two-dimensional array structure of n rows and m columns. This step can automatically adapt to tissue chips of different sizes and different arrangement densities, avoiding the errors and workload caused by manual row and column calibration.

[0138] Specifically, firstly, the position information of each chip is accurately extracted using segmentation algorithms; then, using data processing methods such as clustering, the coordinate features of each row and column of chips are statistically analyzed, and corresponding row and column reference lines are drawn in the graph accordingly.

[0139] The vertical lines are numbered as follows: from left to right, they are numbered 0, 1, 2...; the horizontal lines are numbered as follows: from top to bottom, they are numbered 0, 1, 2...

[0140] By observing the intersection of any vertical and horizontal line, the unique serial number of the corresponding chip in the array can be clearly identified, similar to coordinates in a Cartesian coordinate system. For example, the coordinates (0, 0) and (3, 2) can be marked, representing the position of a specific chip in the row and column dimensions, respectively: the horizontal coordinate corresponds to the vertical line number, and the vertical coordinate corresponds to the horizontal line number, thus achieving clear positioning and numbering management of each chip point.

[0141] This approach retains the original explanation while providing an intuitive analogy similar to a Cartesian coordinate system, helping readers quickly understand the algorithmic logic and coordinate definition rules behind the images.

[0142] S712, Geometric center matching rows and columns:

[0143] After obtaining the row and column clustering results, calculate the geometric center coordinates of each pathological sample image block (automatically cropped by step S2), and calculate the Euclidean distance between the center point and the center of each row and column obtained by the clustering. Take the row and column index corresponding to the minimum value to determine the specific row and column coordinates of the pathological sample in the chip.

[0144] By clustering the x and y coordinates separately, the spatial distribution structure of the chip points can be objectively reflected, avoiding row and column mismatches caused by manual observation or experience-based judgment. Combined with geometric center minimum distance matching, each cropped image block can be uniquely and accurately mapped to the physical location of the chip. The row and column coordinates are linked with the unique identifiers of the chip points generated subsequently, ensuring that when automatically matching staining intensity grading, percentage of positive cells, and clinicopathological information, the interpretation results strictly correspond to the original sample locations, avoiding data misalignment.

[0145] As a preferred embodiment, referring to Table 1, the structured interpretation results include the coordinate information, staining intensity grading, percentage of positive cells, and corresponding clinicopathological feature data for each image block.

[0146] In practice, the structured record of each image block includes the following parts:

[0147] Coordinate information: derived from the geometric center coordinates (X and Y coordinates) and the width and height of the cropped block obtained during step S2, and can be combined with the chip point row and column index determined in steps S711 and S712 to fully describe the position of the image block in the tissue chip and the original WSI.

[0148] Staining intensity grading: After the positive signal region is segmented by the SOLOv2 instance segmentation model in step S5, the intensity level (such as 0~3 or graded according to preset standards) is obtained by image grayscale or color space analysis, which reflects the expression intensity of the target protein in the tissue.

[0149] Percentage of positive cells: Based on the cell binary mask generated by the model and the analysis of staining intensity, the proportion of positive cells in the total number of cells in the image patch is calculated, and the expression level is quantified.

[0150] Corresponding clinical pathological feature data: By establishing naming rules and database association fields, the unique identifier of the image block is bound to the patient's pathology number, clinical diagnosis, stage, treatment plan, follow-up results and other information, so as to realize the automatic matching of interpretation results and clinical information.

[0151] The above four parts of data are stored in the system in a structured form (such as JSON, XML or relational database tables), and the fields are linked by unique identifiers to ensure that the information is complete, the location is accurate and easy to retrieve.

[0152] Table 2 shows the correlation between patient information and model prediction results.

[0153] ;

[0154] Table 2 illustrates that the prediction results are closely related to the degree of differentiation. In moderately differentiated squamous cell carcinoma, the model predictions are highly consistent with the manual annotations, with only slight confusion between adjacent expression levels. In well- to moderately differentiated squamous cell carcinoma, due to the complex tissue structure and high heterogeneity, some cases showed significant grade redistribution. In moderately to poorly differentiated and poorly differentiated squamous cell carcinoma, the overall trend is consistent, but the distinction between weak and moderate expression remains unstable. The consistency is best in atypical hyperplasia cases, and the model is the most robust.

[0155] Overall, the model can reproduce the area distribution of different expression levels well, exhibiting good consistency and repeatability, making it suitable as an auxiliary quantitative tool for pathologists, especially for large-sample analysis. However, manual verification is still required in cases with complex differentiation and high heterogeneity. Future training can improve the discrimination accuracy by adding samples from low-differentiation and boundary regions.

[0156] Coordinate information and row / column indexes allow each interpretation result to be traced back to a specific site on the original slide; staining intensity grading and the percentage of positive cells provide quantitative expression data; and clinicopathological features give the results medical significance. These four elements combine to form a closed-loop data chain. Structured storage allows interpretation results to be directly imported into statistical software or research platforms for cross-case, cross-site expression differential analysis or survival analysis. Furthermore, the data can be reviewed and verified at any time, improving the reproducibility of the research.

[0157] As a preferred embodiment, in order to improve the generalization ability of the SOLOv2 instance segmentation model to tissue microarray samples from different sources, in the model training stage of step S5, the model is further trained by merging multicenter tissue microarray data of at least laryngeal cancer, esophageal cancer and nasopharyngeal cancer, and the SGD optimizer is used to adjust the model parameters to improve the model's adaptability to samples from different sources.

[0158] In practice, immunohistochemically stained sections of the three types of tumor tissue microarrays from different medical institutions or pathology centers are collected and uniformly digitized and preprocessed to obtain a standardized training sample set. This sample set covers microarray images prepared by different staining batches, different scanning devices, and different operators, reflecting the sample diversity that may be encountered in actual clinical practice and scientific research. During training, these multi-center data are merged and input into the SOLOv2 instance segmentation model, allowing the model to be exposed to a wider range of staining intensity distributions, tissue structure morphology, and background noise features during the learning process. The combined training of multi-center and multi-disease data enables the model to learn the common features of staining habits, scanning conditions, and tissue morphology of different institutions, reducing overfitting caused by training with data from a single source and improving the segmentation and quantification accuracy of samples from unknown sources.

[0159] For parameter optimization, the SGD optimizer was selected, along with a suitable learning rate decay strategy. When dealing with large and diverse datasets, the SGD optimizer can suppress gradient oscillations caused by differences in sample sources by updating parameters in batches through random sampling, thus achieving a smoother convergence process and better generalization performance. Compared to some adaptive learning rate optimizers, the SGD optimizer often exhibits more robust adaptability in cases of small samples or distribution shifts, which is particularly important for joint training of multi-center, multi-disease data.

[0160] As a preferred embodiment, when constructing the training set for the automated immunohistochemistry interpretation model, due to subjective differences in manual annotation, the annotation results of two pathologists on the same image region are often not completely consistent. To solve this problem and ensure the annotation quality of the training data, a dual-pathologist annotation consistency verification method is added to the annotation quality control step S3, as follows:

[0161] Two doctors each annotate the same target area on an image, such as a cluster of positive cells or a stained area. The system records the coordinates of the annotated outlines and geometric center points of both doctors. During verification, the Euclidean distance between the two annotated center points is calculated first, and then the IoU (Intersection over Union) of the two annotated regions is calculated, which is the ratio of the number of pixels in the intersection to the number of pixels in the union of the two regions.

[0162] Set a passing threshold: If the center point distance is less than the preset value (e.g., 5-10 pixels, depending on the image resolution) and the IoU is greater than the preset value (e.g., 0.7 or higher), then the annotations of the two people are considered to be basically consistent, and the data of experienced doctors is selected for the training set; if the center point distance is too large or the IoU is too small, then the annotation consistency is deemed insufficient, and it needs to be reviewed and corrected by a third senior pathologist or the original annotator until the consistency requirements are met before it can be used.

[0163] By quantitatively evaluating the geometric overlap and positional deviation of the annotation results, samples with poor consistency are filtered out to avoid noise introduced by annotation differences, ensuring that the features learned by the model truly reflect the target morphology. The dual-doctor cross-verification mechanism can identify differences in individual annotation habits, and by reviewing and correcting to unify standards, the training data becomes more representative in terms of expression intensity and regional range.

[0164] Example 2

[0165] This invention also provides an automated interpretation system for tissue microarray immunohistochemistry results. Through modular design, it automates the entire process from image acquisition to structured interpretation result output. Each module works collaboratively to ensure the accuracy, traceability, and clinical applicability of the interpretation. Specifically, the system includes:

[0166] The image acquisition module is used to acquire whole slide imaging (WSI) of tissue microarrays. It supports digital pathology slide formats of common scanning devices (such as .svs, .ndpi, etc.) to ensure that high-resolution panoramic images are completely imported into the system, providing the raw data foundation for subsequent processing.

[0167] The automatic cropping module is used to segment WSI and extract image patches containing individual pathological samples. The automatic cropping module specifically includes:

[0168] The image scaling unit reduces the original high-resolution WSI by a preset ratio, thus reducing computational complexity.

[0169] The binarization processing unit uses the Otsu thresholding algorithm to binarize the reduced image and separate tissue from the background;

[0170] The noise removal unit combines morphological erosion and dilation operations to remove isolated noise points and smooth boundaries.

[0171] The contour extraction unit uses an edge detection algorithm to obtain the contour and geometric center coordinates of the pathological sample.

[0172] The precise cropping unit maps the contours and coordinates back to the original high-resolution image, completing the precise cropping of a single sample image patch.

[0173] The automatic cropping module improves cropping efficiency while preserving sample edge features through a coarse-to-fine processing flow, providing complete and accurately located image patches for subsequent annotation and model training.

[0174] The annotation module is used to manually annotate target regions (such as positively stained areas or cell clusters) in image patches, generating annotation data. To ensure annotation quality, the system can introduce a dual-doctor annotation consistency check: calculate the Euclidean distance and IoU between the center points of the two annotations; if the distance is too large or the IoU is too low, a review is triggered to ensure the reliability of the training data.

[0175] The data processing module performs format conversion, image preprocessing (such as color normalization), data augmentation (including at least flipping, rotating, mirroring, translating, scaling, and filtering) on ​​the labeled image patches, and data normalization. It also converts the labeled data to COCO format to meet the input requirements of instance segmentation models such as SOLOv2. Data augmentation improves the model's robustness to different coloring and morphological changes.

[0176] The model training module is based on the SOLOv2 instance segmentation model. The model uses ResNet50 as the backbone and FPN as the neck, leveraging multi-scale feature extraction and fusion to improve the ability to recognize targets of different sizes. During training:

[0177] Dice Loss is used as the contour regression loss function to enhance the accuracy of segmentation boundaries;

[0178] Focal Loss is used as the classification loss function to mitigate the impact of positive and negative sample imbalance on model performance;

[0179] The aforementioned data is used to enhance the generalization ability of the strategy.

[0180] In addition, the training data incorporates multi-center tissue microarray samples from at least laryngeal cancer, esophageal cancer, and nasopharyngeal cancer, and the SGD optimizer is used to adjust the parameters to improve the model's adaptability to samples from different P-, P+, P++, and Adjacent sources.

[0181] The prediction module uses a trained interpretation model to automatically predict image patches at various points on the tissue chip, outputting staining intensity grading and the percentage of positive cells. During the prediction phase, the model generates binary masks for cells through feature extraction, S×S grid partitioning, dynamic convolution of KernelBranch and Feature Branch, and introduces Matrix Non-Maximum Suppression (Matrix NMS) to improve detection speed.

[0182] The results matching module automatically matches the predicted results with corresponding clinical information to generate structured interpretation results. The specific process is as follows:

[0183] Cluster analysis was used to determine the row and column coordinates of chip points in the tissue chip;

[0184] Match row and column positions based on the coordinates of the geometric center point;

[0185] Each image block is named according to a preset rule (such as "chip name_antibody name_horizontal index_vertical index_coordinate X_coordinate Y_width_height") and associated with clinical pathological feature data with the pathology number;

[0186] The output is structured data containing coordinate information, staining intensity grading, percentage of positive cells, and clinicopathological features.

[0187] The system provided by this invention achieves full automation from whole-slide imaging to structured interpretation results through the collaborative work of modules including image acquisition, automatic cropping, annotation, data processing, model training, prediction, and result matching. The automatic cropping module extracts samples from coarse to fine and combines double-doctor annotation consistency verification to ensure data quality. The model training module uses multi-center data, composite loss functions, and diversified enhancements to improve adaptability and segmentation accuracy for samples from different sources. The result matching module uses spatial positioning and unique naming rules to accurately associate interpretation data with clinical information to form a closed-loop structure. This effectively overcomes the problems of low efficiency, large subjective differences, and difficulty in standardizing and managing results through traditional manual interpretation, significantly improving interpretation efficiency and accuracy, ensuring the traceability and analyzability of results, and providing an efficient, stable, and reliable technical solution for the clinical application and multi-center research of tissue microarray IHC interpretation.

[0188] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for automatically interpreting tissue microarray immunohistochemistry results, characterized in that, The steps include: S1. Obtain full-slide imaging of the tissue chip; S2. The whole slide image is automatically cropped to segment the image block including a single pathological sample. The automatic cropping step includes: first, the original WSI image is reduced by a preset ratio and then binarized by the Otsu threshold segmentation algorithm, and noise is removed by combining morphological operations. Then, the contour and geometric center coordinates of the pathological sample are extracted by the edge detection algorithm, and then mapped back to the original high-resolution image for accurate cropping. S3. Mark the target region in the cropped image block to generate annotation data, which includes category information of positive and negative regions. S4. Perform data processing on the labeled image blocks; S5. The processed image patch is trained based on the optimized SOLOv2 instance segmentation model. Dice Loss is used as the loss function for contour regression and Focal Loss is used as the classification loss function to obtain a well-trained interpretation model. When training the model, multi-center tissue chip data of at least laryngeal cancer, esophageal cancer and nasopharyngeal cancer are merged for training, and SGD optimizer is used to improve the model's adaptability to samples from different sources. S6. Use the trained interpretation model to automatically predict the image blocks at each point in the tissue chip, and output the immunohistochemical staining intensity grade and the percentage of positive cells for each image block. S7. Automatically match the predicted staining intensity grade and percentage of positive cells with the corresponding clinical information to generate structured interpretation results; the steps for automatically matching the predicted staining intensity grade and percentage of positive cells with the corresponding clinical information are as follows: S71. The row and column coordinates of chip points in the tissue chip are determined through cluster analysis. The specific steps are as follows: S711. First, cluster analysis is performed on the x-axis and y-axis coordinates of each chip point to determine the spatial distribution structure of the chip and obtain a two-dimensional array consisting of n rows and m columns. S712. Then, based on the geometric center point of each pathological sample, calculate the minimum distance between it and the center of the already clustered rows and columns, thereby determining the row and column coordinates of the corresponding pathological sample. S72. Name each chip point, establish an image data identification system according to preset rules, associate it with pathology numbers, and realize a one-to-one mapping between immunohistochemical interpretation results and clinical pathological features.

2. The method for automatic interpretation of tissue microarray immunohistochemistry results according to claim 1, characterized in that, The SOLOv2 instance segmentation model uses ResNet50 as the backbone and FPN as the neck, and combines transfer learning and dynamic convolution techniques to generate cell binary masks.

3. The method for automatic interpretation of tissue microarray immunohistochemistry results according to claim 1 or 2, characterized in that, In step S5, during model training, the SOLOv2 instance segmentation model is optimized using enhancement methods including at least flipping, rotating, mirroring, translating, scaling, and filtering.

4. The method for automatic interpretation of tissue microarray immunohistochemistry results according to claim 1 or 2, characterized in that, The structured interpretation results include the coordinate information, staining intensity grading, percentage of positive cells, and corresponding clinicopathological feature data for each image patch.

5. A method for automatic interpretation of tissue microarray immunohistochemistry results according to claim 1 or 2, characterized in that, To construct a training set for an automated immunohistochemistry interpretation model, at least two pathologists were required to annotate the same target region, which was either a cluster of positive cells or a stained area. The outline and geometric center coordinates of each annotation were recorded. Calculate the Euclidean distance between the center points of the two annotations and the intersection-union ratio (IoU) of the two annotation regions; Set a qualified threshold. If the Euclidean distance is less than a preset value and the IoU is greater than a preset value, it is determined that the annotations are basically consistent, and the annotation data of experienced doctors are selected to enter the training set. If the Euclidean distance is too large or the IoU is too small, the annotation consistency is deemed insufficient. The annotation will be reviewed and corrected by a third senior pathologist or the original annotator until the consistency requirements are met before it can be used.

6. An automated interpretation system for tissue microarray immunohistochemistry results, employing the method described in any one of claims 1-5, characterized in that, include: The image acquisition module is used to acquire full-slide images of the tissue chip; An automatic cropping module is used to segment the whole slide imaging and extract image blocks including individual pathological samples; The annotation module is used to annotate target regions in image blocks and generate annotation data; The data processing module is used to perform format conversion, image preprocessing, data augmentation, and data normalization on the labeled image blocks, and to convert the labeled data into COCO format; The model training module is used to train the processed image patches based on the SOLOv2 instance segmentation model and output the trained interpretation model. The prediction module is used to automatically predict image patches at various points in the tissue chip using a trained interpretation model, and output the staining intensity grade and the percentage of positive cells. The results matching module is used to automatically match the predicted results with the corresponding clinical information to generate structured interpretation results.

Images