An adaptive tree framework-based lung adenocarcinoma pathological image auxiliary diagnosis system and method

Abstract

A pathological image-assisted diagnostic system and method for lung adenocarcinoma based on an adaptive tree framework includes: using a ViT encoder pre-trained with a pathological baseline model to classify tissue regions in whole-slide images and output histological subtype results that conform to the classification criteria for lung adenocarcinoma; calculating the statistical proportions of key pathological features such as the proportion of elegans regions, acinar regions, and high-grade component regions; using a two-layer XGBoost decision tree, combined with the WHO lung adenocarcinoma diagnostic threshold and the IASLC differentiation grading criteria, to complete AIS / MIA / IA subtyping and IA differentiation grading; relying on electronic review feedback from pathology experts, and combining Bayesian optimization to fine-tune the threshold, calibrating the consistency between the model and clinical standards; and interacting with the hospital's PACS system to retrieve patient pathological data, generating a structured report containing treatment recommendations and transmitting it to the EHR, thus achieving a closed-loop medical process.

Description

Technical Field

[0001] This invention relates to computational pathology and artificial intelligence-assisted diagnostic technology, and in particular to a lung adenocarcinoma pathological image-assisted diagnostic system and method based on an adaptive tree framework. Background Technology

[0002] Lung adenocarcinoma (LUAD) is the most common subtype of lung cancer, exhibiting significant morphological heterogeneity, posing a significant challenge to clinicopathological diagnosis. This is especially true in scenarios where pathological data is stored in multiple hospitals and cannot be integrated with hospital pathology information systems (PACS) and electronic health records (EHRs), further exacerbating diagnostic efficiency and data reusability issues. Currently, traditional manual pathological diagnosis relies on visual inspection, which suffers from the following problems: low consistency in pathological diagnosis of lung adenocarcinoma; its differentiation and grading are time-consuming and easily influenced by subjective experience, potentially leading to misjudgments in surgical planning (e.g., excessive lymph node dissection in adenocarcinoma in situ (AIS), or missed diagnosis of high-grade components in invasive adenocarcinoma (IA); and manual entry of diagnostic results into the system makes rapid synchronization to multidisciplinary consultation platforms difficult, posing a risk of delayed treatment decisions. This presents a challenge to accurate histological classification and grading in routine pathological diagnosis. Although some studies have attempted to use deep learning for automated analysis, existing methods are generally insufficient in capturing fine-grained subtypes, and the decision-making process lacks clinical interpretability. They cannot be deeply integrated with the WHO classification guidelines for lung adenocarcinoma and the actual "sampling-diagnosis-reporting-archiving" workflow in pathology departments, nor can they be connected to hospital laboratory information systems to achieve linkage between test data and imaging diagnosis, making it difficult for the technology to be implemented in clinical practice.

[0003] LUAD accounts for 40%-55% of lung cancer cases, and its prognosis is closely related to its histological subtype and differentiation status. The 5-year survival rate for adenocarcinoma in situ (AIS) and minimally invasive adenocarcinoma (MIA) is close to 100%, while the prognosis of invasive adenocarcinoma (IA) depends on the degree of differentiation. Accurate subtyping and grading directly determine clinical treatment pathways—for example, AIS requires no adjuvant therapy after surgery, MIA has an excellent prognosis after surgical resection and usually requires no adjuvant therapy but needs follow-up every 6 months post-surgery, and poorly differentiated IA requires adjuvant chemotherapy after surgery. These are core decision-making bases for developing treatment strategies, predicting patient outcomes, and allocating medical resources. However, the morphological heterogeneity of LUAD leads to the following diagnostic challenges: overlapping features between subtypes (such as acinar and papillary types), the complexity of the tumor microenvironment, and diagnostic inconsistencies caused by intraclass variations. Differences in the application of diagnostic standards by pathology departments at different hospital levels result in insufficient diagnostic consistency between MIA and IA, which may lead to disagreements in clinical treatment strategies (such as the extent of surgery and the choice of adjuvant therapy), thus significantly increasing the patient's medical burden and treatment cycle.

[0004] Accurate histological subtyping and differentiation grading play a crucial role in guiding treatment strategies and predicting patient prognosis. However, the morphological heterogeneity of LUAD, driven by genetic, molecular, and environmental factors, presents significant diagnostic challenges. On the one hand, interclass similarity between IA subtypes often blurs diagnostic boundaries because several patterns share overlapping morphological features. The presence of a complex tumor microenvironment further exacerbates this challenge. On the other hand, intraclass variability is a major obstacle even for experienced pathologists, leading to inconsistencies in grading and classification. Furthermore, manual diagnostic results are often recorded in unstructured text form in pathology reports, failing to automatically link to patients' past genetic testing data and treatment history, making it difficult to support personalized treatment recommendations and failing to meet the needs of precision medicine for the integration of multimodal medical data.

[0005] Artificial intelligence (especially deep learning) has shown the potential to overcome the limitations of traditional pathological diagnosis. Existing methods are mainly divided into two categories: (1) region-level models (such as CNN and ViT) identify local features by analyzing high-resolution image patches; (2) slide-level models (such as multi-instance learning MIL) achieve global diagnosis by aggregating regional information. However, these methods focus on image algorithm optimization and do not consider the actual workflow of the pathology department (such as doctors' review and annotation of suspicious areas and electronic signature of reports), and also lack the ability to interact with the hospital's HIS (Hospital Information System) / LIS system. As a result, the model prediction results need to be manually processed again before they can be converted into clinically usable diagnostic reports, and it is impossible to form a medical data closed loop of "data collection-intelligent analysis-clinical application-feedback optimization". The morphological complexity of LUAD still brings unique challenges: local microstructural differences need to be identified, global features need to be integrated, and the grading process needs to selectively integrate key regional information. The histological subtypes of LUAD show significant heterogeneity at both the spatial and tissue levels. Locally, disease progression manifests as diverse cell and subtissue morphologies, requiring the model to learn compact and discriminative features from a limited field of view. At the global level, accurate diagnosis relies on the model's ability to integrate heterogeneous regional subtypes to infer a coherent and clinically meaningful classification. Furthermore, grading the differentiation of IA requires selectively aggregating regional subtyping results—prioritizing diagnostically relevant patterns while minimizing noise and redundancy. These challenges highlight the need for advanced AI-driven diagnostic frameworks that simultaneously address the fine-grained morphological diversity of LUAD, achieve seamless integration with electronic pathology systems and multidisciplinary consultation platforms, provide reliable and interpretable predictions for clinical decision-making, support the structured management and research reuse of pathology data, and ultimately be implemented as clinical support tools compliant with medical informatics standards to support the clinical pathology diagnostic process.

[0006] It should be noted that the information disclosed in the background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The main objective of this invention is to overcome the shortcomings of the aforementioned background technologies: the consistency of artificial pathological diagnosis of lung adenocarcinoma is low, its differentiation and grading are time-consuming and easily influenced by subjective experience, existing deep learning methods are insufficient in capturing fine-grained subtypes, and the decision-making process lacks clinical interpretability. It also addresses the problems of scattered storage of pathological data across multiple hospital campuses, inability to link with hospital pathology information systems (PACS) and electronic health records (EHR), and the lack of existing technologies to interface with the pathology department's "sampling-diagnosis-reporting-archiving" workflow and hospital HIS / LIS systems, making it difficult to form a closed loop of medical data. This invention provides a lung adenocarcinoma pathological image-assisted diagnostic system and method based on an adaptive tree framework, enabling specialized processing of lung adenocarcinoma pathological data, ultimately becoming a clinical auxiliary tool that conforms to medical informatics standards, supporting a complete medical closed loop from pathological data input to clinical diagnosis and treatment decision support.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect of the present invention, a method for assisted diagnosis of lung adenocarcinoma based on adaptive tree framework pathological images includes the following steps:

[0010] S1. Regional Histological Subtype Classification: The ViT encoder, a visual transformer pre-trained based on a pathological model, is used to perform fine-grained classification of tissue regions in whole slide images and output regional prediction results of multiple histological subtypes according to the classification criteria for lung adenocarcinoma.

[0011] S2. Slide-level feature extraction: Based on the regional prediction results, calculate the statistical proportions of key pathological features used for lung adenocarcinoma subtype classification and differentiation grade grading, including the proportion of lepidic regions, the proportion of high-grade regions, and the proportion of acinar regions.

[0012] S3. Slide-level hierarchical decision-making: The statistical proportions of the key pathological features are input into the XGBoost-based adaptive decision-making module, and the decision is executed sequentially through a two-layer decision tree structure:

[0013] First decision tree (Tree1): Based on the proportion of adherent regions and the World Health Organization's diagnostic threshold for lung adenocarcinoma, slides are classified into adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), or invasive adenocarcinoma (IA).

[0014] The second decision tree (Tree2) classifies the degree of differentiation for slides identified as invasive adenocarcinoma (IA) based on the proportion of high-grade regions, the ratio of acinar to lepidic regions, and the differentiation grading criteria of the International Association for the Study of Lung Cancer (IASLC). A high-grade region proportion ≥20% is output as poorly differentiated, acinar predominance and high-grade region <20% is output as moderately differentiated, and lepidic predominance and high-grade region <20% is output as well-differentiated.

[0015] S4. Clinical Feedback and Adaptive Optimization: Pathology experts review the slide-level prediction results output by the model through the electronic pathology system and provide clinical feedback. Based on this clinical feedback, the classification threshold of the decision tree is iteratively fine-tuned to improve the consistency between the model's prediction results and clinical diagnostic standards. The threshold adjustment process combines the optimal hyperparameter configuration output by Bayesian optimization, calibrates the model's decision boundary through clinical standards, and iteratively optimizes the consistency between the model's prediction results and the pathology diagnostic report specifications.

[0016] S5. Diagnostic Report Generation and Clinical System Interaction: The subtype classification, differentiation degree, and corresponding clinical stage information output by the model are interacted with the hospital pathology information system (PACS) to retrieve the patient's associated pathological data, generating a structured pathological diagnostic report containing subtype classification, differentiation degree, corresponding clinical stage, and treatment recommendations; the structured pathological diagnostic report is then transmitted to the hospital's electronic medical record system (EHR), completing a complete medical closed-loop operation from the input of HE-stained whole slide image data in the pathology department, intelligent analysis, diagnostic report generation to clinical diagnosis and treatment decision support.

[0017] Furthermore, the integration process based on feedback from pathology experts includes:

[0018] The initial slide-level prediction results were reviewed by pathology experts;

[0019] Update the decision threshold based on the review results and refit the XGBoost classifier;

[0020] Repeat the iterations until the predicted results are consistent with clinical standards.

[0021] In a second aspect of the present invention, a lung adenocarcinoma pathological image-assisted diagnostic system based on an adaptive tree framework is provided for performing the aforementioned lung adenocarcinoma pathological image-assisted diagnostic method based on an adaptive tree framework, comprising:

[0022] The image processing module is used to perform regional-level histological subtyping, slide-level feature extraction, and hierarchical decision-making, enabling fine-grained classification and diagnosis of whole slide images;

[0023] The data interface module is used to communicate with the hospital's PACS and EHR systems, retrieve patient WSI data and clinical information, and send back structured diagnostic reports.

[0024] The report generation module is used to generate a preliminary structured pathological diagnosis report that conforms to the HL7 FHIR standard based on the output of the image processing module.

[0025] The doctor review module provides a human-computer interaction interface for pathologists to review, modify, and electronically sign the structured pathology diagnosis reports.

[0026] The system integrates with the hospital's pathology information system through a protocol port, enabling a closed-loop operation throughout the entire process, from pathological image input, intelligent analysis, report generation to clinical decision support.

[0027] Furthermore, the report generation module is configured to automatically fill the predefined structured report template with the multi-dimensional diagnostic elements output by the image processing module to generate the preliminary structured pathological diagnosis report; the multi-dimensional diagnostic elements include one or more of the following: dominant histological subtype classification, differentiation grade determination, percentage of each subtype, proportion of high-grade components, and maximum diameter of infiltrative lesions;

[0028] The data interface module is also configured to: conduct secure and automated data interaction with the hospital's medical record information system through the data integration interface deployed by the hospital;

[0029] The auxiliary diagnostic system realizes a complete closed-loop management of medical care, from pathological data input, intelligent analysis, diagnostic report generation to clinical diagnosis and treatment decision support.

[0030] The present invention has the following beneficial effects:

[0031] This invention proposes a pathological image-assisted diagnostic system and method for lung adenocarcinoma based on an adaptive tree framework. The FigATree (fine-grained adaptive tree) framework is designed, which is a clinically interpretable artificial intelligence pathological diagnostic system used to support the pathological subtype classification and differentiation degree grading of LUAD. The FigATree framework in this invention combines a regional encoder based on a base model with an adaptive pathology decision module based on XGBoost, and the technical logic is fully integrated with medical standards: At the regional level, a ViT encoder pre-trained based on a pathology base model is used, whose classification results match the histological subtypes (adherent, acinar, etc.) in the lung adenocarcinoma classification criteria. This overcomes the bias of natural image pre-training while ensuring consistency with clinical diagnostic criteria. It captures subtle morphological cues at the cellular and subtissue levels through a self-attention mechanism, generating feature vectors to encode local and global morphological features, achieving fine-grained classification of multiple histological subtypes and addressing the shortcomings of traditional models in capturing fine-grained subtypes. At the slide level, an adaptive decision module is constructed using two layers of XGBoost decision trees. Tree1 classifies AIS, MIA, and IA subtypes based on the proportion of adherent regions and the World Health Organization's lung adenocarcinoma diagnostic threshold. Tree2 combines the proportion of high-level regions and acinar / The percentage of apical regions was used to classify the differentiation degree of IA according to the IASLC differentiation grading criteria (≥20% of high-grade regions indicates low differentiation, <20% of high-grade regions with predominant acini indicates moderate differentiation, and <20% of high-grade regions with predominant apical regions indicates high differentiation). This ensured that the classification and grading results fully conformed to clinical diagnostic standards. At the same time, Bayesian optimization was used to adjust hyperparameters, balance regularization terms and loss functions to enhance generalization ability, and integrate pathological prior knowledge with data-driven learning to avoid the limitations of traditional linear models.

[0032] The system transforms qualitative diagnosis into a quantitative paradigm, providing detailed subtype counts and proportions at the regional level and generating an interpretable diagnostic process at the slide level, reducing subjectivity. Furthermore, it iterative optimization of decision thresholds through feedback from pathology experts enhances the consistency between model predictions and clinical standards. This achieves effective feature fusion and accurate diagnosis from the regional to the slide level, providing a novel technical approach for precise pathological diagnosis and clinical treatment decisions in LUAD.

[0033] This invention constructs a comprehensive and standardized database of H&E-stained whole slide images (WSI) for lung pathology. Based on this, 1,186 WSI images from LUAD cases at the Department of Pathology, Fujian Provincial Hospital Affiliated to Fuzhou University were used for training. Comprehensive evaluation showed that: (1) at the regional level, FigATree achieved a classification accuracy of over 95% for the six histological growth patterns; (2) at the slide level, its subtype classification accuracy reached 90%, and its differentiation grade grading accuracy reached 85%, significantly surpassing traditional pathological rule-based methods; (3) the framework exhibits high interpretability at both the regional and slide levels, and its prediction results are consistent with pathological standards and expert evaluations.

[0034] The system of this invention communicates with hospital PACS and EHR systems through a data interface module, automatically retrieving patient WSI data and previous clinical information (such as gene testing reports and historical pathology results) without the need for manual secondary processing. The report generation module can generate structured pathological diagnosis reports conforming to the HL7FHIR standard, and connects with the hospital's electronic medical record system through a protocol port to directly provide clinical treatment recommendations (such as AIS and MIA recommending regular close follow-up; high / moderately differentiated IA requiring comprehensive assessment of high-risk factors to determine whether adjuvant chemotherapy is necessary; low-differentiated IA has a high risk of recurrence and recommends postoperative adjuvant chemotherapy to reduce the probability of recurrence and metastasis). By combining basic model capabilities and clinical decision structures, FigATree provides a precise and reliable tool for LUAD diagnosis, significantly reducing diagnostic time and forming a complete medical data closed loop of pathological data acquisition, intelligent analysis, report generation, clinical application, and expert feedback optimization. This avoids the shortcomings of existing technologies that only optimize algorithms but cannot be applied clinically, directly outputting structured reports to support treatment decisions and establishing a generalizable paradigm for computational pathology in lung cancer and other fields.

[0035] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0036] Figures 1(a) to 1(c) are overview diagrams of the FigATree framework of the method of the embodiment of the present invention.

[0037] Figures 2(a) to 2(j) show the regional histological subtype classification performance and visualization results of the embodiments of the present invention.

[0038] Figure 3 This diagram illustrates the impact of different region sizes on classification performance in an embodiment of the present invention.

[0039] Figures 4(a) to 4(i) show the impact of different dataset sizes on model performance in the embodiments of the present invention.

[0040] Figure 5 This is a slide-level diagnostic result and visualization diagram from an embodiment of the present invention.

[0041] Figure 6 This is a diagram of the ViT architecture of the regional predictor in an embodiment of the present invention.

[0042] Figure 7 This is a flowchart illustrating the attention mechanism calculation process in the ViT encoder according to an embodiment of the present invention.

[0043] Figure 8 This is a structural block diagram of the lung adenocarcinoma pathological image-assisted diagnostic system based on an adaptive tree framework according to the present invention.

[0044] Figure 9 This is a diagram showing the integrated architecture of the diagnostic system and data system of this invention. Detailed Implementation

[0045] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0046] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0047] This invention provides a method for assisted diagnosis of lung adenocarcinoma based on adaptive tree framework pathological images, comprising the following steps:

[0048] Step S1, Regional Histological Subtype Classification: The ViT encoder, a vision transformer pre-trained based on a pathological model, is used to perform fine-grained classification of tissue regions in the whole slide image, and output regional prediction results of multiple histological subtypes according to the classification criteria for lung adenocarcinoma.

[0049] In some embodiments, in step S1: the pathological basic model initializes the ViT encoder weights through pre-training with a large-scale pathological image corpus; the regional predictor is fine-tuned by adding a linear classifier head and a cross-entropy loss function to generate feature vectors to encode the local and global morphological features of the tissue region.

[0050] In some embodiments, the multiple histological subtypes include the following histological subtypes: lepidic, acinar, papillary, high-grade, invasive mucinous adenocarcinoma (IMA), and normal tissue; the high-grade subtype is further subdivided into solid, micropapillary, and cribriform / complex glandular structures.

[0051] Step S2, Slide-level feature extraction: Based on the regional prediction results, calculate the statistical proportions of key pathological features used for lung adenocarcinoma subtype classification and differentiation grade grading, including the proportion of lepidic regions, the proportion of high-grade regions, and the proportion of acinar regions.

[0052] In some embodiments, in step S2: the statistical proportion of the key pathological features is derived from the spatial aggregation of regional prediction results; the regional size is set according to a preset resolution specification and extracted at a predetermined magnification.

[0053] Step S3, Slide-level hierarchical decision-making: The statistical proportions of the key pathological features are input into the XGBoost-based adaptive decision-making module, and the decision is executed sequentially through a two-layer decision tree structure:

[0054] First decision tree (Tree1): Based on the proportion of adherent regions and the World Health Organization's diagnostic threshold for lung adenocarcinoma, slides are classified into adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), or invasive adenocarcinoma (IA).

[0055] The second decision tree (Tree2) classifies the degree of differentiation for slides identified as invasive adenocarcinoma (IA) based on the proportion of high-grade regions, the ratio of acinar to lepidic regions, and the differentiation grading criteria of the International Association for the Study of Lung Cancer (IASLC). A high-grade region proportion ≥20% is output as poorly differentiated, acinar predominance with high-grade regions <20% is output as moderately differentiated, and lepidic predominance with high-grade regions <20% is output as highly differentiated.

[0056] In some embodiments, in step S3:

[0057] The two-layer decision tree structure is implemented using the XGBoost model, whose objective function includes a regularization term to prevent overfitting, specifically:

[0058] Parameters that control the number of nodes in the decision tree can limit the tree depth;

[0059] L2 regularization is used to constrain leaf node values ​​to enhance generalization ability.

[0060] In some embodiments, the objective function of the XGBoost model is:

[0061]

[0062] in Let l be the objective function of the t-th iteration, l be the loss function, and n be the number of samples i. It is the true label of sample i. It is the predicted value for sample i in the previous round. It is the prediction output of the t-th decision tree for sample i. It is a regularization term. For constant terms;

[0063] Where the loss function The specific form is:

[0064]

[0065] It is the predicted probability that sample i belongs to category j.

[0066] The regularization term Calculate according to the following formula:

[0067]

[0068] in It is a parameter that controls the number of leaf nodes to prevent the tree structure from becoming too deep and causing overfitting. It is the penalty coefficient; It is an L2 regularization term that controls the size of the leaf node values ​​to enhance generalization ability. It is the output value of the j-th leaf node;

[0069] The hyperparameters in the XGBoost model are adjusted using Bayesian optimization to construct the performance profile of the surrogate model's approximate hyperparameters and the objective function optimization effect. Furthermore, the model explores and utilizes the acquisition function's balancing mechanism to search for regularization terms. The parameter set that achieves an optimal balance with the loss function l is used to enhance the model's generalization ability.

[0070] In some embodiments, the training of the ViT encoder, which is pre-trained based on the pathological model, employs a data augmentation strategy involving horizontal and vertical flipping; the training of the XGBoost-based adaptive decision module employs an adaptive gradient optimization algorithm, and the model is updated with optimized learning rate and batch size parameters.

[0071] Step S4, Clinical Feedback and Adaptive Optimization: Pathology experts review the slide-level prediction results output by the model through the electronic pathology system and provide clinical feedback. Based on this clinical feedback, the classification threshold of the decision tree is iteratively fine-tuned to improve the consistency between the model's prediction results and clinical diagnostic standards. The threshold adjustment process combines the optimal hyperparameter configuration output by Bayesian optimization, calibrates the model's decision boundary through clinical standards, and iteratively optimizes the consistency between the model's prediction results and the pathology diagnostic report specifications.

[0072] In some embodiments, step S4 includes the following integration process for pathology expert feedback: initial slide-level prediction results are reviewed by pathology experts; decision thresholds are updated based on the review results, and the XGBoost classifier is refitted; the process is repeated iteratively until the prediction results are consistent with clinical standards.

[0073] Step S5: Diagnostic Report Generation and Clinical System Interaction: The subtype classification, differentiation degree, and corresponding clinical stage information output by the model are interacted with the hospital pathology information system (PACS) to retrieve the patient's associated pathological data, generating a structured pathological diagnostic report containing subtype classification, differentiation degree, corresponding clinical stage, and treatment recommendations; the structured pathological diagnostic report is then transmitted to the hospital's electronic medical record system (EHR), completing a complete medical closed-loop operation from the input of HE-stained whole slide image data in the pathology department, intelligent analysis, diagnostic report generation to clinical diagnosis and treatment decision support.

[0074] like Figure 8 As shown, this embodiment of the invention also provides a lung adenocarcinoma pathological image-assisted diagnostic system based on an adaptive tree framework, used to execute the aforementioned lung adenocarcinoma pathological image-assisted diagnostic method based on an adaptive tree framework, including:

[0075] The image processing module is used to perform regional-level histological subtyping, slide-level feature extraction, and hierarchical decision-making, enabling fine-grained classification and diagnosis of whole slide images;

[0076] The data interface module is used to communicate with the hospital's PACS and EHR systems, retrieve patient WSI data and clinical information, and send back structured diagnostic reports.

[0077] The report generation module is used to generate a preliminary structured pathological diagnosis report that conforms to the HL7 FHIR standard based on the output of the image processing module.

[0078] The doctor review module provides a human-computer interaction interface for pathologists to review, modify, and electronically sign the structured pathology diagnosis reports.

[0079] The system integrates with the hospital's pathology information system through a protocol port, enabling a closed-loop operation throughout the entire process, from pathological image input, intelligent analysis, report generation to clinical decision support.

[0080] In some embodiments, the report generation module is configured to: automatically fill the predefined structured report template with the multi-dimensional diagnostic elements output by the image processing module to generate the preliminary structured pathological diagnosis report; the multi-dimensional diagnostic elements include one or more of the following: dominant histological subtype classification, differentiation grade determination, percentage of each subtype, proportion of high-grade components, and maximum diameter of infiltrative lesions;

[0081] The data interface module is also configured to: conduct secure and automated data interaction with the hospital's medical record information system through the data integration interface deployed by the hospital;

[0082] The auxiliary diagnostic system realizes a complete closed-loop management of medical care, from pathological data input, intelligent analysis, diagnostic report generation to clinical diagnosis and treatment decision support.

[0083] The specific implementation includes the following core steps: A ViT encoder, pre-trained using a pathological baseline model, performs fine-grained classification of tissue regions in whole-slide images. The system automatically extracts multi-dimensional diagnostic elements, including dominant histological subtype classification, differentiation grade determination, and key quantitative indicators (such as percentage of each subtype, proportion of high-grade components, and maximum diameter of infiltrating lesions). These elements are then automatically populated into a predefined structured report template to generate a preliminary pathological diagnosis report conforming to clinical standards. The core content of this report includes: subtype classification, differentiation grade, corresponding clinical stage, and treatment recommendations. It also enables secure and automated data interaction with the hospital's medical record information system through a data integration interface module deployed in the hospital. This achieves a complete closed-loop management system for medical care, from pathological data input to clinical decision support.

[0084] This invention proposes an image-assisted diagnostic system and method for lung adenocarcinoma based on an adaptive tree framework. By designing the FigATree framework, it effectively addresses the diagnostic challenges posed by the morphological heterogeneity of lung adenocarcinoma. This framework combines a regional encoder based on a pathological baseline model with an adaptive pathological decision module based on XGBoost, achieving end-to-end subtyping and grading. The regional encoder based on the pathological baseline model, ViT-L, overcomes natural image pre-training bias and captures subtyping discriminative features through a self-attention mechanism, achieving a classification accuracy exceeding 95% for six tissue-based growth patterns. The slide-level XGBoost... Decision trees integrate prior pathological knowledge and, through an adaptive threshold mechanism, avoid the limitations of traditional linear models. Subtype classification accuracy reaches 90%, and differentiation grade accuracy reaches 85%, significantly surpassing traditional pathological rule-based methods. Simultaneously, this framework transforms qualitative diagnosis into a quantitative paradigm, providing detailed counts and proportions of each subtype at the regional level. Slide-level typing and grading generate an interpretable diagnostic process, minimizing subjectivity and providing pathologists with comprehensive and reliable diagnostic evidence. It exhibits high interpretability at both the regional and slide levels, with prediction results consistent with pathological standards and expert assessments. The system can integrate with hospital pathology information systems (PACS) and electronic medical records (EHR) via protocol ports, automatically retrieving patient pathology data and outputting structured diagnostic reports, forming a closed-loop medical system of pathology data input, intelligent analysis, clinical application, and feedback optimization. This not only provides a precise and reliable tool for diagnosing lung adenocarcinoma but also establishes a generalizable paradigm for computational pathology in lung cancer and other fields.

[0085] The following further describes specific embodiments and experimental verifications of the present invention.

[0086] To address the shortcomings of existing technologies (such as insufficient capture of fine-grained subtypes, lack of clinical interpretability in decision-making, and difficulties in deep integration of diagnostic guidelines and actual workflows), this invention presents a lung adenocarcinoma pathological image-assisted diagnostic system and method based on an adaptive tree framework. The FigATree framework (Figures 1(a), 1(b), and 1(c)) is a novel hybrid deep learning framework specifically designed for clinical pathological diagnosis and conforming to medical informatics standards for LUAD subtyping and grading. It is a fine-grained adaptive tree capable of achieving accurate regional-level histological subtyping as well as slide-level histological subtyping and differentiation grading. This invention also constructs and uses a large dataset containing 1,186 LUADWSI images, covering three LUAD histological subtypes (AIS, MIA, IA) and three differentiation levels (high, intermediate, and low differentiation). The dataset annotation follows WHO standards and IASLC guidelines to ensure the clinical applicability of the data. Key contributions include:

[0087] An end-to-end LUAD subtype classification and grading pathological diagnosis framework was proposed, which is aligned with the WHO classification criteria for lung adenocarcinoma and the IASLC differentiation grading criteria, to achieve comprehensive regional-level histological subtype classification, slide-level histological subtype classification, and differentiation degree grading.

[0088] A region encoder based on a pathological model was developed to avoid the pathological feature capture bias caused by pre-training on natural images, and to achieve efficient region encoding and accurate fine-grained subtype classification.

[0089] An adaptive slide decision module based on XGBoost was designed. The decision threshold is dynamically adjusted according to clinical guidelines to achieve effective feature fusion from region to slide. The output results can be directly mapped to clinical staging to support the selection of treatment plan.

[0090] This invention constructs a comprehensive and standardized H&E staining WSI database for lung pathology, which is directly compatible with hospital pathology information systems (PACS). The superiority of the clinical case validation method ensures that the technology can be applied in clinical practice.

[0091] The implementation process of the core model construction and experimental verification of the pathological image-assisted diagnosis of lung adenocarcinoma in this method is described in detail below.

[0092] Dataset

[0093] Source: 111 WSI (regional level) images from 94 patients at the Provincial Hospital Affiliated to Fuzhou University, and 1075 WSI (slide level) images from 714 patients. The regional cohort contains 90,083 labeled regions, covering six major histological subtypes of lung adenocarcinoma in situ, minimally invasive adenocarcinoma, and invasive adenocarcinoma (lepidic, acinar, papillary, solid, micropapillary, and cribriform / complex adenoid structures).

[0094] Labelling standards: Following WHO 5th edition and the International Association for the Study of Lung Cancer guidelines. Tissue slides were digitized using the TEKSQRAY imaging system at 40x magnification.

[0095] Category distribution: Male 34.83%, Female 65.17%, AIS 18.89%, MIA 18.21%, High differentiation 15.26%, Medium differentiation IA 27.40%, Low differentiation 19.31%.

[0096] Preprocessing: Extract a 1024×1024 pixel area under 40x magnification.

[0097] FigATree frame

[0098] Regional predictor

[0099] Figure 6 The ViT architecture diagram of the region-level predictor in this embodiment of the invention is shown. The backbone is a ViT-L (24-layer Transformer) pre-trained on a pathology baseline model, with each layer equipped with a 16-head multi-head self-attention mechanism. Position embeddings are added to the input block embeddings to maintain spatial structure throughout the encoding process. In previous studies, the initial parameters of the encoder were mostly derived from pre-training on natural images, leading to a mismatch with the feature distribution of pathology images and introducing semantic ambiguity. In this invention, FigATree utilizes a baseline model pre-trained on a large-scale pathology image corpus to initialize the ViT encoder. This enhances the encoder's ability to extract discriminative features, laying the foundation for identifying significant differences between subtypes with relatively large inter-class variations and small intra-class variations.

[0100] Fine-tuning: Add a linear classification head and optimize the cross-entropy loss function.

[0101] Figure 7 This paper illustrates the attention mechanism calculation process in the ViT encoder according to an embodiment of the present invention. The attention mechanism within each Transformer layer operates as follows:

[0102] (1)

[0103] Where Q, K, and V represent the query, key, and value matrices derived from the input features, and d_{k} is the dimension of the key vector.

[0104] For model optimization, the cross-entropy loss function is used to supervise subtype prediction:

[0105] (2)

[0106] Where N is the number of samples, C is the number of histological subtypes, and y_{i, c} is the true label indicator (1 or 0) for sample i belonging to category c. It is the predicted probability of class c obtained through the softmax function.

[0107] Features: 1024-dimensional vector encoding of local and global morphological features.

[0108] Slide-level classifier

[0109] Structure: A two-layer XGBoost decision tree (Tree1 subtyping → Tree2 IA grading). The first decision tree (Tree1) primarily performs subtyping based on the proportion of elegans regions; the second decision tree (Tree2) refines the grading by considering the proportion of high-level regions and the ratio of elegans to acinar regions. To make LUAD diagnosis more aligned with the diagnostic patterns of pathologists, this framework combines domain-specific pathological knowledge derived from established clinical guidelines (such as the relative proportions of growth patterns in different tissues) with data-driven learning, dynamically adjusting decision thresholds to improve classification accuracy and clinical relevance. This tree-structured ensemble model consists of multiple decision trees, each branch corresponding to a specific pathological feature (including growth pattern distribution, cell morphology, and spatial arrangement). By aggregating the outputs of all decision trees, a final prediction is generated, allowing for adaptive threshold adjustment based on the complexity and ambiguity of the input pattern.

[0110] Input features: proportion of efflorescent cells, proportion of high-level regions, proportion of acinar cells.

[0111] Optimization: Bayesian optimization adjusts hyperparameters (see Table 1), and regularization term prevents overfitting.

[0112] The objective function of the XGBoost model is:

[0113]

[0114] in Let l be the objective function of the t-th iteration, l be the loss function, and n be the number of samples i. It is the true label of sample i (such as histological subtype or differentiation category). It is the predicted value for sample i in the previous round. It is the prediction output of the t-th decision tree for sample i. It is a regularization term. For constant terms;

[0115] Where the loss function The specific form is:

[0116]

[0117] It is the predicted probability that sample i belongs to category j.

[0118] The regularization term Calculate according to the following formula:

[0119]

[0120] in It is a parameter that controls the number of leaf nodes to prevent the tree structure from becoming too deep and causing overfitting. It is the penalty coefficient; It is an L2 regularization term that controls the size of the leaf node values ​​to enhance generalization ability. It is the output value of the j-th leaf node.

[0121] Table 1 shows the details of the hyperparameters obtained by XGBoost in FigATree.

[0122] Table 1

[0123]

[0124] This structure improves robustness by learning residuals through multiple decision trees and focusing on samples that are difficult to classify, while the addition of a regularization term effectively mitigates overfitting. Compared to rule-based methods based on pathology guidelines, FigATree enables the model to adaptively update its parameters to prevent overfitting. Compared to linear fitting methods, FigATree offers more flexible parameter optimization, making it easier to reach the optimal solution.

[0125] Bayesian optimization is used to adjust the complex non-convex hyperparameters involved in the XGBoost objective function (such as...). , By constructing a surrogate model (usually a Gaussian process) to approximate the performance pattern of hyperparameters and objective function optimization, and by exploring and utilizing the balancing function, the search is guided to find the parameter set that achieves the optimal balance between the regularization term Ω(f_t) and the loss function l, thereby enhancing the model's generalization ability and preventing overfitting.

[0126] Process: Initial prediction → Expert feedback → Threshold iterative update. Expert review can be performed by experienced pulmonary pathologists to review the region of interest (ROI) prediction results generated by the model, provide clinical feedback, and ensure interpretability.

[0127] Training details

[0128] Hardware: 8×NVIDIA 3090 GPUs.

[0129] Software: PyTorch 2.5.1, CUDA 12.4. Operating system: Ubuntu 22.04. Training was accelerated using cuDNN. Parameters: 20 epochs, batch size 256, learning rate 0.0001, Adam optimizer. Augments: Horizontal / vertical flipping, all input images were resized to 1024x1024 pixels.

[0130] Experimental results

[0131] Regional histological subtyping

[0132] FigATree achieved an overall accuracy of 95.37% in the classification task of six histological subtypes (adherent, acinar, papillary, high-grade, IMA, and normal tissue) (Fig. 2(a)). It performed fine-grained classification across six different histological subtypes: adherent, acinar, papillary, high-grade, invasive mucinous adenocarcinoma, and normal tissue, significantly outperforming the ImageNet-pretrained ResNet and ViT baseline models. Its performance was particularly outstanding in identifying adherent (97.00%), acinar (91.71%), and high-grade (95.85%) subtypes, and it also showed significant improvement in the challenging papillary subtype (79.94%). The confusion matrix (Fig. 2(j)) shows that the main misclassifications occurred between the papillary and high-grade subtypes. These results demonstrate that FigATree effectively captures subtle morphological cues at the cellular and subtissue levels by utilizing region encoding based on the base model and transforms them into subtype-specific diagnostic features, thereby achieving accurate fine-grained classification. The high-grade subtype was further divided into solid type (87.60%), micropapillary type (90.30%), and cribriform / complex glandular structure (performance limited), likely due to greater morphological diversity within this subtype. Grad-CAM visualization (Fig. 2(b)–Fig. 2(f)) confirmed that the model could locate key pathological areas (such as alveolar wall tumor cells and acinar structures). In some cases (Fig. 2(e)), attention was also focused on highly differentiated cells, although irrelevant areas were occasionally highlighted. Misclassification analysis (Fig. 2(g)–Fig. 2(h)) indicated that artifacts caused by alveolar collapse were the main source of error; for example, the presence of alveolar collapse within adenocarcinoma in situ could resemble pseudoalveolar structures, such as alveolar septal hyperplasia or true alveoli (Fig. 2(g)). Furthermore, disruption of alveolar septa, whether due to disease progression or intraoperative traction, could lead to a pseudopapillary appearance, which could be confused with true papillary adenocarcinoma in the absence of a fibrovascular axis (Fig. 2(h)). Nevertheless, FigATree demonstrates strong generalization ability in diverse histological contexts and accurately identifies most tissue growth patterns.

[0133] Slide-level classification and grading

[0134] To perform slide-level histological subtyping and differentiation grading, FigATree first classifies slides into adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), and invasive adenocarcinoma (IA) based on regional classification results. For IA slides, differentiation is further graded based on the proportion of high-grade regions and the acinar / embedded ratio, resulting in high, intermediate, or low differentiation categories. At the slide level, FigATree classifies cases into AIS / MIA / IA subtypes and grades IA for differentiation (high / intermediate / low differentiation). Compared to baseline methods such as guideline-based fixed models (GFM) and linear fitting models (LFM), the thresholds in clinical guidelines are usually derived from expert consensus and prioritize interpretability over statistical robustness, making them unsuitable for direct use in deep learning frameworks. Slide-level classification involves aggregating predictions of thousands of regions, and small region-level errors can accumulate and significantly reduce accuracy. Most guideline-based systems assume a linear relationship between region features and slide-level labels, failing to capture the complex nonlinear patterns inherent in tumor heterogeneity. FigATree achieved 89.30% accuracy in subtyping (a 51.85% improvement over GFM) and 84.34% accuracy in grading (a 36.57% improvement) (Table 1). Compared to LFM, FigATree further improved subtyping accuracy by 47.86% and grading accuracy by 38.96%. Its performance was balanced across all subtypes (AIS 85.42%, MIA 86.30%, IA 91.41%). In histological subtyping, its accuracy exceeded 85% in AIS, MIA, and IA, significantly surpassing the suboptimal model LFM. Regarding grading, FigATree demonstrated superior performance in highly differentiated and moderately differentiated IA, with an accuracy approaching 85%, while GFM, although achieving 92.34% accuracy in low-differentiated types, sacrificed performance in other categories.

[0135] Table 2 compares the accuracy of the three methods in the classification of histological subtypes and the grading of differentiation at the slide level.

[0136] Table 2

[0137]

[0138] Table 3 compares the accuracy of the three methods in slide-level histological subtyping and differentiation grading (corrected to reflect actual performance). Bold values ​​indicate best performance.

[0139] Table 3

[0140]

[0141] Diagnostic enhancement and interpretability

[0142] Optimize and integrate pathology expert feedback through iterative thresholding. Figure 5 Visualize their aggregated predictions to validate the performance of results from the regional level to the whole slide, such as Figure 5 As shown, the pie chart illustrates the block-level classification distribution of representative slides, demonstrating FigATree's strong generalization ability across different cases. The decision tree structure achieves clinical interpretability, and the feedback loop iterates multiple times, significantly improving consistency with clinical decision-making criteria. The full-slide heatmap visualization shows the consistency between the model's focused area and pathological criteria (Table 2). These macroscopic visualizations provide insights into the model's decision-making patterns and enhance its interpretability and reliability in clinical diagnostic applications.

[0143] Ablation Research

[0144] Region size: Optimal performance is achieved with a 1024×1024 pixel region. Figure 3 These results demonstrate that moderately large regions enable the model to capture more comprehensive contextual features, thereby improving the encoding of LUAD pathological images. Notably, the ResNet baseline, used as a control, exhibits the worst performance at a 1024-pixel setting, further highlighting the advantages of using a base model for region-level feature extraction.

[0145] Data size: Peak performance was achieved with 15,000 region samples (see Figures 4(a) to 4(i)). Excessive size led to overfitting, and the results revealed a non-linear relationship between data size and diagnostic accuracy. Excessively reducing the training size impaired the model's ability to capture key morphological patterns, resulting in limited generalization ability.

[0146] Clinical application closed loop and system integration

[0147] After completing the construction and experimental verification of the core model for pathological image-assisted diagnosis of lung adenocarcinoma, this invention, through the clinical feedback optimization in step S4 and the clinical system interaction in step S5, combined with the modular design of a dedicated auxiliary diagnostic system, forms a complete medical data closed loop serving clinical diagnosis and treatment. The specific implementation is as follows:

[0148] In step S4, the clinical feedback and adaptive optimization process, this invention does not rely solely on the algorithm's autonomous iteration. Instead, it constructs a "pathology expert-model" clinical feedback mechanism: pathology experts review the slide-level prediction results output by the model through the electronic pathology system, providing clinical feedback on questionable subtyping (such as boundary cases between MIA and IA) or grading deviations (such as disputes over the determination of moderate and low differentiation). The system iteratively fine-tunes the classification threshold of the XGBoost decision tree based on this feedback. The threshold adjustment process is combined with the optimal hyperparameter configuration output by Bayesian optimization, and the model's decision boundary is calibrated using clinical standards such as the WHO classification standard for lung adenocarcinoma and the IASLC differentiation grading standard. The specific integration process is as follows: after the initial slide-level prediction results are reviewed by pathology experts, the decision threshold is updated based on the review opinions, and the XGBoost classifier is refitted. This iteration is repeated until the prediction results are completely consistent with clinical standards, ultimately achieving a high degree of consistency between the model's prediction results and the relevant pathology diagnosis report standards, avoiding a disconnect between algorithm optimization and clinical practice.

[0149] In step S5, the diagnostic report generation and clinical system interaction, this invention focuses on the clinical translation and system linkage of medical data: First, the subtype classification (AIS / MIA / IA), differentiation degree (high / intermediate / low differentiation), and corresponding clinical stage information output by the model are safely and automatically interacted with the hospital pathology information system (PACS) to retrieve relevant pathological data such as the patient's previous pathology slide reports and lesion measurement data; then, the above information is integrated, and a structured pathological diagnostic report containing subtype classification, differentiation degree, corresponding clinical stage, and treatment recommendations (such as regular follow-up for AIS and adjuvant chemotherapy after surgery for low-differentiation IA) is generated with reference to WHO lung adenocarcinoma diagnosis and treatment recommendations; finally, the structured report is transmitted to the hospital's electronic medical record system (EHR) through a standardized interface, fully covering the entire process from the input of HE-stained whole slide image data from the pathology department, intelligent model analysis, diagnostic report generation to clinical diagnosis and treatment decision support.

[0150] To support the implementation of the aforementioned clinical closed loop, the present invention provides a lung adenocarcinoma pathology image-assisted diagnostic system (…). Figure 8The system features multi-module division of labor and adaptability to medical systems: the image processing module is responsible for performing regional-level subtyping, slide-level feature extraction, and hierarchical decision-making, providing accurate diagnostic basis for subsequent clinical applications; the data interface module communicates specifically with the hospital's PACS and EHR systems to retrieve patient WSI data and clinical information and transmit structured reports, while ensuring interactive security through the data integration interface deployed by the hospital; the report generation module automatically fills predefined templates with multi-dimensional diagnostic elements such as dominant histological subtype, degree of differentiation, and percentage of each subtype output by the image processing module, generating structured pathology diagnostic reports that conform to the HL7 FHIR standard; the physician review module provides a human-computer interaction interface for pathologists to review and modify reports and complete electronic signatures, ensuring the clinical compliance of reports; the entire system is deeply integrated with the hospital's pathology information system through protocol ports, ultimately achieving closed-loop management of the entire process from pathological image input to clinical decision support.

[0151] Figure 9 The invention demonstrates the integrated architecture of the diagnostic system and data system in the system embodiment of the present invention. From the underlying distributed technical architecture level, it provides highly available data storage, computing processing and multi-terminal collaborative support for the image processing, data interface, report generation and other functional modules of the auxiliary diagnostic system of the present invention, ensuring the efficient flow of medical data between modules and the stable operation of clinical closed-loop business.

[0152] This architecture employs a highly available distributed design, forming a complete collaborative link from data storage to terminal applications: The data storage layer reliably stores medical data such as whole-slide images (WSI) through multiple object storage nodes, providing stable data access capabilities to the business processing layer via highly available addresses; The business processing layer deploys multiple business servers, which on the one hand receive pathology diagnosis requests from multiple terminals (accessed via highly available addresses) such as computers, apps, and web pages, and perform operations such as regional-level typing and slide-level decision-making based on core algorithms, outputting the results as pathology diagnosis results; on the other hand, fine-tuning / feedback of pathology results (such as diagnostic corrections after pathology expert review and threshold adjustments) can be fed back into the business servers, driving iterative optimization of models and processes. The data persistence layer ensures the consistency and high reliability of patient pathology data, model parameters, structured reports, and other information through a bidirectional synchronization mechanism between master / slave databases and master / slave servers, providing data persistence and state synchronization support for the business servers. The overall architecture relies on highly available addresses and distributed components to meet the concurrent access and data interaction needs of multiple terminals and multiple medical systems (such as PACS and EHR). Through a closed-loop design of storage-computation-feedback-persistence, it ensures the stable and efficient operation of the entire process of lung adenocarcinoma pathological diagnosis from data input to clinical decision support.

[0153] In summary, this invention proposes a lung adenocarcinoma pathological image-assisted diagnostic system and method based on an adaptive tree framework. The innovation of the FigATree framework in this system and method is reflected in:

[0154] 1. Regional level: The ViT-L encoder based on the pathological model overcomes the pre-training bias of natural images, captures subtype discrimination features through a self-attention mechanism, and the classification results match the WHO histological subtype criteria for lung adenocarcinoma, providing basic data support that conforms to guidelines for subsequent clinical subtyping.

[0155] 2. Slide-level: XGBoost decision tree integrates prior pathological knowledge (such as WHO lung adenocarcinoma diagnostic thresholds and IASLC differentiation grading indicators). The adaptive threshold mechanism avoids the limitations of traditional linear models. The decision-making process can be traced back to clinical diagnostic standards. The subtyping (AIS / MIA / IA) and grading (high / intermediate / low differentiation) results directly correspond to the basis for determining clinical treatment pathways.

[0156] 3. Transforming qualitative diagnosis into a quantitative paradigm (regional proportion statistics + hierarchical decision-making path). At the regional level, it provides detailed counts and proportions for each subtype, while slide-level subtype classification and grading produce an interpretable diagnostic process. In other words, FigATree minimizes subjectivity. While it cannot replace pathologists, it provides them with comprehensive and reliable diagnostic evidence, and the output quantitative data can be directly used to populate structured reports that conform to relevant pathology diagnostic reporting standards, supporting data interaction with hospital electronic medical record (EHR) systems.

[0157] The FigATree framework of this invention can also be used to develop organ-specific encoders, such as encoders built on large-scale lung cancer pathology datasets, to more effectively capture lung cancer-specific pathological features, introduce graph neural networks to model spatial relationships, and integrate large language models to generate pathology reports. This framework can be extended to pathological tasks such as squamous cell carcinoma and small cell carcinoma of the lung, and maintains compatibility with hospital pathology information systems (PACS) during the extension process, ensuring that newly developed functions can be seamlessly integrated into the clinical pathology workflow. This framework can also be extended to pathological tasks such as squamous cell carcinoma and small cell carcinoma of the lung, and after extension, it can still be bound to the corresponding clinical diagnostic criteria for the cancer type, maintaining a closed loop of medical data from data input to intelligent analysis to report generation to clinical application, complying with medical informatics standards.

[0158] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A method for assisted diagnosis of lung adenocarcinoma based on adaptive tree framework pathological images, characterized in that, Includes the following steps: S1. Regional Histological Subtype Classification: The ViT encoder, a visual transformer pre-trained based on a pathological model, is used to perform fine-grained classification of tissue regions in whole slide images and output regional prediction results of multiple histological subtypes that conform to the classification criteria for lung adenocarcinoma. S2. Slide-level feature extraction: Based on the regional prediction results, calculate the statistical proportions of key pathological features used for lung adenocarcinoma subtype classification and differentiation grade grading, including the proportion of lepidic regions, the proportion of high-grade regions, and the proportion of acinar regions; the statistical proportions of the key pathological features are derived from the spatial aggregation of the regional prediction results, and multi-dimensional diagnostic elements including the maximum diameter of the infiltrative lesions are extracted; the region size is set according to the preset resolution specifications and extracted at a predetermined magnification. S3. Slide-level hierarchical decision-making: The statistical proportions of the key pathological features are input into the XGBoost-based adaptive decision-making module, and the decision is executed sequentially through a two-layer decision tree structure: First decision tree: Based on the proportion of adherent regions and the World Health Organization's diagnostic threshold for lung adenocarcinoma, slides are classified into AIS, MIA, or IA, where AIS is adenocarcinoma in situ, MIA is microinvasive adenocarcinoma, and IA is invasive adenocarcinoma. The second decision tree: For slides identified as invasive adenocarcinoma, the degree of differentiation is graded based on the proportion of high-grade regions, the ratio of acinar to lepidic regions, and the IASLC differentiation grading criteria. A high-grade region proportion ≥20% is output as a poorly differentiated category, acinar predominance with high-grade regions <20% is output as a moderately differentiated category, and lepidic predominance with high-grade regions <20% is output as a highly differentiated category. The IASLC refers to the International Association for the Study of Lung Cancer. S4. Clinical Feedback and Adaptive Optimization: Pathology experts review the slide-level prediction results output by the model through the electronic pathology system and provide clinical feedback. Based on this clinical feedback, the classification threshold of the decision tree is iteratively fine-tuned to improve the consistency between the model's prediction results and clinical diagnostic standards. The threshold adjustment process combines the optimal hyperparameter configuration output by Bayesian optimization, calibrates the model's decision boundary through clinical standards, and iteratively optimizes the consistency between the model's prediction results and the pathology diagnostic report specifications. S5. Diagnostic Report Generation and Clinical System Interaction: The subtype classification, differentiation degree, and corresponding clinical stage information output by the model are interacted with the PACS to retrieve the patient's associated pathological data, generating a structured pathological diagnostic report containing subtype classification, differentiation degree, corresponding clinical stage, and treatment recommendations; the structured pathological diagnostic report is then transmitted to the EHR, completing a complete medical closed-loop operation from the input of HE-stained whole slide image data from the pathology department, intelligent analysis, diagnostic report generation to clinical diagnosis and treatment decision support; wherein, the PACS is the hospital pathology information system, and the EHR is the hospital electronic medical record system.

2. The method as described in claim 1, characterized in that, In step S1: The pathological basic model initializes the ViT encoder weights through pre-training on a large-scale pathological image corpus. By fine-tuning the regional predictor with an additional linear classifier head and cross-entropy loss function, feature vectors are generated to encode the local and global morphological features of the organization region.

3. The method according to any one of claims 1-2, characterized in that, In step S3: The two-layer decision tree structure is implemented using the XGBoost model, whose objective function includes a regularization term to prevent overfitting, specifically: Parameters that control the number of nodes in the decision tree can limit the tree depth; L2 regularization is used to constrain leaf node values ​​to enhance generalization ability.

4. The method as described in claim 3, characterized in that: The objective function of the XGBoost model is: in Let l be the objective function of the t-th iteration, l be the loss function, and n be the number of samples i. It is the true label of sample i. It is the predicted value for sample i in the previous round. It is the prediction output of the t-th decision tree for sample i. It is a regularization term. For constant terms; Where the loss function The specific form is: It is the predicted probability that sample i belongs to category j. The regularization term Calculate according to the following formula: in It is a parameter that controls the number of leaf nodes to prevent the tree structure from becoming too deep and causing overfitting. It is the penalty coefficient; It is an L2 regularization term that controls the size of the leaf node values ​​to enhance generalization ability. It is the output value of the j-th leaf node; The hyperparameters in the XGBoost model are adjusted using Bayesian optimization to construct the performance profile of the surrogate model's approximate hyperparameters and the objective function optimization effect. Furthermore, the model explores and utilizes the acquisition function's balancing mechanism to search for regularization terms. The parameter set that achieves an optimal balance with the loss function l is used to enhance the model's generalization ability.

5. The method according to any one of claims 1-2, characterized in that: The training of the ViT encoder, a visual transformer pre-trained based on a pathological model, employs a data augmentation strategy involving horizontal and vertical flipping. The training of the XGBoost-based adaptive decision module employs an adaptive gradient optimization algorithm, and the model is updated with optimized learning rate and batch size parameters.

6. The method according to any one of claims 1-2, characterized in that, In step S4: The integration process based on feedback from pathology experts includes: The initial slide-level prediction results were reviewed by pathology experts; Update the decision threshold based on the review results and refit the XGBoost classifier; Repeat the iterations until the predicted results are consistent with clinical standards.

7. The method according to any one of claims 1-2, characterized in that: The various histological subtypes include the following histological subtypes: lepidic, acinar, papillary, high-grade, invasive mucinous adenocarcinoma, and normal tissue; The higher-level type is further subdivided into solid type, micropapillary type, cribriform structure, and complex glandular structure.

8. A lung adenocarcinoma pathological image-assisted diagnostic system based on an adaptive tree framework, used to execute the lung adenocarcinoma pathological image-assisted diagnostic method based on an adaptive tree framework as described in any one of claims 1 to 7, characterized in that, include: The image processing module is used to perform regional-level histological subtyping, slide-level feature extraction, and hierarchical decision-making, enabling fine-grained classification and diagnosis of whole slide images; The data interface module is used to communicate with the hospital's PACS system and EHR system, retrieve patient WSI data and clinical information, and send back structured diagnostic reports; wherein, the WSI is a whole slide image; The report generation module is used to generate a preliminary structured pathological diagnosis report that conforms to the HL7 FHIR standard based on the output of the image processing module. The doctor review module provides a human-computer interaction interface for pathologists to review, modify, and electronically sign the structured pathology diagnosis reports. The system integrates with the hospital's pathology information system through a protocol port, enabling a closed-loop operation throughout the entire process, from pathological image input, intelligent analysis, report generation to clinical decision support.

9. The lung adenocarcinoma pathological image-assisted diagnostic system according to claim 8, characterized in that: The report generation module is configured to automatically fill the predefined structured report template with the multi-dimensional diagnostic elements output by the image processing module to generate the preliminary structured pathological diagnosis report; the multi-dimensional diagnostic elements include one or more of the following: dominant histological subtype classification, differentiation grade determination, percentage of each subtype, proportion of high-grade components, and maximum diameter of infiltrative lesions; The data interface module is also configured to: conduct secure and automated data interaction with the hospital's medical record information system through the data integration interface deployed by the hospital; The auxiliary diagnostic system realizes a complete closed-loop management of medical care, from pathological data input, intelligent analysis, diagnostic report generation to clinical diagnosis and treatment decision support.

Images