Small Cancerous Lesion Anchor Point Warning System Based on Digital Pathology Images

Abstract

This invention discloses a small cancer lesion anchor point warning system based on digital pathological images, comprising: a feature extraction module, used to extract multi-level features from digital pathological images using a deep learning model to obtain feature extraction results; the feature extraction results include morphological features, texture features, and color features; an anchor point detection module, used to locate candidate regions of small cancer lesions based on the feature extraction results using a target detection model to obtain anchor point detection results; an optimization module, used to optimize the anchor point detection results, including merging, deduplication, and classification of candidate regions; and a generation module, used to determine the anchor point region matrix coordinate information based on the digital pathological image coordinate system to generate a list of small cancer lesion anchor point warning islands. This system enables rapid location and marking of small cancer lesions in pathological sections, automatically analyzes the section images, identifies and captures cancer lesion anchor points, and displays them in the form of warning islands, improving the efficiency and accuracy of pathological analysis.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a small cancerous lesion anchor point warning system based on digital pathological images. Background Technology

[0002] Digital pathology technology is a crucial tool for pathological diagnosis in modern medicine. Traditional pathology relies on manual microscopic observation and analysis, which is easily influenced by the observer's subjective experience and fatigue, and is difficult to standardize. However, digital pathology, by converting slides into high-resolution images and combining them with artificial intelligence (AI) algorithms, can achieve rapid, efficient, and accurate pathological diagnosis, particularly excelling in cancer screening and lesion identification. Small cancerous lesions refer to early-stage cancers or tiny lesion areas, which are often difficult to accurately identify with the naked eye or traditional imaging techniques. Efficiently and accurately capturing small cancerous lesions is essential for early diagnosis and treatment. However, due to the high resolution and large data volume of digital pathology images, existing image processing and analysis methods face many challenges, such as the complexity, diversity, and heterogeneity of lesion regions, as well as high annotation costs.

[0003] Manual annotation and traditional image processing methods are currently one of the fundamental approaches to digital pathology image analysis. This involves experienced pathologists manually annotating pathology images, combined with traditional image processing techniques such as edge detection and region segmentation for analysis. While this method was feasible in its early stages, it has significant limitations. Manual annotation is time-consuming and labor-intensive, and easily influenced by the observer's subjective experience, leading to inconsistent annotation quality. Traditional image processing techniques are poorly adaptable to noise and the diversity of small cancerous lesions in pathology images, failing to meet the demand for accurate identification of complex lesion regions. Furthermore, traditional methods are inefficient when processing large-scale pathology image data, making it difficult to meet the high efficiency and accuracy requirements of modern pathology diagnosis. The existing technologies suffer from the following technical problems: 1. Accuracy in identifying small cancerous lesions: Traditional methods are poorly adaptable to the complexity, diversity, and heterogeneity of small cancerous lesion regions, making it difficult to accurately capture small cancerous lesions. 2. High-resolution pathology image data processing issues: Pathology image data is massive, and existing methods are inefficient when processing high-resolution images. 3. Intelligent automatic capture of lesion regions: Existing technologies rely heavily on manual annotation, have low automation levels, and fail to meet efficiency requirements. 4. Model training challenges with limited labeled data: Obtaining labeled data for small cancer lesions is difficult, and existing deep learning methods are highly dependent on large-scale labeled data, making it difficult to handle insufficient samples. 5. Challenges in multimodal data fusion and registration: Current techniques suffer from data mismatch and inconsistent standards in the fusion analysis of pathological images and other modalities. 6. Equipment dependence and computational resource consumption: Existing solutions require high-performance computing equipment, increasing the cost and complexity of practical applications. 7. Noise interference and data quality issues: Pathological images may contain artifacts, noise, and other interference, affecting the accuracy of anchor point capture. Summary of the Invention

[0004] This invention aims to at least partially solve one of the technical problems in the aforementioned technologies. Therefore, the purpose of this invention is to propose a small cancer lesion anchor point warning system based on digital pathological images, enabling rapid location and marking of small cancer lesions in pathological sections. It can automatically analyze section images, identify and capture cancer lesion anchor points, and display them in the form of warning islands, thereby improving the efficiency and accuracy of pathological analysis.

[0005] To achieve the above objectives, embodiments of the present invention propose a small cancerous lesion anchor point warning system based on digital pathological images, comprising: The feature extraction module is used to extract multi-level features from digital pathological images using a deep learning model, and obtain feature extraction results; the feature extraction results include morphological features, texture features and color features. The anchor point detection module is used to locate candidate regions of small cancer lesions based on feature extraction results and through the target detection model to obtain anchor point detection results; The optimization module is used to optimize the anchor point detection results, including merging candidate regions, deduplication, and classification. The generation module is used to determine the coordinate information of the anchor point region matrix based on the digital pathological image coordinate system and generate a list of small cancerous lesion anchor point warning islands.

[0006] According to some embodiments of the present invention, an image preprocessing module is also included, which is used to preprocess the digital pathological image before the feature extraction module extracts multi-level features from the digital pathological image using a deep learning model, including denoising, artifact removal and image normalization.

[0007] According to some embodiments of the present invention, the feature extraction module includes: The first extraction module is used to learn the spatial relationships in the image based on the structure of the convolutional neural network using a deep learning model, and extract morphological features; the morphological features include the shape, size, and boundaries of objects in the image. The second extraction module is used to extract texture features based on a deep learning model; the texture features describe the patterns or arrangement rules of local regions in the image; The third extraction module is used to extract color features based on a deep learning model through color space conversion and color histogram statistics; the color features are the color distribution and changes of pixels in the image.

[0008] According to some embodiments of the present invention, the anchor point detection module includes: The first training module is used to train the object detection model using incomplete labeled data based on weakly supervised learning techniques, and to make up for the lack of labeled data by introducing additional constraints or assumptions. The second training module is used to expand the training set using unlabeled images based on semi-supervised learning techniques, and to guide the labeling process of unlabeled images by constructing pseudo-labels or using the prediction results of object detection models. The localization module is used to locate candidate regions for small cancerous lesions based on feature extraction results using a trained target detection model, and obtain anchor point detection results.

[0009] According to some embodiments of the present invention, the optimization module includes: The merging and deduplication module is used to merge and deduplicate candidate regions based on the nonmaximum suppression algorithm. The regional hierarchical screening module is used to set different screening criteria or thresholds, and divide candidate regions into different levels or categories based on the first classification model. A classification model calibration module is used to adjust the output of the first classification model based on a calibration method; the calibration method includes at least one of Platt calibration, Isotonic regression calibration, histogram calibration, and neural network calibration.

[0010] According to some embodiments of the present invention, the optimization module includes: The merging module is used to calculate the overlapping area of ​​each pair of candidate regions and compare it with a preset overlapping area threshold. Pairs of candidate regions with an overlapping area greater than the preset overlapping area threshold are merged into a new region. The bounding box of the new region is the smallest bounding rectangle of all merged regions. The deduplication module is used to calculate the shape similarity of each pair of candidate regions and compare it with a preset shape similarity threshold. When the shape similarity is determined to be greater than the preset shape similarity threshold, one of the candidate regions in the pair is removed. The classification module is used to extract the feature vector of each candidate region, input it into the pre-trained second classification model, and output the classification result.

[0011] According to some embodiments of the present invention, the generation module includes: The conversion module is used to convert the position of the cancerous lesion region in the digital pathology image into coordinate information in the anchor point matrix based on the characteristics of the cancerous lesion region in the coordinate system of the digital pathology image and the coordinate transformation method. The characteristics of the cancerous lesion region include area, perimeter and position. The viewport scaling module is used to record the scaling ratio and position information of each anchor point region while generating anchor point matrix coordinate information.

[0012] According to some embodiments of the present invention, it further includes: The multimodal data fusion module is used to register and fuse the anchor point detection results with imaging modal data or genomic modal data based on the deep registration algorithm after the anchor point detection module obtains the anchor point detection results, so as to obtain the processed anchor point detection results. The optimization module is used to optimize the processed anchor point detection results.

[0013] According to some embodiments of the present invention, it further includes: an adjustment module, used to compare the digital pathological image with a location standard pathological image before the feature extraction module extracts multi-level features from the digital pathological image using a deep learning model, and adjust the recognition angle of the digital pathological image according to the comparison result to obtain an adjusted digital pathological image.

[0014] This invention proposes a small cancer lesion anchor point warning system based on digital pathological images, improving the accuracy and stability of small cancer lesion identification and reducing the rates of missed and false detections. It optimizes the data processing workflow to achieve rapid analysis and anchor point capture, meeting real-time clinical needs. It reduces reliance on manual intervention, achieving automated lesion area capture. By combining weakly supervised learning with semi-supervised learning methods, it effectively utilizes limited labeled data to improve the system's learning and recognition capabilities. Optimized multimodal fusion technology enhances the comprehensiveness and accuracy of anchor point capture. Algorithm optimization and lightweight design reduce the system's dependence on hardware, improving its versatility and deployment flexibility. Image preprocessing and feature enhancement techniques improve data quality and enhance the robustness of the recognition results.

[0015] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a block diagram of a small cancerous lesion anchor point warning system based on digital pathology images according to an embodiment of the present invention; Figure 2 This is a block diagram of a feature extraction module according to an embodiment of the present invention; Figure 3 This is a block diagram of an anchor point detection module according to an embodiment of the present invention. Detailed Implementation

[0018] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0019] like Figure 1 As shown, this embodiment of the invention proposes a small cancerous lesion anchor point warning system based on digital pathological images, comprising: The feature extraction module is used to extract multi-level features from digital pathological images using a deep learning model, and obtain feature extraction results; the feature extraction results include morphological features, texture features and color features. The anchor point detection module is used to locate candidate regions of small cancer lesions based on feature extraction results and through the target detection model to obtain anchor point detection results; The optimization module is used to optimize the anchor point detection results, including merging candidate regions, deduplication, and classification. The generation module is used to determine the coordinate information of the anchor point region matrix based on the digital pathological image coordinate system and generate a list of small cancerous lesion anchor point warning islands.

[0020] The working principle of the above technical solution is as follows: Deep learning models are used to extract multi-level features from digital pathology images. These features include, but are not limited to, morphological features of cancer lesions (such as shape, size, and edges), texture features (such as image roughness and directionality), and color features (such as hue, saturation, and brightness). Based on the feature extraction results, the anchor point detection module uses target detection models (such as YOLO, Faster R-CNN, etc.) to locate candidate regions of small cancer lesions in the image. These candidate regions are parts that may contain cancer lesions. The anchor point detection results are optimized, mainly including merging candidate regions (merging adjacent and similar regions into a larger region), deduplication (removing duplicate detected regions), and classification (further subdividing the type or severity of cancer lesions based on features). Optimization reduces false positives and false negatives, improving the system's recognition accuracy and practicality. Based on the coordinate system of the digital pathology image, the generation module determines the matrix coordinate information of each anchor point region and generates a list of small cancer lesion anchor point warning islands. This list contains information such as the location, size, and feature description of each warning island. It visually demonstrates the location and characteristics of potential small cancerous lesions in the images, providing important reference for doctors' further diagnosis.

[0021] The beneficial effects of the above technical solution are as follows: The small cancer lesion anchor point warning system based on digital pathological images achieves automatic identification and warning of small cancer lesions through deep learning, target detection and optimization processing, and realizes rapid positioning and marking of small cancer lesions in pathological sections. It can automatically analyze the section images, identify and capture cancer lesion anchor points and display them in the form of warning islands, thereby improving the efficiency and accuracy of pathological analysis.

[0022] According to some embodiments of the present invention, an image preprocessing module is also included, which is used to preprocess the digital pathological image before the feature extraction module extracts multi-level features from the digital pathological image using a deep learning model, including denoising, artifact removal and image normalization.

[0023] The working principle of the above technical solutions is as follows: Noise Reduction: Eliminating noise in the image, such as random noise and speckle noise, to improve image clarity and readability. Artifact Removal: Identifying and removing artifacts in the image. These artifacts may be caused by interference during scanning, transmission, or storage, and they affect image quality and diagnostic accuracy. Image Normalization: Adjusting the image's brightness, contrast, and other attributes to conform to specific standards or ranges. Normalization helps eliminate differences between different images, improving system consistency and stability.

[0024] The beneficial effects of the above technical solution are as follows: Preprocessed images are clearer and more accurate, providing better input for the feature extraction module. Deep learning models can better learn useful features in images during training, thereby improving the system's recognition ability. Since the image preprocessing module improves image quality, this will positively impact subsequent anchor point detection, optimization, and generation modules. More accurate images will reduce the possibility of false positives and false negatives, improving the overall performance of the system.

[0025] like Figure 2 As shown, according to some embodiments of the present invention, the feature extraction module includes: The first extraction module is used to learn the spatial relationships in the image based on the structure of the convolutional neural network using a deep learning model, and extract morphological features; the morphological features include the shape, size, and boundaries of objects in the image. The second extraction module is used to extract texture features based on a deep learning model; the texture features describe the patterns or arrangement rules of local regions in the image; The third extraction module is used to extract color features based on a deep learning model through color space conversion and color histogram statistics; the color features are the color distribution and changes of pixels in the image.

[0026] The working principle of the above technical solution is as follows: The first extraction module utilizes the Convolutional Neural Network (CNN) structure in a deep learning model to learn the spatial relationships in an image and extract morphological features. These features describe key information such as the shape, size, and boundaries of objects in the image. By training the CNN model, it is able to recognize and extract key morphological information from the image. The convolutional layers, pooling layers, and fully connected layers in the CNN work together to capture and extract features at different scales and levels in the image. Morphological features can provide doctors with an intuitive display of key information such as the shape and size of lesion areas, helping them to make more accurate diagnoses.

[0027] The second extraction module extracts texture features from the image based on a deep learning model. These features describe the patterns or arrangements of local regions in the image, such as the thickness, direction, and density of the texture. Deep learning models (such as CNNs and ResNets) can learn texture information in images and extract and represent it through specific network structures (such as convolutional layers and pooling layers). Furthermore, texture analysis methods (such as Gray-Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP)) can be combined for more in-depth texture feature extraction. Texture features can reflect the microstructure and tissue changes of lesion areas, providing doctors with richer diagnostic information.

[0028] The third extraction module utilizes deep learning models to extract color features from images through methods such as color space transformation and color histogram statistics. These features describe the color distribution and variations of pixels in the image, such as hue, saturation, and brightness. The image is converted from the RGB color space to other color spaces (such as HSV, Lab, etc.) to better capture color information. Then, deep learning models (such as convolutional neural networks CNN) or statistical methods (such as color histograms) are used to extract and represent the color features. Color features can provide doctors with intuitive information about color changes in lesion areas, helping them determine the type, severity, and possible pathological processes of lesions.

[0029] The beneficial effects of the above technical solution are: by using deep learning models to extract multi-level and multi-dimensional features (including morphological features, texture features and color features) from images, it can provide doctors with more accurate and comprehensive information, which helps to improve the accuracy and efficiency of doctors' pathological diagnosis.

[0030] like Figure 3 As shown, according to some embodiments of the present invention, the anchor point detection module includes: The first training module is used to train the object detection model using incomplete labeled data based on weakly supervised learning techniques, and to make up for the lack of labeled data by introducing additional constraints or assumptions. The second training module is used to expand the training set using unlabeled images based on semi-supervised learning techniques, and to guide the labeling process of unlabeled images by constructing pseudo-labels or using the prediction results of object detection models. The localization module is used to locate candidate regions for small cancerous lesions based on feature extraction results using a trained target detection model, and obtain anchor point detection results.

[0031] The working principle of the above technical solutions is as follows: Weakly supervised learning techniques include image-level labeling (rather than precise annotation at the bounding box level) and the use of the Multiple Instance Learning (MIL) framework. The model learns to extract useful information from images containing target objects and gradually improves its localization ability. Weakly supervised learning reduces the dependence on large amounts of precisely labeled data, enabling the model to learn and train effectively even with limited annotation resources.

[0032] Semi-supervised learning techniques expand the training set with unlabeled images, thereby improving the generalization ability of object detection models. Semi-supervised learning combines the advantages of labeled and unlabeled data, guiding the labeling process of unlabeled images by constructing pseudo-labels or utilizing the model's predictions. Semi-supervised learning includes self-training methods, where the model is first trained on labeled data, then used to label unlabeled data using these predictions, and the model is iteratively updated; or it uses consistency regularization or adversarial training to enhance the model's robustness. Semi-supervised learning can fully utilize unlabeled data, improving the model's generalization performance and robustness, especially when labeled data is scarce.

[0033] The localization module receives feature maps or feature vectors from the feature extraction module and applies a trained object detection model (such as Faster R-CNN, YOLO, etc.) to predict the location and size of small cancerous lesions. These predictions constitute anchor point detection results, i.e., candidate regions in the image that may contain small cancerous lesions. The localization module transforms the predictive power of the deep learning model into specific, interpretable auxiliary diagnostic information, providing doctors with intuitive indications of lesion location.

[0034] The beneficial effects of the above technical solution are as follows: The anchor point detection module, by combining weakly supervised learning, semi-supervised learning, and object detection techniques, achieves accurate localization of candidate regions for small cancerous lesions under limited labeled data conditions. This integrated approach not only improves the accuracy and robustness of the object detection model but also reduces the dependence on large amounts of precisely labeled data, providing strong support for digital pathological image analysis.

[0035] According to some embodiments of the present invention, the optimization module includes: The merging and deduplication module is used to merge and deduplicate candidate regions based on the nonmaximum suppression algorithm. The regional hierarchical screening module is used to set different screening criteria or thresholds, and divide candidate regions into different levels or categories based on the first classification model. A classification model calibration module is used to adjust the output of the first classification model based on a calibration method; the calibration method includes at least one of Platt calibration, Isotonic regression calibration, histogram calibration, and neural network calibration.

[0036] The working principle of the above technical solution is as follows: The merging and deduplication module is used to process overlapping or adjacent candidate regions output by the object detection model. Non-Maximum Suppression (NMS) algorithm is used for merging and deduplication. The basic idea of ​​the NMS algorithm is to select the candidate region with the highest score and suppress other candidate regions with high overlap around it. The specific implementation steps include calculating the Intersection over Union (IoU) ratio between candidate regions, and then deciding whether to suppress a candidate region based on a set IoU threshold. Merging and deduplication reduces redundant candidate regions, improves the accuracy and efficiency of object detection, and provides more accurate input for subsequent region classification and model correction.

[0037] The regional grading screening module allows for the setting of different screening criteria or thresholds, classifying candidate regions based on their characteristics (such as size, shape, and color) and the prediction scores of the classification model. These screening criteria can be customized to suit different application scenarios. Regional grading screening helps doctors identify potential lesion areas more quickly and prioritize them according to their importance or risk level, thereby improving the efficiency and accuracy of diagnosis.

[0038] The classification model calibration module can employ various calibration methods, including Platt calibration, Isotonic regression calibration, histogram calibration, and neural network calibration. The core idea of ​​these methods is to utilize additional validation data to adjust the output probabilities or scores of the classification model, making them closer to the true class distribution. Classification model calibration can improve the performance of classification models, especially when the dataset is imbalanced or noisy. Through calibration, the accuracy and reliability of object detection can be further improved, providing doctors with more accurate auxiliary diagnostic information.

[0039] The beneficial effects of the above technical solution are as follows: The optimization module is a multifunctional component that integrates deduplication, regional hierarchical screening, and classification model correction. By processing candidate regions after target detection, it reduces redundant information, improves classification accuracy, and provides doctors with more accurate and reliable auxiliary diagnostic information, thus helping to improve the performance and efficiency of digital pathology image analysis.

[0040] According to some embodiments of the present invention, the optimization module includes: The merging module is used to calculate the overlapping area of ​​each pair of candidate regions and compare it with a preset overlapping area threshold. Pairs of candidate regions with overlapping areas greater than the preset overlapping area threshold are merged into a new region. The bounding box of the new region is the smallest bounding rectangle of all merged regions. The deduplication module is used to calculate the shape similarity of each pair of candidate regions and compare it with a preset shape similarity threshold. When the shape similarity is determined to be greater than the preset shape similarity threshold, one of the candidate regions in the pair is removed. The classification module is used to extract the feature vector of each candidate region, input it into the pre-trained second classification model, and output the classification result.

[0041] The working principle of the above technical solution is as follows: The merging module first calculates the overlap area of ​​each pair of candidate regions, which is usually achieved by calculating the intersection area of ​​the bounding boxes of the two candidate regions. Then, these overlap areas are compared with a preset overlap area threshold. If the overlap area of ​​a pair of candidate regions is greater than the threshold, the two candidate regions are considered to be merged. The bounding box of the merged new region is set as the smallest bounding rectangle of all merged regions to ensure that the new region can completely cover the original candidate regions. By merging overlapping candidate regions, redundant information in target detection can be reduced, improving the accuracy and reliability of the detection results.

[0042] The deduplication module first calculates the shape similarity of each pair of candidate regions. This is achieved by comparing features such as the bounding box shape, area, and aspect ratio of the two candidate regions. Then, these shape similarities are compared to a preset shape similarity threshold. If the shape similarity of a pair of candidate regions is greater than the threshold, the two candidate regions are considered similar, and one of them should be removed. In practice, candidate regions with higher scores or more accurate locations can be selected and retained as needed. By removing candidate regions with similar shapes, duplicate detections in object detection can be reduced, improving the accuracy and efficiency of the detection results.

[0043] The classification module first extracts feature vectors for each candidate region. These feature vectors can include various information such as color, texture, and shape. Then, these feature vectors are input into a pre-trained secondary classification model. This model can be based on deep learning methods, such as convolutional neural networks (CNNs). The classification model outputs a classification result based on the input feature vectors, indicating whether each candidate region belongs to the target category. Through the classification module, candidate regions can be accurately classified, thereby filtering out the true target regions.

[0044] The beneficial effects of the above technical solution are as follows: The optimization module is a component that processes candidate regions after target detection. It includes three key modules: merging, deduplication, and classification. These modules improve the accuracy and efficiency of target detection by reducing redundant information, removing duplicate detections, and accurately classifying candidate regions, thus contributing to the performance improvement of digital pathology image analysis.

[0045] According to some embodiments of the present invention, the generation module includes: The conversion module is used to convert the position of the cancerous lesion region in the digital pathology image into coordinate information in the anchor point matrix based on the characteristics of the cancerous lesion region in the coordinate system of the digital pathology image and the coordinate transformation method. The characteristics of the cancerous lesion region include area, perimeter and position. The viewport scaling module is used to record the scaling ratio and position information of each anchor point region while generating anchor point matrix coordinate information.

[0046] The working principle of the above technical solution is as follows: The conversion module extracts features of the cancerous lesion region in the digital pathology image, including key information such as area, perimeter, and location. Then, based on these features and predefined coordinate transformation methods (such as affine transformation and projection transformation), the module calculates the corresponding coordinates of the cancerous lesion region in the anchor point matrix. This conversion process ensures that the cancerous lesion region in the digital pathology image can be accurately mapped to the anchor point matrix, thereby facilitating subsequent processing and analysis.

[0047] When the conversion module transforms the cancerous lesion region into coordinates in the anchor point matrix, the viewport scaling module simultaneously records the scaling ratio of each anchor point region relative to the original digital pathology image. Furthermore, the module records the specific location information of each anchor point region within the digital pathology image, which facilitates backtracking to the original image for further analysis or verification when needed. The viewport scaling module ensures that the coordinate information in the anchor point matrix not only has precise mathematical meaning but also maintains a consistent physical meaning with the original digital pathology image.

[0048] The beneficial effects of the above technical solution are as follows: The generation module is a complex component integrating a conversion module and a viewport scaling module. It extracts features of cancerous lesions in digital pathology images and converts them into coordinate information in an anchor point matrix, while simultaneously recording the scaling ratio and position information of each anchor point region. This comprehensive processing method enables the effective digitization and quantification of cancerous lesions in digital pathology images, facilitating data analysis.

[0049] According to some embodiments of the present invention, it further includes: The multimodal data fusion module is used to register and fuse the anchor point detection results with imaging modal data or genomic modal data based on the deep registration algorithm after the anchor point detection module obtains the anchor point detection results, so as to obtain the processed anchor point detection results. The optimization module is used to optimize the processed anchor point detection results.

[0050] The working principle of the above technical solution is as follows: the deep registration algorithm can automatically learn the correspondence between different modalities of data, thereby achieving accurate image registration. After registration is completed, the module fuses the anchor point detection results with the registered radiological or genomic data. The significance of the multimodal data fusion module lies in its ability to integrate data from different modalities, providing more comprehensive and accurate information.

[0051] The beneficial effects of the above technical solution are as follows: the multimodal data fusion module can integrate data from different modalities, providing more comprehensive and accurate information; while the optimization module can optimize the processed anchor point detection results, improving their accuracy and reliability. The combined effect of these two modules will enable the system to provide more accurate and reliable pathological analysis.

[0052] In one embodiment, model quantization, pruning techniques, and edge computing optimization improve the system's versatility and real-time performance in practical applications. Through algorithm optimization and model compression, the system's hardware dependence is reduced, achieving a lightweight design suitable for various terminal devices.

[0053] According to some embodiments of the present invention, it further includes: an adjustment module, used to compare the digital pathological image with a location standard pathological image before the feature extraction module extracts multi-level features from the digital pathological image using a deep learning model, and adjust the recognition angle of the digital pathological image according to the comparison result to obtain an adjusted digital pathological image.

[0054] The working principle and beneficial effects of the above technical solution are as follows: The location standard pathological image serves as the optimal recognition angle for the feature extraction module when using a deep learning model for feature extraction. Before the feature extraction module extracts multi-level features from the digital pathological image using the deep learning model, the digital pathological image is compared with the location standard pathological image. Based on the comparison results, the recognition angle of the digital pathological image is adjusted to ensure consistency with the location standard pathological image. By adjusting the recognition angle of the digital pathological image through the adjustment module, the accuracy of feature extraction by the deep learning model can be ensured. This reduces the difficulty of subsequent feature extraction and processing, thereby improving the overall processing efficiency.

[0055] According to some embodiments of the present invention, the adjustment module includes: The segmentation module is used to segment digital pathology images into several sub-digital pathology images; and to segment location standard pathology images into several sub-location standard pathology images; the sub-digital pathology images and the sub-location standard pathology images are of equal size; The matching module is used for: One sub-digital pathology image is randomly selected from several sub-digital pathology images and used as a reference sub-digital pathology image; Calculate the first matching degree between the reference sub-digital pathological image and several sub-location standard pathological images respectively; ; in, The first matching degree between the reference sub-digital pathology image and the standard pathology image at the i-th sub-location; The length of the image pixels in the reference digital pathology image; The width of the image pixels for reference sub-digital pathology images; The value of the pixel in the s-th row and t-th column of the reference digital pathology image; The average pixel value of the reference digital pathology image; Let be the value of the pixel in the s-th row and t-th column of the standard pathological image of the i-th sub-location; Let be the mean value of the pixels of the standard pathological image at the i-th sub-location, where i = 1, 2, 3...M, and M is the number of standard pathological images at the sub-location. The standard pathological image of the sub-position with the highest first matching degree is selected as the target sub-position standard pathological image; the reference sub-digital pathological image is rotated at different angles to obtain several adjusted images; the second matching degree between the several adjusted images and the target sub-position standard pathological image is calculated; the rotation angle of the adjusted image with the highest second matching degree is selected as the adjustment angle. Calculate the coordinate adjustment amount of the reference sub-digital pathology image; ; ; in, This is the horizontal coordinate adjustment amount. This is the vertical coordinate adjustment amount. To determine the row number of the reference sub-digital pathology image's midpoint within the overall digital pathology image, To determine the column number of the reference sub-digital pathology image's midpoint within the overall digital pathology image, This represents the row number of the midpoint of the sublocation standard pathology image within the location standard pathology image. This represents the column number of the midpoint of the sub-location standard pathology image within the location standard pathology image; The digital pathology image is adjusted by adjusting the angle and coordinates to obtain the adjusted digital pathology image. ; ; in, and To determine the number of rows and columns of the pixels to be adjusted in the digital pathology image, and This represents the row and column number of the pixel to be adjusted in the adjusted digital pathology image; To adjust the angle.

[0056] The working principle of the above technical solution is as follows: A segmentation module is used to divide the digital pathology image into several sub-digital pathology images. Simultaneously, the location standard pathology image is divided into several sub-location standard pathology images of the same size as the sub-digital pathology images. One sub-digital pathology image is randomly selected from these sub-digital pathology images as a reference sub-digital pathology image. The first matching degree between the reference sub-digital pathology image and each sub-location standard pathology image is calculated, and the sub-location standard pathology image with the highest first matching degree is selected as the target sub-location standard pathology image. The reference sub-digital pathology image is rotated at different angles to obtain several adjusted images. The second matching degree between each adjusted image and the target sub-location standard pathology image is calculated. The rotation angle corresponding to the adjusted image with the highest second matching degree is selected as the final adjustment angle, and the digital pathology image is adjusted according to the adjustment angle and coordinate adjustment amount.

[0057] The beneficial effects of the above technical solution are as follows: the adjustment module can precisely adjust the angle and position of the digital pathology image to achieve optimal alignment with the location standard pathology image, making the deep learning model more accurate in feature extraction. This reduces the difficulty of subsequent feature extraction and processing, thereby improving overall processing efficiency.

[0058] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A small cancerous lesion anchor point warning system based on digital pathological images, characterized in that, include: The feature extraction module is used to extract multi-level features from digital pathological images using a deep learning model to obtain feature extraction results. The feature extraction results include morphological features, texture features, and color features; The anchor point detection module is used to locate candidate regions of small cancer lesions based on feature extraction results and through the target detection model to obtain anchor point detection results; The optimization module is used to optimize the anchor point detection results, including merging candidate regions, deduplication, and classification. The generation module is used to determine the coordinate information of the anchor point region matrix based on the digital pathological image coordinate system and generate a list of small cancerous lesion anchor point warning islands.

2. The small cancerous lesion anchor point warning system based on digital pathological images as described in claim 1, characterized in that, It also includes an image preprocessing module, which is used to preprocess digital pathological images before the feature extraction module uses a deep learning model to extract multi-level features from digital pathological images, including denoising, artifact removal and image normalization.

3. The small cancerous lesion anchor point warning system based on digital pathological images as described in claim 1, characterized in that, The feature extraction module includes: The first extraction module is used to learn the spatial relationships in the image based on the structure of the convolutional neural network using a deep learning model, and extract morphological features; the morphological features include the shape, size, and boundaries of objects in the image. The second extraction module is used to extract texture features based on a deep learning model; the texture features describe the patterns or arrangement rules of local regions in the image; The third extraction module is used to extract color features based on a deep learning model through color space conversion and color histogram statistics; the color features are the color distribution and changes of pixels in the image.

4. The small cancerous lesion anchor point warning system based on digital pathological images as described in claim 1, characterized in that, The anchor point detection module includes: The first training module is used to train the object detection model using incomplete labeled data based on weakly supervised learning techniques, and to make up for the lack of labeled data by introducing additional constraints or assumptions. The second training module is used to expand the training set using unlabeled images based on semi-supervised learning techniques, and to guide the labeling process of unlabeled images by constructing pseudo-labels or using the prediction results of object detection models. The localization module is used to locate candidate regions for small cancerous lesions based on feature extraction results using a trained target detection model, and obtain anchor point detection results.

5. The small cancerous lesion anchor point warning system based on digital pathological images as described in claim 1, characterized in that, The optimization module includes: The merging and deduplication module is used to merge and deduplicate candidate regions based on the nonmaximum suppression algorithm. The regional hierarchical screening module is used to set different screening criteria or thresholds, and divide candidate regions into different levels or categories based on the first classification model. A classification model calibration module is used to adjust the output of the first classification model based on a calibration method; the calibration method includes at least one of Platt calibration, Isotonic regression calibration, histogram calibration, and neural network calibration.

6. The small cancerous lesion anchor point warning system based on digital pathological images as described in claim 1, characterized in that, The optimization module includes: The merging module is used to calculate the overlapping area of ​​each pair of candidate regions and compare it with a preset overlapping area threshold. Pairs of candidate regions with overlapping areas greater than the preset overlapping area threshold are merged into a new region. The bounding box of the new region is the smallest bounding rectangle of all merged regions. The deduplication module is used to calculate the shape similarity of each pair of candidate regions and compare it with a preset shape similarity threshold. When the shape similarity is determined to be greater than the preset shape similarity threshold, one of the candidate regions in the pair is removed. The classification module is used to extract the feature vector of each candidate region, input it into the pre-trained second classification model, and output the classification result.

7. The small cancerous lesion anchor point warning system based on digital pathological images as described in claim 1, characterized in that, The generation module includes: The conversion module is used to convert the position of the cancerous lesion region in the digital pathology image into coordinate information in the anchor point matrix based on the characteristics of the cancerous lesion region in the coordinate system of the digital pathology image and the coordinate transformation method. The characteristics of the cancerous lesion region include area, perimeter and position. The viewport scaling module is used to record the scaling ratio and position information of each anchor point region while generating anchor point matrix coordinate information.

8. The small cancerous lesion anchor point warning system based on digital pathological images as described in claim 1, characterized in that, Also includes: The multimodal data fusion module is used to register and fuse the anchor point detection results with imaging modal data or genomic modal data based on the deep registration algorithm after the anchor point detection module obtains the anchor point detection results, so as to obtain the processed anchor point detection results. The optimization module is used to optimize the processed anchor point detection results.

9. The small cancerous lesion anchor point warning system based on digital pathological images as described in claim 1, characterized in that, Also includes: The adjustment module is used to compare the digital pathological image with the location standard pathological image before the feature extraction module extracts multi-level features from the digital pathological image using a deep learning model. Based on the comparison results, the recognition angle of the digital pathological image is adjusted to obtain the adjusted digital pathological image.

Images