Lung cancer pathological section analysis method and system based on artificial intelligence

Abstract

The invention belongs to the technical field of pathological diagnosis, and discloses an artificial intelligence-based lung cancer pathological section analysis and system, which introduces identification basis visual information generated by an artificial intelligence model, and combines the correction operation of a plurality of pathologists on a model identification result and the acquisition of correction reasons, so that the pathological section analysis of the lung cancer is realized. Further performing consistency analysis on the correction results to generate uniform data, and finally performing incremental training on an artificial intelligence model by using the data, thereby constructing a closed-loop feedback optimization mechanism combining artificial intelligence and expert knowledge. The model performance problem caused by insufficient training data coverage, expert labeling efficiency and result inconsistency of an artificial intelligence model in pathological diagnosis is solved, and model learning and performance improvement are achieved.

Description

Technical Field

[0001] The present application relates to the technical field of pathological diagnosis, and more specifically, to an artificial intelligence-based lung cancer pathological section analysis method and system. Background Art

[0002] In medical diagnostic practice, the pathological diagnosis of lung cancer is a critical step in determining treatment plans and assessing prognosis. Pathologists examine lung tissue sections under a microscope to identify the morphology, arrangement, and interrelationship of cancer cells with surrounding tissue. Accurate detection of early-stage invasive lesions plays a crucial role in patient prognosis. However, these lesions are often small, have subtle morphological features, and exhibit a high degree of morphological similarity to carcinoma in situ or benign lesions, making identification difficult and prone to missed or misdiagnosis, presenting significant challenges to clinical diagnosis.

[0003] With the development of digital pathology technology, pathology slides are scanned as high-resolution whole slide images (WSIs), providing a data foundation for AI-assisted diagnosis. AI-based auxiliary diagnosis systems have been introduced into the pathology diagnostic process, using deep learning models to analyze whole slide images, aiming to automatically identify lesions and assist pathologists in diagnosis.

[0004] However, existing AI models still face numerous challenges in practical clinical application. Initial training datasets often have limited coverage, which impacts diagnostic accuracy and reliability. To improve model performance, continuous training data expansion and iterative model optimization are essential. This requires pathologists to manually review the AI ​​system's diagnostic results and accurately annotate cases that the system failed to accurately identify. This new, clinically valuable, expertly annotated data is crucial for model learning and understanding the characteristics of complex lesions.

[0005] However, the precise labeling of early-stage invasive lesions is time-consuming and highly dependent on expert experience, significantly increasing the workload of pathologists. Furthermore, for some difficult cases with ambiguous morphology or unclear boundaries, different pathologists may have subtle differences in their understanding of diagnostic criteria, leading to inconsistencies in labeling results. The scarcity of expert labeling resources and the subjective variability of labeling results constitute a bottleneck in the acquisition of high-quality training data. Due to the inefficiency of acquiring expert-labeled data and the difficulty in ensuring consistency, AI models are unable to achieve rapid and effective iterative updates, resulting in a lag in the speed at which the models learn and improve from clinical practice, making it difficult to adapt to ever-changing clinical needs.

[0006] Furthermore, the inherent inconsistency of expert annotation results, even among experienced pathologists, can introduce noise into the training data. When the model is trained using this data, which contains subjective differences, it may not learn the universal characteristics of early-stage invasive lesions, but may instead learn the preferences of specific experts or in specific scenarios. This further reduces the model's generalization ability and recognition accuracy when faced with real-world data from different hospitals, different experts, or different diagnostic practices, making it difficult to provide stable and reliable auxiliary diagnostic support in multi-center clinical settings. This data quality issue directly affects the ultimate clinical practicality and credibility of the AI ​​model, and a technical solution that can effectively address this issue is urgently needed.

[0007] In view of the above problems, the existing technology is in urgent need of improvement. Summary of the Invention

[0008] The purpose of this application is to provide an artificial intelligence-based lung cancer pathology section analysis method and system, which can improve the accuracy, reliability and clinical practicality of artificial intelligence-assisted diagnosis.

[0009] In a first aspect, the present application provides an artificial intelligence-based lung cancer pathology section analysis method, which identifies early invasive lesions in lung cancer pathology sections based on an artificial intelligence model. The method comprises the following steps: A1. Obtain full slide images of lung cancer pathology sections; A2. Analyze the whole slide image based on an artificial intelligence model to identify early invasive lesion areas and generate visualization information based on the identification of the early invasive lesion areas; A3. Receive correction operations performed by multiple pathologists on the early invasive lesion region based on the identification visualization information, and obtain correction reasons corresponding to the correction operations to obtain multiple correction results; A4. performing consistency analysis on the correction results to generate final early invasive lesion area identification result information; A5. Incrementally train the artificial intelligence model based on the final early invasive lesion area identification result information.

[0010] In a second aspect, the present application provides an artificial intelligence-based lung cancer pathology section analysis system, which identifies early invasive lesions in lung cancer pathology sections based on an artificial intelligence model. The system includes: An image acquisition module, used to acquire full-slide images of lung cancer pathological sections; an image analysis module for analyzing the whole slide image based on an artificial intelligence model, identifying early invasive lesion areas, and generating visual information based on the identification of the early invasive lesion areas; a correction feedback module, configured to receive correction operations performed by multiple pathologists on the early invasive lesion region based on the identification basis visualization information, and obtain correction reasons corresponding to the correction operations to obtain multiple correction results; A consistency analysis module, configured to perform consistency analysis on the correction results to generate final early invasive lesion area recognition result information; The model training module is used to perform incremental training on the artificial intelligence model based on the final early invasive lesion area identification result information.

[0011] Beneficial effects: The present application provides an artificial intelligence-based lung cancer pathology section analysis and system, which introduces visualization information of the recognition basis generated by the artificial intelligence model, and combines the correction operations of multiple pathologists on the model recognition results and the acquisition of the reasons for the corrections, further performs consistency analysis on these correction results to generate unified data, and finally uses these data to perform incremental training on the artificial intelligence model, thereby constructing a closed-loop feedback optimization mechanism combining artificial intelligence and expert knowledge, solving the model performance problems of the artificial intelligence model in pathological diagnosis caused by insufficient training data coverage, expert labeling efficiency and inconsistent results, and realizing model learning and performance improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 Flowchart of the artificial intelligence-based lung cancer pathology section analysis method provided in the embodiment of the present application.

[0013] Figure 2 Schematic diagram of the structure of the artificial intelligence-based lung cancer pathology section analysis system provided in an embodiment of the present application.

[0014] Explanation of numbers: 1. Image acquisition module; 2. Image analysis module; 3. Correction feedback module; 4. Consistency analysis module; 5. Model training module. DETAILED DESCRIPTION

[0015] The technical model in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of the embodiments. The components of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for protection, but merely represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work fall within the scope of protection of this application.

[0016] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and should not be understood as indicating or implying relative importance.

[0017] refer to Figure 1 This application proposes an artificial intelligence-based lung cancer pathology section analysis method, which identifies early invasive lesions in lung cancer pathology sections based on an artificial intelligence model. The method includes the following steps: A1. Obtain full slide images of lung cancer pathology sections; A2. Analyze the whole slide image based on an artificial intelligence model to identify early invasive lesion areas and generate visualization information based on the identification of the early invasive lesion areas; A3. Receive correction operations performed by multiple pathologists on the early invasive lesion region based on the identification visualization information, and obtain correction reasons corresponding to the correction operations to obtain multiple correction results; A4. performing consistency analysis on the correction results to generate final early invasive lesion area identification result information; A5. Incrementally train the artificial intelligence model based on the final early invasive lesion area identification result information.

[0018] Among them, the artificial intelligence model refers to a computational model obtained through algorithm training that can analyze and recognize patterns in image data. It can use deep learning networks, convolutional neural networks or recurrent neural networks, etc., and is mainly used to automatically identify specific areas in pathological images.

[0019] Among them, recognition-based visualization information refers to the data that graphically presents the internal decision-making process or focus area of ​​the artificial intelligence model when identifying the lesion area. It can be generated in the form of heat maps, saliency maps or attention maps, and is mainly used to display the judgment logic of the model and assist doctors in understanding and evaluating the model results.

[0020] Among them, the correction operation refers to the adjustment behavior of the early invasive lesion area made by the pathologist based on the recognition results and visualization information provided by the model. It can include adding, deleting, adjusting the boundaries or modifying the classification of the area. It is mainly used to correct the recognition bias of the model and introduce expert experience.

[0021] Among them, the correction reason refers to the explanatory description or classification label provided by the pathologist when performing the correction operation. It can be obtained by preset vocabulary selection or text input. It is mainly used to record the logic of the expert correction and provide semantic information for model learning.

[0022] Among them, consistency analysis refers to the process of comparing and evaluating the correction results provided by multiple pathologists. It can be processed by voting, consensus mechanism or statistical methods. It is mainly used to resolve subjective differences among experts and generate unified data.

[0023] Among them, incremental training refers to the process of continuously learning and optimizing the model based on the existing model using new data. It can be achieved using technologies such as transfer learning, online learning or reinforcement learning, and is mainly used to improve the model's ability to recognize new types or complex lesions.

[0024] The core innovation of this application lies in introducing visualization information of the recognition basis generated by the artificial intelligence model, combining the correction operations of multiple pathologists on the model recognition results and the acquisition of the reasons for the corrections, further analyzing the consistency of these correction results to generate unified data, and finally using these data to perform incremental training on the artificial intelligence model, thereby constructing a closed-loop feedback optimization mechanism combining artificial intelligence and expert knowledge, solving the model performance problems of the artificial intelligence model in pathological diagnosis caused by insufficient training data coverage, expert labeling efficiency and inconsistent results, and realizing model learning and performance improvement.

[0025] Specifically, this method establishes a closed-loop feedback system that collaborates between artificial intelligence (AI) and pathologists, enabling continuous optimization of the AI ​​model for identifying early invasive lesions in lung cancer pathology slides. First, the system acquires whole-slide images of lung cancer pathology slides as input for the AI ​​model's analysis. The AI ​​model then analyzes these images, automatically identifying potential early invasive lesion regions and simultaneously generating visualizations of the model's basis for identifying these lesions. This visualization reveals the model's focus and decision-making logic within the images, providing a reference for human intervention. Next, the system receives corrections from multiple pathologists to the model's initial identification results based on this visualization. These corrections include adding, deleting, adjusting boundaries, or modifying classifications of lesion regions. The system also captures the pathologists' reasons for these corrections, which provide semantic information about the expert judgments. By aggregating the corrections and reasons from multiple experts, the system generates multiple corrections. To address potential diagnostic discrepancies between experts, the system performs consistency analysis on these corrections, identifying and reconciling differences in opinion between experts. Ultimately, a single early invasive lesion identification result is generated, either through expert consensus or system fusion. Finally, this verified and unified recognition result information is used to incrementally train the artificial intelligence model, enabling the model to continuously learn and adapt from actual clinical cases and expert feedback, thereby improving its recognition accuracy and generalization ability for complex early invasive lesions.

[0026] As a preferred embodiment, the solution of this application is implemented as follows: First, digitized whole-slide images of lung cancer pathology sections are acquired using a digital pathology scanner. Next, these whole-slide images are analyzed using an artificial intelligence model trained on a deep convolutional neural network to automatically detect and mark suspected early-stage invasive lesions. During the recognition process, the model simultaneously generates an attention map based on a gradient-weighted class activation map. This map is superimposed on the original image as a heat map, with color depth indicating the model's attention to different image regions, serving as visual information for recognition. The system then displays the model recognition results and attention map to multiple pathologists via a web interface or dedicated client software. Each pathologist can use interface tools to interactively modify the lesion region (hereinafter referred to as the "early-stage invasive lesion region") identified by the model, for example by dragging a bounding box to adjust the lesion's extent and selecting a reason for the modification from a preset list of pathological features, such as "insufficient cellular atypia," "atypical acinar architecture," or "indistinct invasive margins." The system collects each pathologist's modification results and corresponding reasons. These revised results are then compared, for example by calculating the Jaccard similarity coefficient and classification consistency of the revised areas (i.e., the revised early invasive lesion areas) by different doctors, to identify discrepant lesion areas. For these discrepant areas, the system can trigger an expert consultation process or generate final lesion identification information based on preset fusion rules. Ultimately, this dataset, which has been revised and reconciled by multiple experts, is used for incremental training of the AI ​​model, for example through fine-tuning or continuous learning strategies, to improve the model's recognition performance for new or difficult lesions.

[0027] Through the above scheme, this application solves the problems of insufficient training data coverage, poor expert annotation efficiency and consistency in the identification of early invasive lesions in lung cancer pathology sections by artificial intelligence models. This scheme improves the transparency of model decision-making by providing visual information on the basis of recognition, and assists pathologists in making corrections efficiently. At the same time, the introduction of correction operations of multiple pathologists and the acquisition of correction reasons enrich the semantic information of model learning and solve the limitations of a single model or a single expert judgment. The consistency analysis of the correction results eliminates the data noise caused by subjective differences among experts and ensures the quality of data used for model training. Finally, the artificial intelligence model is incrementally trained using high-quality recognition result information, so that the model can continue to learn and optimize from challenging cases and expert feedback encountered in actual clinical practice, thereby improving the model's recognition accuracy and generalization ability for complex early invasive lesions, enhancing pathologists' trust in the artificial intelligence-assisted diagnosis system, and improving the efficiency and accuracy of clinical diagnosis.

[0028] In some embodiments, step A2 comprises: A201. Analyze the whole slide image based on the artificial intelligence model to identify early invasive lesion areas and obtain a preliminary identification image including the boundaries of the early invasive lesion areas; A202. Extracting response information of the artificial intelligence model to each region in the whole slide image during the process of identifying the early invasive lesion region; A203. Based on the response information, an attention map is generated and superimposed on the preliminary identification image to form a graphical identification image, which serves as visual information for identifying the early invasive lesion area; the color depth of the attention map indicates the degree of attention of the artificial intelligence model to each area in the whole slide image.

[0029] Among them, response information refers to data such as the activation degree, feature extraction intensity or gradient change of different areas of the input image by the internal layers of the artificial intelligence model when processing the whole slide image. It can be achieved by using data generated by interpretable methods such as the activation map of the middle layer of the model, gradient weighted class activation mapping (Grad-CAM) or LRP (Layer-wise Relevance Propagation, layer-by-layer propagation of relevance scores).

[0030] Among them, the attention map refers to a visual heat map or pseudo-color map generated based on the response information of the artificial intelligence model, which is used to intuitively represent the model's attention level or importance weight in different areas of the image. It can be achieved by converting the response information into a visually recognizable intensity distribution and superimposing it with the image.

[0031] The solution in this application aims to provide a more transparent and interpretable AI-assisted diagnosis process, thereby improving the efficiency and data quality of pathologists' corrections to model recognition results. Specifically, when analyzing whole-slide images to identify early-stage invasive lesions, the AI ​​model first performs preliminary image processing, identifies the lesion area, and generates a preliminary recognition image containing its boundaries. This preliminary recognition image provides the pathologist with an intuitive understanding of the model's location and extent, serving as the basis for subsequent interaction. Furthermore, to enhance the transparency of the model's decision-making, the system further delves into the AI ​​model's internal operating mechanisms. During the lesion identification process, the system extracts information about the model's response to each region in the whole-slide image. This response information reflects which parts of the image had a critical impact on the model's recognition decision. By acquiring this internal data, this solution can reveal the model's "thinking process," providing data support for subsequent visual interpretation. The system then generates an attention map based on the extracted response information. This attention map visually represents the AI ​​model's attention to each image region through color depth, indicating which areas the model considers most important for identifying early invasive lesions. Overlaying this attention map on the initial recognition image creates a graphical recognition image. This graphical recognition image serves as a visual representation of the early invasive lesion region. It not only displays the lesion regions identified by the model but, more importantly, clearly reveals the basis for the model's judgment. When reviewing the recognition results, pathologists can simultaneously see the key areas of focus of the model, allowing them to more intuitively understand the model's decision-making logic. This graphical recognition image enables pathologists to make corrections based on their understanding of the model's decision-making process. For example, if the model's attention areas align with key pathological features identified by the pathologist, this can enhance the pathologist's confidence in the model's results. Conversely, if the model focuses on non-lesion areas or the focus points do not align with the pathological features, the doctor needs to conduct a thorough review.

[0032] This transparency greatly assists pathologists in performing efficient and accurate correction operations, and can provide high-quality correction reasons, thereby effectively improving the data quality and efficiency of subsequent incremental training of artificial intelligence models, thereby promoting the auxiliary value and continuous improvement capabilities of artificial intelligence models in complex pathology diagnosis, and solving the problem of difficulty in correction and limited data quality caused by insufficient model interpretability.

[0033] Preferably, step A203 may include: determining, based on the response information, a contribution value of each pixel region in the full slide image to an early invasive lesion region recognition result; Normalizing the contribution values, and selecting corresponding colors from a preset color gradient according to the normalized contribution values ​​to generate the attention map; The focus map is superimposed on the preliminary identification image to form a graphical identification image, which serves as visual information for identifying the early invasive lesion area.

[0034] The response information reflects the model's focus on different regions of the input image or the strength of its feature extraction. Based on this, the contribution value quantifies the response information and represents the contribution of each pixel region in the full-slide image to the final recognition result. This value can be calculated based on gradients, activation maps, or specific interpretation algorithms to measure the importance of each pixel or region in the model's decision-making.

[0035] Among them, normalization refers to the process of converting the original contribution value into a unified, preset numerical range, such as [0,1] or [0,255]. This can eliminate the differences in the dimension and range of the contribution value between different images or different model outputs, and ensure the consistency and comparability of subsequent color mapping.

[0036] Among them, the preset color gradient refers to a series of colors arranged in a specific order with smooth transitions between these colors, which is used to map the normalized contribution value to visually intuitive color depth or hue changes, so that the size of the contribution value can be intuitively presented through color.

[0037] The solution of this application solves the problem of lack of accuracy and consistency in attention map generation by introducing specific methods for quantifying, standardizing, and visually mapping the response information of the artificial intelligence model, thereby improving the accuracy and interpretability of the visual information used as the basis for identification. The solution first determines the contribution value of each pixel area in the whole slide image to the identification result of the early invasive lesion area based on the response information of the artificial intelligence model. This step converts the abstract response information within the model into a specific, quantifiable "contribution value", clarifying which areas in the image play a key role in the model's recognition decision, thereby providing an accurate quantitative basis for the subsequent generation of the attention map and solving the problem that the response information is difficult to directly convert into a visual basis. Next, the contribution value is normalized, and the corresponding color is selected from a preset color gradient based on the normalized contribution value to generate the attention map. By normalizing the contribution value, it ensures that the contribution values ​​of different cases or different model outputs can be compared and mapped on a unified scale, eliminating the dimensional differences that may exist in the original contribution value and ensuring the consistency of the visualization results. On this basis, the corresponding color is selected from the preset color gradient based on the normalized contribution value, providing a standardized mapping rule for the generation of the attention map. This preset color gradient establishes a clear and intuitive correspondence between color depth and contribution value, allowing pathologists to quickly and accurately understand the degree of attention or contribution of the artificial intelligence model to different areas, thereby improving the interpretability and practicality of the attention map. Finally, the generated attention map is superimposed on the preliminary recognition image to form a graphical recognition image, which serves as the visual information for the identification of early invasive lesions.

[0038] By overlaying the precisely quantified and standardized attention map with the preliminary recognition image, pathologists can visually see the basis for the AI ​​model's recognition judgment, namely the key areas of focus and their contribution, while reviewing the recognition results. This graphical recognition image provides pathologists with transparent and traceable diagnostic information, helping them quickly understand the model's decision-making logic and efficiently perform manual review and correction operations.

[0039] As a preferred embodiment, when determining the contribution of each pixel region in a whole-slide image to the early invasive lesion region identification result, a gradient-weighted approach, such as the Grad-CAM (Gradient-weighted Class Activation Mapping) or LRP (Layer-wise Relevance Propagation) algorithm, can be employed. Specifically, the gradient of the predicted score for the target class (early invasive lesion) relative to a specific feature map (for example, if a convolutional neural network is used, the feature map output by the model at a specific convolutional or fully connected layer can be extracted, which reflects the activation strength of the model for different image regions). These gradients are used as weights to perform a weighted summation of the feature maps, thereby generating a raw heatmap representing the contribution of each pixel region. When normalizing the contribution values, the raw heatmap obtained above can be subjected to min-max normalization, for example, scaling all pixel values ​​in the heatmap to a range of 0 to 255, where 0 represents the lowest contribution and 255 represents the highest contribution. When selecting the corresponding color from the preset color gradient based on the normalized contribution value, a color gradient from transparent (or low-saturation blue) to high-saturation red (or yellow) can be used. Areas with lower contribution values ​​can be mapped to transparent or light blue, while areas with higher contribution values ​​can be mapped to dark red or bright yellow to intuitively represent the increasing degree of attention. In this way, an attention map that matches the size of the full-slide image can be generated. When superimposing the attention map on the preliminary identification image, the generated attention map can be superimposed on the preliminary identification image in a semi-transparent manner through image fusion technology, such as alpha blending. For example, the transparency of the attention map can be set to 0.5 so that it can clearly display the area of ​​interest without completely blocking the details of the original pathology image.

[0040] In some embodiments, step A3 comprises: A301. Displaying the identification basis visualization information to the pathologist as reference information for performing correction operations on the early invasive lesion area; A302. Receive the pathologist's correction operation on the early invasive lesion region; the correction operation includes at least one of adding, deleting, adjusting the boundary, and modifying the classification of the early invasive lesion region; A303. According to the type of correction operation and the image features of the corresponding correction area, a list of candidate correction reasons related to the correction operation is extracted from a preset pathological feature vocabulary and displayed to the pathologist; A304. Receive the correction reason selected by the pathologist from the candidate correction reason list, and / or receive the supplementary correction reason input by the pathologist; A305. Normalize the selected correction reason and / or the input supplementary correction reason to obtain a structured correction reason; A306. Associate the structured correction reason with the correction operation and record it to obtain a correction result.

[0041] Among them, the preset pathological feature vocabulary library refers to a pre-built collection of standard pathological terms and concepts, which can be organized in the form of medical ontology, controlled vocabulary or professional field knowledge graph. For example, it can include "benign lesions", "inflammatory response", "fibrosis", "necrosis", "cell atypia", "insufficient infiltration depth" and other entries. These entries have been reviewed and classified by experts and aim to provide accurate and consistent correction reason options.

[0042] Among them, normalization refers to the process of converting unstructured or semi-structured text information into a unified, machine-readable structured format. It can be achieved by using natural language processing technology (such as named entity recognition, semantic analysis), rule matching or ontology-based mapping methods. For example, "the lesion is benign" can be normalized into "benign lesion", or "unclear boundaries" can be normalized into "fuzzy boundaries".

[0043] Among them, the structured correction reason refers to the correction reason that has been normalized and expressed in a predefined data format (for example, a key-value pair, a JSON object, or a database record). The correction result obtained by associating it with the correction operation record may include the correction type, the image feature label of the correction area, and the corresponding standard pathology reason code or text. For example, it can be expressed as {"operation type":"deletion","image feature":"inflammation","cause":"benign inflammatory lesion"}, which facilitates storage, retrieval, and analysis by the system.

[0044] This solution forms an efficient and structured feedback loop by fine-tuning the process of pathologists' correction operations and correction reason acquisition. First, before the pathologist performs a correction operation on the early invasive lesion area, the system will display visual information of the identification basis to the pathologist. This allows the pathologist to intuitively understand the internal logic and focus of the artificial intelligence model's judgment, so that they can more accurately and quickly locate the deviation of the model's recognition, and use this as reference information to perform correction operations such as addition, deletion, boundary adjustment or classification modification. This transparent information presentation improves the accuracy and efficiency of the pathologist's correction operation.

[0045] After receiving the pathologist's correction operation, the system does not simply wait for free text input. Instead, based on the specific type of correction operation and the image features of the corrected area, it intelligently filters and extracts a list of candidate correction reasons related to the current correction situation from the preset pathology feature vocabulary. This mechanism reduces the burden of pathologists thinking and entering correction reasons, while ensuring that the options provided are professional and standard. Pathologists can quickly select the reasons that are in line with the actual situation from this intelligently generated list, which not only improves operational efficiency, but also effectively avoids the problem of inconsistent descriptions of reasons due to differences in personal expression habits among different doctors.

[0046] To accommodate special circumstances and deeper expert insights, the system offers standard selections while also allowing pathologists to enter supplemental correction reasons. Regardless of whether these are selected or supplemented, all correction reasons received are further normalized and converted into unified, machine-readable structured correction reasons. This process eliminates linguistic ambiguity and discrepancies in presentation, ensuring that all expert feedback data can be accurately understood and utilized by the system. Ultimately, these structured correction reasons are linked to the specific correction actions performed by the pathologist, forming a complete correction result.

[0047] The coordinated operation of this series of steps enables the AI ​​model to receive high-quality and valuable supervisory data. By providing visual information supporting identification, pathologists can make corrections more effectively, while a guided, structured process for eliciting the cause ensures standardization and consistency of the correction reasons. These meticulously processed correction results provide the foundation for subsequent targeted incremental training of the AI ​​model. For example, when the model misses a specific type of lesion, analysis of the structured correction reasons can clearly identify specific causes such as "lack of cellular atypia" or "background inflammatory interference." This guides the model to focus more on learning these subtle features during incremental training, improving the model's generalization and recognition accuracy. This mechanism enables the AI ​​model to continuously learn and improve from clinical practice, forming a self-optimizing closed loop and enhancing the practicality and credibility of the auxiliary diagnosis system.

[0048] Preferably, step A303 may include: Preliminarily extracting a list of candidate correction reasons related to the correction operation from a preset pathological feature vocabulary library according to the type of the correction operation and the image features of the corresponding correction area; Based on the image features of the correction area, analyzing the degree of correlation between each cause in the list of candidate correction reasons initially extracted and the image features, and obtaining a feature correlation degree of each candidate correction cause; Obtaining co-occurrence frequency information of correction reasons of early invasive lesion areas with similar image features to the correction area in historical correction data; Prioritizing or grouping the initially extracted candidate correction reason list based on the feature correlation and the co-occurrence frequency information to obtain an optimized candidate correction reason list; The optimized list of candidate revision reasons is displayed to the pathologist.

[0049] The image features of the corrected area refer to the visual information presented by the pathologist in the early invasive lesion area where the correction is performed, such as cell morphology, tissue structure, staining intensity, lesion boundary clarity, nuclear-cytoplasm ratio, degree of atypia, and infiltration pattern. These features can be acquired and quantified using image processing technologies such as deep learning-based feature extractors, texture analysis algorithms, and color histogram analysis.

[0050] Feature relevance refers to the degree of match or correlation between the initially extracted candidate correction reasons and the image features of the correction area. This can be calculated in a variety of ways, such as by using a pre-trained semantic similarity calculation model to evaluate the semantic difference between the cause description and the image feature description, or by using a machine learning model to learn the mapping relationship between image features and correction reasons and output a relevance score.

[0051] Co-occurrence frequency information refers to the statistical frequency or probability of selecting a specific correction reason for a lesion region with similar image features to the current correction region in historical correction data. This information can be obtained using statistical analysis methods, such as constructing a co-occurrence matrix, calculating conditional probabilities, or using association rule mining algorithms to quantify the degree of co-occurrence between specific image features and correction reasons.

[0052] To improve the efficiency of extracting correction reasons after receiving a pathologist's correction operation on an early invasive lesion, this solution first extracts a preliminary list of candidate correction reasons related to the correction operation from a pre-set pathology feature vocabulary based on the type of correction operation and the image features of the corresponding correction area. This preliminary extraction ensures that the scope of the initial list is roughly consistent with the current correction behavior and image content, laying the foundation for subsequent refined processing. Based on this, the system further analyzes the degree of correlation between each candidate correction reason and the image features of the correction area, thereby deriving the feature correlation degree of each candidate correction reason. This analysis step deeply explores the visual information of the correction area. By quantifying the degree of match between each candidate reason and the current image features, the system can identify which reasons are most semantically consistent with the specific image manifestations observed by the pathologist. This provides an objective basis for subsequent optimization based on image content. Furthermore, to incorporate the accumulated experience and historical patterns of the expert community, the system obtains information on the co-occurrence frequency of correction reasons for early invasive lesions in historical correction data that have similar image features to the correction area. By analyzing a large amount of historical revision data, the system can understand which revision reasons are frequently selected by pathologists in cases with image features similar to the current revision area. This reflects common clinical issues, diagnostic consensus, or typical revision reasons corresponding to specific pathological features, providing an empirically based statistical basis for optimization. Subsequently, this solution prioritizes or groups the initially extracted list of candidate revision reasons based on feature correlation and co-occurrence frequency information, resulting in an optimized list of candidate revision reasons. This is the core mechanism of this solution. By comprehensively considering the visual relevance of the reason to the current image and the frequency of selection in similar historical cases, the system can re-rank or classify the candidate reasons according to intelligent strategies. For example, revision reasons that are both highly correlated with the current image features and frequently appear in historical data are given higher priority or more prominent placement. This two-dimensional optimization ensures that the list presented to the pathologist is both visually accurate and consistent with empirical patterns in clinical practice, significantly increasing the likelihood that the pathologist will quickly identify and select the correct revision reason. Finally, the system displays the optimized list of candidate revision reasons to the pathologist. By presenting a list that has been intelligently filtered and sorted, pathologists no longer need to waste time searching through redundant or irrelevant options, but can instead select directly from a highly relevant and easy-to-understand list. This series of steps, working together, enables pathologists to provide corrective feedback more efficiently and accurately, thereby providing higher-value, more consistent structured data for the continuous learning and improvement of AI models.Compared with the basic solution of only preliminarily extracting the list, this significantly improves the value and efficiency of data collection, and thus improves the ability of the artificial intelligence model to conduct effective incremental training through these feedbacks, solving the problem of limiting model performance improvement due to insufficient feedback data quality.

[0053] In one specific embodiment, when a pathologist performs a correction operation on an early invasive lesion region identified by an AI model, for example, revising a region labeled benign by the model to an early invasive lesion, the system first identifies the correction operation as a "classification correction" and obtains image features of the correction region, such as at least one of its nuclear size, mitotic figures, glandular structure, and invasion depth. Based on this information, the system initially selects candidate correction reasons related to "classification correction" and "early invasive lesion" from a pre-set pathological feature vocabulary, such as "invasive adenocarcinoma," "micropapillary structure," and "atypical lepidic growth pattern." Next, based on the image features of the correction region, the system utilizes a pre-trained image feature analysis module to evaluate the degree of match between each candidate correction reason initially extracted and the current image features. For example, if the correction region has a micropapillary structure, the feature relevance of the cause "micropapillary structure" will be calculated as a higher value. This relevance can be a value between 0 and 1, indicating the degree of semantic and visual match. At the same time, the system queries a database containing a large number of historical revision records. This database records the revisions performed by pathologists on early-stage invasive lesions with similar image features, along with the corresponding revision reasons. Using an image feature similarity matching algorithm, the system identifies historical cases with image features highly similar to those of the currently revised area and calculates the co-occurrence frequency of each revision reason selected in these similar cases. For example, if "micropapillary structure" is frequently selected in revisions of lesions with micropapillary structures in the historical data, its co-occurrence frequency information is recorded. The system then combines the calculated feature correlation with the acquired co-occurrence frequency information. For example, a weighted summation approach can be used to combine the feature correlation and co-occurrence frequency information to generate a comprehensive evaluation value. Revision reasons with higher comprehensive evaluation values ​​are given higher priority. For example, if "micropapillary structure" has both a high feature correlation and a high co-occurrence frequency, it will be placed at the top of the list. If a cause has a high correlation with its features but a low co-occurrence frequency (usually representing a rare condition), or if the features have a low correlation but a high co-occurrence frequency (usually representing a common but non-specific cause), the system will balance them according to the preset weighting rules. Ultimately, the system will prioritize the list of candidate correction causes that were initially extracted based on these comprehensive evaluation values ​​to form an optimized list. Finally, this optimized list of candidate correction causes will be displayed to the pathologist through the user interface. The pathologist can quickly select a cause that is highly consistent with the current correction situation from this list that has been sorted by intelligent strategies, such as directly clicking to select "micropapillary structure" to complete the feedback on the correction cause.

[0054] Preferably, the step of prioritizing or grouping the initially extracted candidate correction reason list based on the feature correlation and the co-occurrence frequency information to obtain an optimized candidate correction reason list may include: determining statistical sparsity of image features of the correction region; Determining, based on the statistical sparsity, contribution weights of the feature relevance and the co-occurrence frequency information in generating the optimized candidate correction reason list; According to the contribution weights, combined with the feature relevance and the co-occurrence frequency information, the initially extracted candidate correction reason list is prioritized or grouped to obtain the optimized candidate correction reason list.

[0055] Among them, statistical sparsity refers to the degree to which the image features of the correction area appear less frequently in the historical correction data. It can be achieved by counting the number of occurrences of the image features in the historical data set and comparing it with a preset threshold, or by analyzing its percentile ranking in the feature distribution.

[0056] Among them, the contribution weight refers to the relative importance or influence of the feature correlation and co-occurrence frequency information when generating the optimized list of candidate correction reasons. It can be dynamically adjusted using adaptive algorithms, machine learning models or based on preset rules to ensure that the reliability differences of different information sources are reflected.

[0057] Among them, when performing priority sorting or grouping processing, each reason in the preliminary extracted candidate correction reason list can be weighted and scored according to the determined contribution weight. For example, the final score of each candidate correction reason can be calculated as: score = (feature correlation * feature correlation weight) + (co-occurrence frequency information * co-occurrence frequency information weight). The system can sort the candidate correction reasons in descending order according to these scores, or group them according to the preset score range in which the scores fall (for example, if the score falls within the preset high score range, the corresponding candidate correction reason is classified as a "highly recommended" group; if the score falls within the preset medium score range, the corresponding candidate correction reason is classified as a "moderately recommended" group; if the score falls within the preset low score range, the corresponding candidate correction reason is classified as a "generally recommended" group), thereby obtaining an optimized candidate correction reason list and displaying it to the pathologist.

[0058] The method first determines the statistical sparsity of the image features in the correction region based on co-occurrence frequency information. This step assesses their statistical prevalence or rarity by analyzing the frequency of occurrence of the image features in the current correction region in historical data. For image features with low frequency in historical data, i.e., high statistical sparsity, this indicates that the corresponding historical co-occurrence frequency information may be insufficient or unrepresentative, thus providing a basis for subsequent weight assignment. Next, based on the determined statistical sparsity, the contribution of feature correlation and co-occurrence frequency information to the generated optimized list of candidate correction reasons is dynamically determined. This means that the system no longer simply combines the two types of information in a fixed ratio, but instead adjusts their relative importance based on the characteristics of the current case. For example, when the statistical sparsity of the image features in the correction region is high, the system assigns a higher weight to the feature correlation derived from the current image feature analysis, as historical co-occurrence frequency information may not provide a reliable reference. Conversely, for common cases with low statistical sparsity, historical co-occurrence frequency information is assigned a higher weight to fully utilize its statistical reliability. This adaptive weight distribution mechanism enables the system to more intelligently utilize information from different sources, avoiding the impact of the limitations of a single information source on the accuracy of the results. Finally, based on the determined contribution weights, combined with feature correlation and co-occurrence frequency information, the initially extracted list of candidate correction reasons is prioritized or grouped to obtain an optimized list of candidate correction reasons. Through this weighted combination approach, the system can more accurately sort or group candidate correction reasons, ensuring that the list of correction reasons presented to the pathologist not only includes options that are highly relevant to the current image features, but also takes into account the accumulation of historical experience and can be intelligently adjusted according to the rarity of the case.

[0059] This solution, combined with the previous technical solution, can further improve the performance of the entire lung cancer pathology section analysis method. Specifically, in the step of receiving the pathologist's correction operation and obtaining the correction reason, this solution dynamically adjusts the combination of feature correlation and co-occurrence frequency information, so that the preliminary extracted candidate correction reason list can be more accurately optimized. This optimization directly acts on the candidate correction reason list displayed to the pathologist, so that the pathologist can obtain more targeted and reliable reference information when performing the correction operation. This not only improves the efficiency of the pathologist in selecting the correction reason, but also ensures the accuracy of the selected correction reason, especially when dealing with rare or atypical lesions, it can effectively avoid misjudgments caused by insufficient historical data. Ultimately, these high-quality correction results will be used for incremental training of the artificial intelligence model, so that the model can learn from more accurate and representative expert feedback, and continuously improve its recognition and generalization capabilities for early invasive lesions, forming an efficient and adaptive closed-loop optimization system.

[0060] In some embodiments, step A4 comprises: A401. Compare the boundary information and classification information of each early invasive lesion region in the multiple correction results, identify the early invasive lesion region with correction differences, and determine the type of the correction differences; A402. Extracting, based on the correction difference type, the structured correction reason corresponding to the early invasive lesion region having the correction difference; A403. Assess the clinical importance of the revised difference based on the revised difference type and the extracted structured revision reason; A404. Based on the clinical significance of the revised differences, displaying the early invasive lesion regions with revised differences and the corresponding extracted structured reasons for revisions to the multiple pathologists to assist the pathologists in discussion; A405. Receive the consensus opinion formed after the discussion among the pathologists, or fuse the multiple correction results according to the preset fusion rules and the clinical importance of the correction differences to generate the final early invasive lesion area identification result information.

[0061] Correction discrepancies refer to inconsistencies in the correction operations (e.g., boundary adjustments, classification modifications, additions, or deletions) performed by different pathologists on the same early invasive lesion region. These discrepancies can manifest as insufficient boundary overlap, inconsistent classification labels, or a lesion being added by some doctors but not others. These discrepancies can be addressed using image processing algorithms (e.g., IoU calculation and pixel-level difference analysis) and text semantic analysis (e.g., classification label comparison).

[0062] The clinical importance of revision discrepancies is a quantitative assessment of the potential clinical impact of differences between revisions by different pathologists. It reflects the criticality of these differences to patient diagnosis, treatment selection, or prognosis. This can be achieved through expert scoring, weighting based on clinical guidelines, or prediction using machine learning models.

[0063] The preset fusion rules refer to strategies used by the system to integrate multiple revisions based on pre-defined logic when pathologists fail to reach a consensus or when automated processing is needed. For example, weighted fusion can be based on the clinical significance of the revision differences, the pathologist's qualifications, the pathologist's historical diagnostic accuracy, or majority voting. This can be achieved using algorithms such as weighted averaging, Bayesian fusion, decision trees, or rule engines.

[0064] The operational logic of this solution is to first receive multiple corrections performed by multiple pathologists on early invasive lesion regions and obtain the reasons for the corrections, resulting in multiple correction results. By comparing the boundary information and classification information of each early invasive lesion region in these correction results, the system can identify areas with correction discrepancies between different pathologists and further determine the specific types of these discrepancies. This comparison process is the foundation for subsequent processing, clarifying inconsistencies that require attention and resolution, avoiding blind processing of all results, and improving processing efficiency and targeting. Next, based on the identified correction discrepancy types, the system extracts structured correction reasons corresponding to the early invasive lesion regions with correction discrepancies. This step is important because it links the correction operation with the pathologist's root cause for the correction. By obtaining structured correction reasons, the system not only understands "where" the discrepancy occurs, but also "why" the discrepancy occurs, providing important semantic information for subsequent discrepancy assessment and resolution. This is due to the fact that the system has already received and normalized the correction reasons selected or entered by the pathologist when capturing the correction operation, giving these reasons a structured form that can be analyzed and utilized by the system. Subsequently, based on the type of corrected discrepancy and the extracted structured reasons for the correction, the system assesses the clinical significance of the corrected discrepancy. This assessment step is a key innovation of this solution. It goes beyond simple discrepancy detection to quantify the potential impact of the discrepancy from a clinical perspective. For example, critical discrepancies that could impact a patient's treatment plan or prognosis are assigned a higher clinical significance. This assessment helps distinguish between "minor discrepancies" and "critical discrepancies requiring resolution," ensuring that resources are focused on resolving valuable discrepancies. Based on the assessed clinical significance of the corrected discrepancy, the system then displays the early invasive lesion areas with the corrected discrepancy and the corresponding structured reasons for the correction to multiple pathologists to facilitate discussion. This step demonstrates the advantages of human-machine collaboration. The system identifies and quantifies the discrepancy and provides contextual information (the reasons for the correction). Based on this information, pathologists can engage in targeted discussions, leveraging their expertise and experience to reach consensus. This guided discussion mechanism improves the efficiency and quality of expert consensus building. Finally, the system receives the consensus reached by the pathologists after discussion, or fuses multiple revised results based on preset fusion rules and the clinical significance of the revised differences to generate the final early invasive lesion area identification result. This step is the end of the entire consistency analysis process, ensuring that a single, high-quality, clinically consensus-based or intelligently fused identification result is ultimately produced from multiple potentially inconsistent revised results.Whether reaching consensus through expert discussion or automated processing through intelligent fusion rules, the clinical importance of the differences is fully taken into consideration, thereby ensuring the accuracy and reliability of the final results and providing a solid foundation for the continuous learning and optimization of the artificial intelligence model.

[0065] Overall, this solution effectively compensates for the lack of consistency in multi-expert annotated data in existing technologies by introducing a consistency analysis mechanism for multi-source correction results and combining correction reasons and clinical importance for difference processing. Through systematic difference identification, cause extraction, importance assessment, expert-assisted discussion, and intelligent fusion, this solution can transform the originally scattered and potentially conflicting expert correction opinions into unified and high-quality final recognition result information. This not only solves the problem of reduced model training effectiveness due to data inconsistency, but more importantly, it incorporates the professional experience and judgment of pathologists into the data refinement process, making the final generated data more clinically valuable and reliable, thereby providing a solid foundation for the continuous incremental training of artificial intelligence models, improving the generalization ability and recognition accuracy of the model in complex clinical scenarios, and enhancing its practicality and credibility in multi-center clinical environments.

[0066] In a specific embodiment, this solution can be applied to process the revisions submitted by three pathologists (e.g., Doctor A, Doctor B, and Doctor C) regarding the same early-stage invasive lesion identified by an AI model. First, during the comparison phase, the system receives the revisions submitted by Doctors A, B, and C. Suppose Doctor A labels the lesion "adenocarcinoma" and outlines his or her boundaries. Doctor B also labels it "adenocarcinoma," but his or her outlined boundaries partially overlap with Doctor A's, but also differ. Doctor C labels the lesion "atypical adenomatous hyperplasia," and his or her boundaries differ from those of the previous two doctors. The system automatically compares the boundary information (e.g., by calculating pixel-level differences or Intersection over Union (IoU) values) and classification information (e.g., by directly comparing classification labels) of the lesion region in these three revisions. The comparison results reveal two types of revision discrepancies: "boundary inconsistency" and "classification inconsistency." Next, based on the identified revision discrepancy types, the system automatically extracts the structured revision reasons associated with each revision submitted by Doctors A, B, and C. For example, Doctor A's reason for boundary revision was "lack of obvious cellular infiltration at the lesion edge," Doctor B's reason for boundary revision was "interference from local inflammatory response," and Doctor C's reason for classification revision was "insufficient cellular atypia, not meeting the diagnostic criteria for adenocarcinoma." These structured reasons provide a deeper understanding of the discrepancies. The system then assesses the clinical importance of these differences based on the type of revision and the extracted structured reasons. For example, because the clinical treatment options for "adenocarcinoma" and "atypical adenomatous hyperplasia" differ significantly, the system would assess the clinical importance of "classification discrepancy" as high. Minor boundary differences, if they do not affect the overall assessment of the lesion's nature, might be assessed as moderate or low. This assessment can be made by combining information about the lesion's pathological characteristics (e.g., size, location, cell density) with the patient's clinical background information (e.g., age, medical history). The system then displays images of early invasive lesions with discrepancies in the corrections, along with the corrections from Doctors A, B, and C, and the corresponding structured reasons for the corrections, all on a single interactive interface based on the assessed clinical importance. Discrepancies of high clinical importance are highlighted. For example, the system highlights "classification inconsistency" as a high-importance discrepancy and displays Doctor C's reason for the correction, "insufficient cellular atypia." This facilitates targeted discussions among the three pathologists. For example, they can review the original slide images together and, drawing on their respective expertise, conduct a detailed analysis of the lesion's cellular morphology and histological structure. Finally, after the discussion, the system receives the pathologists' consensus. For example, after discussion, the three doctors ultimately reach a consensus that the lesion should be diagnosed as "atypical adenomatous hyperplasia" and jointly determine a final boundary. The system uses this consensus as the final early invasive lesion region identification result information.Alternatively, if pathologists fail to reach a clear consensus, the system can automatically process the differences based on pre-set fusion rules. For example, for classification differences of high clinical importance, the system may prioritize the majority expert opinion. For boundary differences, the system can use weighted averaging or an intelligent fusion algorithm based on morphological features based on their clinical importance to generate a comprehensive boundary, thereby obtaining the final identification result information for the early invasive lesion area.

[0067] Preferably, step A403 may include: Obtain pathological characteristics of early invasive lesions with corrected differences and clinical background information of the corresponding patients; Determining the evaluation indicator weights of the corrected difference in different clinical dimensions from a preset clinical evaluation rule library based on the pathological feature information and the clinical background information; Based on the type of correction difference, the structured correction reason and the evaluation index weight, the clinical importance of the correction difference is comprehensively calculated.

[0068] Among them, pathological characteristic information refers to data describing the morphological, histological, and cytological characteristics of early invasive lesions, which can be represented by data in multiple dimensions, including lesion size, shape, degree of cell atypia, number of mitotic figures, depth of infiltration, vascular invasion, distribution of necrotic areas, and immunohistochemical labeling results.

[0069] Among them, clinical background information refers to data related to the patient's individual health status, disease development process and treatment. It can be represented by various data types such as the patient's age, gender, medical history (such as smoking history, family cancer history), concomitant diseases, imaging examination results, blood biochemical indicators and accepted treatment plans.

[0070] Among them, the preset clinical evaluation rule base refers to a knowledge system constructed by expert knowledge, clinical guidelines and historical case data. It can be implemented in various forms such as rule-based expert systems, decision tree models, Bayesian networks or machine learning models. It is used to derive the importance evaluation results of the corrected differences in different clinical dimensions (i.e., the evaluation indicator weights) based on the input pathological feature information and clinical background information.

[0071] Among them, the evaluation indicator weight refers to the relative importance coefficient assigned to different clinical dimensions (such as the impact on diagnostic accuracy, the impact on treatment plan selection, and the impact on patient prognosis) when evaluating the clinical importance of the corrected difference. It can be expressed in the form of numbers, percentages, or grades, reflecting the contribution of each factor to the final evaluation result under specific pathological and clinical scenarios.

[0072] Among them, comprehensive calculation refers to the integration and processing of multi-source information such as the correction difference type, structured correction reasons, and evaluation indicator weights to obtain a unified and quantitative clinical importance assessment result. It can be achieved by using various algorithms such as weighted summation, multivariate regression analysis, or machine learning-based fusion models.

[0073] This method, building on existing assessments based on the type of correction discrepancy and structured correction cause, further incorporates pathological features of early invasive lesions and the patient's clinical background. Specifically, the system first acquires pathological features of early invasive lesions with correction discrepancies. This information directly reflects the biological behavior and potential malignancy of the lesions, such as size, morphology, and cellular atypia. Simultaneously, the system acquires clinical background information for the corresponding patients, providing individualized risk factors and prognostic considerations, such as age, medical history, and treatment regimen. The inclusion of this data enables the assessment of correction discrepancies to move beyond the type and superficial causes of the discrepancy and delve deeper into the specific circumstances of the lesion and patient, enabling a more accurate assessment of the true clinical significance of the discrepancy. Based on this acquired pathological and clinical background information, the system determines the weights of evaluation indicators for the correction discrepancy across different clinical dimensions from a pre-defined clinical evaluation rule library. This process enables intelligent and personalized assessment, as different pathological features and clinical backgrounds can lead to vastly different clinical significance for the same type of correction discrepancy. By leveraging a pre-defined rule base and incorporating specific case characteristics, the system dynamically assigns appropriate weights to various evaluation indicators (e.g., impact on diagnosis, treatment, and prognosis), ensuring clinical relevance and targetedness of the evaluation results. Finally, the system comprehensively calculates the clinical importance of the corrected discrepancy based on the type of corrected discrepancy, the structured reason for the correction, and the dynamically determined weights of the evaluation indicators. This step integrates all relevant information to produce a comprehensive and instructive assessment of clinical importance. By combining the type of corrected discrepancy (e.g., margin discrepancy, classification discrepancy), the structured reason for the correction provided by the pathologist (e.g., "disagreement in the proportion of invasive adenocarcinoma component"), and the weights of the evaluation indicators determined based on pathological and clinical information, the system performs a multi-dimensional, weighted calculation. This comprehensive approach avoids the limitations of single-factor evaluation and ensures that the evaluation results fully reflect the potential impact of the corrected discrepancy in actual clinical decision-making.

[0074] By combining these steps, this method expands the existing assessment process from a single correction difference to encompass the inherent characteristics of the lesion and the individual patient's condition, making the results more clinically valuable. This more refined and personalized assessment effectively distinguishes the true impact of different correction differences in actual clinical diagnosis and treatment, providing a solid and reliable basis for subsequent discussion and consensus among pathologists, or for the system to automatically integrate the results, significantly improving the accuracy of the final identification results.

[0075] Preferably, the step of comprehensively calculating the clinical importance of the revised difference based on the revised difference type, the structured revision reason, and the evaluation indicator weight may include: Converting the correction difference type and the structured correction reason into corresponding quantitative importance values ​​respectively; Performing weighted processing on the quantitative importance value according to the evaluation index weight; The weighted quantitative importance values ​​are aggregated to obtain the clinical importance of the corrected difference.

[0076] Among them, the quantitative importance value refers to converting the originally non-numerical correction difference type and structured correction reason into an indicator with numerical significance through preset mapping rules, scoring system or machine learning model, which can be achieved by numerical scoring, grade mapping or probability assignment.

[0077] Among them, weighted processing refers to applying different weight coefficients to quantitative importance values ​​according to different evaluation dimensions to reflect their relative contributions in the overall evaluation. It can be achieved by multiplication weighting, proportional distribution or statistical model-based methods.

[0078] Aggregation refers to integrating multiple weighted quantitative importance values ​​into a single, comprehensive value through specific mathematical operations or logical rules to represent the overall clinical importance of the corrected difference. It can be achieved by summation, average calculation or multi-factor comprehensive evaluation model.

[0079] This protocol aims to provide an objective, standardized method for comprehensively calculating the clinical importance of corrected discrepancies, addressing the difficulty in quantifying and effectively aggregating qualitative information during the assessment process, thereby improving the accuracy and reliability of the assessment. Specifically, the corrected discrepancy type and structured correction reason are first converted into a quantitative importance value. This step addresses the issue of converting originally qualitative or semi-qualitative information into a calculable value, laying the foundation for subsequent comprehensive calculations and ensuring the objectivity of the assessment. Subsequently, the quantitative importance values ​​are weighted according to the assessment indicator weights. These weights reflect the importance or impact of the corrected discrepancy across different clinical dimensions. This weighting ensures that the relative importance of different clinical dimensions is fully considered when calculating clinical importance, making the final assessment more consistent with clinical needs. Finally, the weighted quantitative importance values ​​are aggregated to produce a single, comprehensive value representing the clinical importance of the corrected discrepancy. This aggregation approach integrates previously dispersed, multi-dimensional information into a unified metric, facilitating pathologists' rapid understanding and comparison of the severity of different corrected discrepancies. The implementation of this program has shifted the revised difference assessment of early invasive lesion areas from qualitative judgment to quantitative analysis, providing an objective and quantitative basis for subsequent pathologist discussions, consensus formation, or automated fusion processing, thereby improving the efficiency and accuracy of the entire analysis process.

[0080] In one embodiment, when converting the correction difference type and the structured correction reason into the corresponding quantitative importance value, an initial importance score can be preset for each correction difference type (e.g., "boundary adjustment", "classification modification", "addition", "deletion") and stored in a lookup table. For example, "boundary adjustment" can be assigned 3 points, and "classification modification" can be assigned 5 points. Similarly, for the structured correction reasons (e.g., "lesion atypicality", "boundary fuzziness", "cell atypia"), corresponding scores can also be preset, for example, "lesion atypicality" is 4 points, and "boundary fuzziness" is 2 points. These scores can be determined based on clinical expert consensus or historical data analysis. When weighting the quantitative importance value according to the evaluation index weight, the evaluation index weight for different clinical dimensions can be first obtained. The method for obtaining these weights is referred to above. Then, the previously obtained quantitative importance value is multiplied by the corresponding evaluation index weight to obtain a weighted value (each quantitative importance value corresponds to the evaluation index weight of each clinical dimension, and a corresponding weighted quantitative importance value is obtained). When aggregating the weighted quantitative importance values, all weighted quantitative importance values ​​can be simply summed or calculated using a weighted average to obtain a final composite value of the clinical importance of the modified difference. For example, the final clinical importance can be expressed as the sum of (the quantitative value of the modified difference type multiplied by the corresponding weight) plus (the quantitative value of the structured modified reason multiplied by the corresponding weight).

[0081] refer to Figure 2 The present application provides an artificial intelligence-based lung cancer pathology section analysis system, which identifies early invasive lesions in lung cancer pathology sections based on an artificial intelligence model. The system includes: Image acquisition module 1, used to acquire a full slide image of a lung cancer pathology section (for details, please refer to step A1 above); Image analysis module 2, configured to analyze the whole slide image based on an artificial intelligence model, identify early invasive lesion areas, and generate visual information based on the identification of the early invasive lesion areas (for the specific process, please refer to step A2 above); Correction feedback module 3 is used to receive correction operations performed by multiple pathologists on the early invasive lesion area based on the identification basis visualization information, and obtain correction reasons corresponding to the correction operations to obtain multiple correction results (the specific process can be referred to step A3 above); The consistency analysis module 4 is used to perform consistency analysis on the correction results to generate final early invasive lesion area recognition result information (the specific process can be referred to the above step A4); The model training module 5 is used to perform incremental training on the artificial intelligence model based on the final early invasive lesion area recognition result information (the specific process can be referred to step A5 above).

[0082] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, various modifications and variations of the present application are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A lung cancer pathology section analysis method based on artificial intelligence, which identifies early invasive lesions in lung cancer pathology sections based on an artificial intelligence model, characterized in that: The steps of the method include: A1. Obtain full slide images of lung cancer pathology sections; A2. Analyze the whole slide image based on an artificial intelligence model to identify early invasive lesion areas and generate visualization information based on the identification of the early invasive lesion areas; A3. Receive correction operations performed by multiple pathologists on the early invasive lesion region based on the identification visualization information, and obtain correction reasons corresponding to the correction operations to obtain multiple correction results; A4. performing consistency analysis on the correction results to generate final early invasive lesion area identification result information; A5. Incrementally train the artificial intelligence model based on the final early invasive lesion area identification result information.

2. The method for analyzing lung cancer pathological sections based on artificial intelligence according to claim 1, characterized in that: Step A2 includes: A201. Analyze the whole slide image based on the artificial intelligence model to identify early invasive lesion areas and obtain a preliminary identification image including the boundaries of the early invasive lesion areas; A202. Extracting response information of the artificial intelligence model to each region in the whole slide image during the process of identifying the early invasive lesion region; A203. Based on the response information, an attention map is generated and superimposed on the preliminary identification image to form a graphical identification image, which serves as visual information for identifying the early invasive lesion area; the color depth of the attention map indicates the degree of attention of the artificial intelligence model to each area in the whole slide image.

3. The method for analyzing lung cancer pathological sections based on artificial intelligence according to claim 2, wherein: Step A203 includes: determining, based on the response information, a contribution value of each pixel region in the full slide image to an early invasive lesion region recognition result; Normalizing the contribution values, and selecting corresponding colors from a preset color gradient according to the normalized contribution values ​​to generate the attention map; The focus map is superimposed on the preliminary identification image to form a graphical identification image, which serves as visual information for identifying the early invasive lesion area.

4. The method for analyzing lung cancer pathological sections based on artificial intelligence according to claim 1, wherein: Step A3 includes: A301. Displaying the identification basis visualization information to the pathologist as reference information for performing correction operations on the early invasive lesion area; A302. Receive the pathologist's correction operation on the early invasive lesion region; the correction operation includes at least one of adding, deleting, adjusting the boundary, and modifying the classification of the early invasive lesion region; A303. According to the type of correction operation and the image features of the corresponding correction area, a list of candidate correction reasons related to the correction operation is extracted from a preset pathological feature vocabulary and displayed to the pathologist; A304. Receive the correction reason selected by the pathologist from the candidate correction reason list, and / or receive the supplementary correction reason input by the pathologist; A305. Normalize the selected correction reason and / or the input supplementary correction reason to obtain a structured correction reason; A306. Associate the structured correction reason with the correction operation and record it to obtain a correction result.

5. The method for analyzing lung cancer pathological sections based on artificial intelligence according to claim 4, characterized in that: Step A303 includes: Preliminarily extracting a list of candidate correction reasons related to the correction operation from a preset pathological feature vocabulary library according to the type of the correction operation and the image features of the corresponding correction area; Based on the image features of the correction area, analyzing the degree of correlation between each cause in the list of candidate correction reasons initially extracted and the image features, and obtaining a feature correlation degree of each candidate correction cause; Obtaining co-occurrence frequency information of correction reasons of early invasive lesion areas with similar image features to the correction area in historical correction data; Prioritizing or grouping the initially extracted candidate correction reason list based on the feature correlation and the co-occurrence frequency information to obtain an optimized candidate correction reason list; The optimized list of candidate revision reasons is displayed to the pathologist.

6. The method for analyzing lung cancer pathological sections based on artificial intelligence according to claim 5, characterized in that: The step of prioritizing or grouping the initially extracted candidate correction reason list based on the feature correlation and the co-occurrence frequency information to obtain an optimized candidate correction reason list includes: determining statistical sparsity of image features of the correction region; Determining, based on the statistical sparsity, contribution weights of the feature relevance and the co-occurrence frequency information in generating the optimized candidate correction reason list; According to the contribution weights, combined with the feature relevance and the co-occurrence frequency information, the initially extracted candidate correction reason list is prioritized or grouped to obtain the optimized candidate correction reason list.

7. The method for analyzing lung cancer pathological sections based on artificial intelligence according to claim 4, characterized in that: Step A4 includes: A401. Compare the boundary information and classification information of each early invasive lesion region in the multiple correction results, identify the early invasive lesion region with correction differences, and determine the type of the correction differences; A402. Extracting, based on the correction difference type, the structured correction reason corresponding to the early invasive lesion region having the correction difference; A403. Assess the clinical importance of the revised difference based on the revised difference type and the extracted structured revision reason; A404. Based on the clinical significance of the revised differences, displaying the early invasive lesion regions with revised differences and the corresponding extracted structured reasons for revisions to the multiple pathologists to assist the pathologists in discussion; A405. Receive the consensus opinion formed after the discussion among the pathologists, or fuse the multiple correction results according to the preset fusion rules and the clinical importance of the correction differences to generate the final early invasive lesion area identification result information.

8. The method for analyzing lung cancer pathological sections based on artificial intelligence according to claim 7, characterized in that: Step A403 includes: Obtain pathological characteristics of early invasive lesions with corrected differences and clinical background information of the corresponding patients; Determining the evaluation indicator weights of the corrected difference in different clinical dimensions from a preset clinical evaluation rule library based on the pathological feature information and the clinical background information; Based on the type of correction difference, the structured correction reason and the evaluation index weight, the clinical importance of the correction difference is comprehensively calculated.

9. The method for analyzing lung cancer pathological sections based on artificial intelligence according to claim 8, characterized in that: The step of comprehensively calculating the clinical importance of the revised difference based on the revised difference type, the structured revision reason, and the evaluation indicator weight includes: Converting the correction difference type and the structured correction reason into corresponding quantitative importance values ​​respectively; Performing weighted processing on the quantitative importance value according to the evaluation index weight; The weighted quantitative importance values ​​are aggregated to obtain the clinical importance of the corrected difference.

10. An artificial intelligence-based lung cancer pathology section analysis system, which identifies early invasive lesions in lung cancer pathology sections based on an artificial intelligence model, characterized in that: The system includes: An image acquisition module, used to acquire full-slide images of lung cancer pathological sections; an image analysis module for analyzing the whole slide image based on an artificial intelligence model, identifying early invasive lesion areas, and generating visual information based on the identification of the early invasive lesion areas; a correction feedback module, configured to receive correction operations performed by multiple pathologists on the early invasive lesion region based on the identification basis visualization information, and obtain correction reasons corresponding to the correction operations to obtain multiple correction results; A consistency analysis module, configured to perform consistency analysis on the correction results to generate final early invasive lesion area recognition result information; The model training module is used to perform incremental training on the artificial intelligence model based on the final early invasive lesion area identification result information.

Images