An AI-based multi-modal endometrial cancer diagnosis system

Abstract

This invention relates to the field of diagnostic system technology and discloses an AI-based multimodal diagnostic system for endometrial cancer. The system includes a multimodal diagnostic system for collaborative analysis and decision support of multi-source medical data in the medical diagnosis of endometrial cancer. The multimodal diagnostic system comprises: a multi-source heterogeneous data fusion unit for receiving and standardizing heterogeneous endometrial cancer-related data from medical imaging equipment, electronic medical records, and genomics testing equipment in real time; achieving deep collaborative analysis of cross-modal information by fusing multi-source medical data; employing incremental learning and federated learning techniques to ensure continuous system evolution and privacy security; and combining uncertainty quantification and interpretable output to provide accurate, safe, and traceable intelligent diagnostic decision support for endometrial cancer, effectively improving the scientific rigor of early screening, disease assessment, and treatment optimization.

Description

Technical Field

[0001] This invention relates to the field of diagnostic system technology, specifically to an AI-based multimodal diagnostic system for endometrial cancer. Background Technology

[0002] Endometrial cancer is a common gynecological malignancy. Clinical diagnosis often requires the integration of multiple medical data, including imaging examinations, laboratory tests, physiological monitoring and genomics information. However, these data usually come from different medical devices and have significant heterogeneity, with obvious differences in acquisition time, spatial resolution and data format. According to CN118629634A, a multimodal data fusion intelligent diagnostic system is disclosed. This technology discloses "a multimodal data fusion intelligent diagnostic system, relating to the field of artificial intelligence diagnostic system technology, including a data fusion module, a diagnostic algorithm module, a personalized diagnostic module, and a collaborative diagnostic module. Its features include: the data output of the data fusion module serves as the database source for the diagnostic algorithm module; the personalized parameter suggestion output of the personalized diagnostic module serves as the data input for the diagnostic algorithm module; and the collaborative diagnostic module is used to perform data sharing diagnosis on the user information collected by the personalized diagnostic module and output collaborative diagnostic opinions to the data fusion module and the diagnostic algorithm module." This system has the technical effect of "achieving artificial intelligence-based, personalized diagnosis and treatment suggestions based on individual data through the design of a personalized diagnostic module, solving the problem of how to adaptively adjust the diagnostic model according to individual differences, improving the accuracy and individual applicability of the diagnosis, and enhancing patients' trust and satisfaction with the diagnostic results." The core problem facing existing endometrial cancer medical diagnostic systems is the difficulty in effectively integrating multi-source heterogeneous medical data (such as imaging, physiological monitoring, and genomics) from different devices and detection technologies. These data vary significantly in terms of acquisition time, spatial resolution, and format, making it impossible to achieve accurate spatiotemporal alignment and standardized processing. At the same time, traditional systems lack dynamic scalability, making it impossible to adapt to new data modalities generated by new medical devices, and also difficult to achieve collaborative knowledge optimization across medical institutions while ensuring data privacy and security. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides an AI-based multimodal diagnostic system for endometrial cancer. By fusing multi-source medical data, it achieves deep collaborative analysis of cross-modal information; it employs incremental learning and federated learning techniques to ensure continuous system evolution and privacy security; and by combining uncertainty quantification and interpretable output, it provides accurate, safe, and traceable intelligent diagnostic decision support for endometrial cancer, effectively improving the scientific rigor of early screening, disease assessment, and treatment optimization.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an AI-based multimodal diagnostic system for endometrial cancer, comprising a multimodal diagnostic system and used for collaborative analysis and decision support of multi-source medical data in the medical diagnosis of endometrial cancer. The multimodal diagnostic system includes: A multi-source heterogeneous data fusion unit is used to receive and standardize heterogeneous data related to endometrial cancer from medical imaging equipment, electronic medical records and genomics testing equipment in real time; The dynamic weight allocation unit uses an attention mechanism to dynamically adjust the contribution weights of different modalities in the diagnosis of endometrial cancer. The cross-modal feature extraction unit contains parallel 3D-CNN and Transformer branches, which respectively handle spatiotemporal features and long-range dependencies; The uncertainty quantification module generates probability distribution outputs and calculates diagnostic confidence levels using Monte Carlo Dropout.

[0005] Preferably, the multi-source heterogeneous data fusion unit includes: The multimodal spatiotemporal alignment module uses a dynamic time warping algorithm to unify the time reference of data from various modalities and performs non-rigid registration on medical images such as endometrial ultrasound and pelvic CT / MRI to achieve spatial alignment. The intelligent missing data processing module constructs a dedicated completion model based on generative adversarial networks for different modal characteristics. It uses a context-aware PatchGAN generator for missing regions of medical images and an LSTM-attention hybrid interpolator for missing physiological signal periods. The standardization module performs adaptive normalization of data from each modality. Medical images are standardized using Z-score after window width and window level adjustment. Genomic data is converted into a unified code through a location-specific score matrix. The genomic data includes ER, PR, p53 immunohistochemical indicators and related gene detection results.

[0006] Preferably, the dynamic weight allocation unit includes: The modal quality assessment module calculates the original weights based on the signal-to-noise ratio and the proportion of missing data. The clinical scenario adaptation module automatically adjusts the modal weight coefficients according to different stages of endometrial cancer diagnosis and treatment, including screening, diagnosis and recurrence monitoring. The real-time feedback learning mechanism module continuously optimizes weight parameters by having doctors correct records, thereby strengthening the correlation weight between typical symptom data and lesion feature data.

[0007] Preferably, the cross-modal feature extraction unit employs: Cascaded cross-modal attention fusion layers achieve pixel-level alignment in the feature space; Learnable modal embedding vectors are used to identify different data sources; The feature distillation compression engine compresses cross-modal feature dimensions through a teacher-student network, preserving key diagnostic features such as lesion size, infiltration depth, and abnormal gene expression.

[0008] Preferably, the multimodal diagnostic system further includes: The incremental learning engine unit supports new modal data through dynamic architecture expansion without retraining the entire model; The federated learning interface unit enables each medical node to participate in model optimization under the protection of data privacy, and realizes collaborative training of multi-center endometrial cancer case data.

[0009] Preferably, the incremental learning engine unit includes: The modality-aware neural architecture search module automatically generates network branches that adapt to new modalities. The knowledge solidification module employs flexible weight solidification technology to prevent catastrophic forgetting; The cross-modal knowledge transfer module transfers the feature extraction capabilities of existing modalities to new modalities.

[0010] Preferably, the uncertainty quantification module unit includes: A Bayesian confidence calculation engine that calculates the confidence score of endometrial cancer diagnosis results based on a Bayesian neural network. A multimodal conflict arbitrator triggers an alert when the difference between conclusions from different modalities exceeds a threshold. The interpretable output module generates heatmap annotations for sources of uncertainty, highlighting blurred lesion areas or contradictory gene detection indicators in images.

[0011] Preferably, the multimodal diagnostic system further includes: The real-time decision tree engine unit maps deep learning outputs to traceable clinical decision-making paths, linking them to standard procedures for the diagnosis and treatment of endometrial cancer. The evidence-based medicine validation unit automatically links to the latest clinical guidelines and literature evidence for endometrial cancer, and marks the evidence-based basis for diagnostic conclusions; The risk prediction unit dynamically displays the spatiotemporal prediction results of endometrial cancer progression based on multimodal features, including the time trend of recurrence risk and metastasis probability.

[0012] This invention provides an AI-based multimodal diagnostic system for endometrial cancer. Compared with existing technologies, it has the following advantages: 1. By using multi-source heterogeneous data fusion technology, the challenges of spatiotemporal alignment, missing data completion, and standardization of medical multimodal data are solved; the dynamic weight allocation mechanism intelligently adjusts the contribution of each modality based on data quality and clinical value; the cross-modal feature extraction network fully explores the feature correlations of different data through parallel 3D-CNN and Transformer branches; it breaks through the limitations of traditional single-modal diagnosis, enabling multi-dimensional information such as imaging, genomics, and physiological parameters to work synergistically, significantly improving diagnostic accuracy and providing a more comprehensive and reliable basis for clinical decision-making.

[0013] 2. It adopts an incremental learning engine to achieve seamless access to new modal data, automatically expands processing capabilities through neural architecture search, and protects existing diagnostic knowledge with elastic weight solidification technology; the federated learning interface achieves collaborative optimization of multi-center medical knowledge under the premise of strictly protecting data privacy; it can continuously adapt to the development of medical technology, constantly absorb the latest diagnostic and treatment experience, form a dynamically evolving diagnostic capability, and provide medical institutions with a long-term available, safe and reliable intelligent diagnostic solution.

[0014] 3. The uncertainty quantification module provides probabilistic diagnostic conclusions, the multimodal conflict detection mechanism automatically identifies contradictory evidence, and the interpretable output intuitively displays the basis for judgment; the decision tree engine transforms deep learning results into treatment paths that conform to clinical thinking, the evidence-based medicine validation unit links to the latest guidelines in real time, and the risk prediction model provides visual analysis of disease progression; a transparent and reliable decision support system is constructed, which retains the analytical depth of AI while conforming to clinical work habits, effectively improving diagnostic efficiency and accuracy. Attached Figure Description

[0015] Figure 1 This is a system block diagram of the present invention; Figure 2 This is a block diagram of the multi-source heterogeneous data fusion unit in this invention; Figure 3 This is a block diagram of the dynamic weight allocation unit in this invention; Figure 4 This is a block diagram of the cross-modal feature extraction unit in this invention; Figure 5 This is a block diagram of the uncertainty quantification module unit in this invention; Figure 6 This is a block diagram of the incremental learning engine unit in this invention.

[0016] In the diagram: 1. Multimodal diagnostic system; 11. Multi-source heterogeneous data fusion unit; 111. Multimodal spatiotemporal alignment module; 112. Intelligent missing data processing module; 113. Standardization module; 12. Dynamic weight allocation unit; 121. Modal quality assessment module; 122. Clinical context adaptation module; 123. Real-time feedback learning mechanism module; 13. Cross-modal feature extraction unit; 131. Cascaded cross-modal attention fusion layer; 132. Learnable modal embedding vector; 133. 14. Feature distillation and compression engine; 14. Uncertainty quantification module; 141. Bayesian confidence calculation engine; 142. Multimodal conflict arbitrator; 143. Interpretability output module; 15. Incremental learning engine; 151. Modality-aware neural architecture search module; 152. Knowledge solidification module; 153. Cross-modal knowledge transfer module; 16. Federated learning interface unit; 17. Real-time decision tree engine unit; 18. Evidence-based medicine validation unit; 19. Risk prediction unit. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 - Figure 6 This invention provides a technical solution: an AI-based multimodal diagnostic system for endometrial cancer, comprising a multimodal diagnostic system 1 and used for collaborative analysis and decision support of multi-source medical data in the medical diagnosis of endometrial cancer. The multimodal diagnostic system 1 includes: The multi-source heterogeneous data fusion unit 11 is used to receive and standardize heterogeneous data related to endometrial cancer from medical imaging equipment, electronic medical records and genomics testing equipment in real time. Dynamic weight allocation unit 12 uses an attention mechanism to dynamically adjust the contribution weights of different modal data in the diagnosis of endometrial cancer. The cross-modal feature extraction unit 13 includes parallel 3D-CNN branches and Transformer branches, which respectively handle spatiotemporal features and long-range dependencies; Uncertainty quantification module unit 14 generates probability distribution output and calculates diagnostic confidence through Monte Carlo Dropout.

[0019] This implementation plan enables collaborative analysis of multimodal medical data, overcoming the information limitations of traditional single-modal diagnostic systems; it significantly improves the robustness of diagnostic results through intelligent weight allocation and uncertainty assessment; and its modular design allows the system to flexibly adapt to the diagnostic and treatment needs of different specialties, providing clinicians with comprehensive and reliable intelligent decision support, effectively reducing the risk of misdiagnosis in complex cases, and improving diagnostic and treatment efficiency.

[0020] Specifically, the multi-source heterogeneous data fusion unit 11 includes: The multimodal spatiotemporal alignment module 111 uses a dynamic time warping algorithm to unify the time reference of each modality of data and performs non-rigid registration on medical images such as endometrial ultrasound and pelvic CT / MRI to achieve spatial alignment. The intelligent missing data processing module 112 constructs a dedicated completion model based on generative adversarial network for different modal characteristics. It uses a context-aware PatchGAN generator for missing regions of medical images and an LSTM-attention hybrid interpolator for missing physiological signal periods. Standardization module 113 performs adaptive normalization of data from each modality. Medical images are standardized using Z-score after window width and window level adjustment. Genomic data are converted into a unified code through a position-specific score matrix. Genomic data includes ER, PR, p53 immunohistochemical indicators and related gene detection results.

[0021] In this embodiment, a multimodal spatiotemporal alignment module 111 is used to achieve accurate matching of cross-device data. This module uses a dynamic time warping algorithm to unify the time axis of physiological signals from wearable devices and electronic medical records, ensuring temporal correlation such as "time of onset of vaginal bleeding symptoms - simultaneous ultrasound examination". At the same time, a non-rigid registration technique based on deep learning is used to eliminate spatial differences between different imaging devices, ensuring that multimodal data achieves sub-millimeter alignment in anatomical structures and accurately tracks changes in endometrial lesions. The intelligent missing data processing module 112 addresses the common problem of partial missing data in medical data by developing a modality-specific intelligent completion solution: for missing areas in ultrasound / CT / MRI images, it uses a generative adversarial network that combines local anatomical semantics for context-aware repair; for physiological parameters that are interrupted in the collection of wearable devices, it uses a temporal attention mechanism to reconstruct the most likely signal curve shape. Standardization module 113 implements a medical-specific normalization strategy to establish a unified feature representation space for different modalities: medical imaging data retains key diagnostic density intervals through adaptive window width and window level adjustments; genome sequencing data is transformed into a standardized functional activity scoring matrix.

[0022] Specifically, the dynamic weight allocation unit 12 includes: Modal quality assessment module 121 calculates the original weights based on signal-to-noise ratio and the proportion of missing data; The clinical scenario adaptation module 122 automatically adjusts the modal weight coefficients according to different stages of endometrial cancer diagnosis and treatment, including screening, diagnosis and recurrence monitoring. The real-time feedback learning mechanism module 123 continuously optimizes the weight parameters through doctor correction records, strengthening the correlation weight between typical symptom data and lesion feature data.

[0023] In this embodiment, the modal quality assessment module 121 uses an adaptive signal processing algorithm to analyze the noise distribution characteristics and integrity of each modal data, and automatically downweights MRI sequences with motion artifacts or ECG data with signal interruptions. The clinical scenario adaptation module 122 has a built-in disease-modal association knowledge graph that automatically adjusts modal weight coefficients according to different stages of endometrial cancer diagnosis and treatment (screening / diagnosis / recurrence monitoring). For example, it increases the weight of ultrasound images in the screening stage and increases the weight of pathological genomic data in the diagnosis stage. The real-time feedback learning mechanism module 123 constructs a closed-loop optimization system. When clinicians correct the system's suggestions, it automatically traces the key modalities that caused the differences and adjusts the corresponding weight parameters through online learning algorithms.

[0024] Specifically, the cross-modal feature extraction unit 13 employs: A cascaded cross-modal attention fusion layer 131 is used to achieve pixel-level alignment in the feature space; Learnable modal embedding vectors 132 are used to identify different data sources; The Feature Distillation Compression Engine 133 compresses cross-modal feature dimensions through a teacher-student network, preserving key diagnostic features such as lesion size, infiltration depth, and abnormal gene expression.

[0025] In this embodiment, precise feature alignment is achieved through a cascaded cross-modal attention fusion layer 131. This fusion layer employs a multi-scale attention mechanism to establish cross-modal feature associations at different depth levels of the convolutional neural network, which can automatically correct anatomical structure shifts caused by different imaging devices and achieve pixel-level feature matching. The learnable modality embedding vectors 132 assign unique feature identifiers to each data modality. These vectors are dynamically optimized during training, preserving the unique characteristics of each modality while constructing a unified feature representation space. The feature distillation and compression engine 133 extracts complex cross-modal interaction features into compact discriminative representations by constructing a teacher-student network architecture. The teacher network learns comprehensive cross-modal associations, while the student network focuses on retaining the feature combinations that are most discriminative for clinical diagnosis.

[0026] Specifically, the multimodal diagnostic system 1 also includes: Incremental learning engine unit 15 supports new modal data through dynamic architecture expansion without retraining the entire model; Federated learning interface unit 16 enables each medical node to participate in model optimization under the protection of data privacy, and realizes collaborative training of multi-center endometrial cancer case data.

[0027] In this embodiment, the incremental learning engine unit 15 enables the system to continuously evolve. This engine adopts dynamic neural network architecture technology. When new medical imaging equipment or testing instruments are connected, it can automatically analyze the feature distribution of the new data modalities, intelligently expand dedicated processing branches, and maintain the integrity of the original network structure, ensuring that the system does not damage the established diagnostic capabilities while absorbing new knowledge. The federated learning interface unit 16 constructs a secure collaborative training framework. Each medical institution node can participate in model optimization through encrypted gradient exchange without leaving the private network. The system uses differential privacy protection technology to process noise in the uploaded parameters and designs a modality isolation update mechanism to ensure that specific specialty data is only used to optimize the processing capabilities of the corresponding modality.

[0028] Specifically, the incremental learning engine unit 15 includes: Modality-aware neural architecture search module 151 automatically generates network branches that adapt to new modalities; The knowledge solidification module 152 employs flexible weight solidification technology to prevent catastrophic forgetting. The cross-modal knowledge transfer module 153 transfers the feature extraction capabilities of existing modalities to new modalities.

[0029] In this embodiment, the modality-aware neural architecture search module 151 automatically analyzes the data characteristics of the new modality and intelligently generates a matching network topology. This module can identify the spatiotemporal feature patterns of the new modality and automatically configure key parameters such as the optimal convolutional kernel size and the number of attention heads, achieving plug-and-play modality expansion capabilities. The knowledge solidification module 152 uses elastic weight solidification technology to dynamically lock the key network parameters of the existing modality. By calculating the importance of the parameters in the Fisher information matrix of historical tasks, it assigns differentiated regularization constraint strengths to different connections to ensure that the learned diagnostic knowledge is not destroyed when training the new modality. The cross-modality knowledge transfer module 153 constructs feature mapping relationships between modalities and transfers the feature extraction capabilities obtained from the training of the existing modality to the new modality branch through parameter sharing, feature projection, and other methods, significantly reducing the requirement for the amount of labeled data for the training of the new modality.

[0030] Specifically, the uncertainty quantification module unit 14 includes: Bayesian confidence calculation engine 141 calculates the confidence of endometrial cancer diagnosis results based on Bayesian neural network; The multimodal conflict arbitrator 142 triggers an early warning when the difference between conclusions from different modalities exceeds a threshold. The interpretability output module 143 generates heat map annotations of sources of uncertainty, highlighting blurred lesion areas or contradictory gene detection indicators in images.

[0031] In this embodiment, the diagnostic results are probabilistically evaluated by a Bayesian confidence calculation engine 141. This engine uses Monte Carlo sampling technology to simulate the distribution of network parameters and captures the output variability of the model for the same input data in multiple forward propagations, generating a diagnostic probability report containing confidence intervals. A multimodal conflict arbitrator 142 monitors the intermediate features and output results of each modality subsystem in real time. When there is a clinically significant contradiction between imaging features and cross-modal evidence such as genomic markers, a graded early warning mechanism is automatically triggered, and the specific data fragments of the conflicting modality are highlighted on the doctor's workstation. An interpretability output module 143 generates an uncertainty heatmap by integrating gradient algorithms and class activation mapping technology on the original medical images. It visually marks the key anatomical areas that lead to doubts about the diagnosis using visual coding, and generates text descriptions pointing out technical factors (such as image artifacts) or clinical factors (such as rare case features) that may affect the judgment.

[0032] Specifically, the multimodal diagnostic system 1 also includes: The real-time decision tree engine unit 17 maps the deep learning output to a traceable clinical decision path, which is associated with the standard steps of endometrial cancer diagnosis and treatment. Evidence-based medicine verification unit 18 automatically links to the latest clinical guidelines and literature evidence for endometrial cancer, and marks the evidence-based basis for diagnostic conclusions; Risk prediction unit 19 dynamically displays the spatiotemporal prediction results of endometrial cancer progression based on multimodal features, including the time trend of recurrence risk and metastasis probability.

[0033] In this embodiment, the multimodal diagnostic system maps the complex features output by the deep learning model into decision paths that conform to clinical diagnosis and treatment logic through the real-time decision tree engine unit 17. The engine has a built-in interpretable conversion layer, which can convert the black box output of the neural network into a tree structure containing key diagnostic evidence, differential diagnosis points and treatment suggestions. Each decision node is labeled with the corresponding medical feature evidence, which supports clinicians to backtrack the diagnostic reasoning process. The Evidence-Based Medicine Validation Unit 18 continuously connects to authoritative medical knowledge bases such as UpToDate and PubMed. When the system outputs diagnostic suggestions, it automatically retrieves the latest matching clinical research evidence, guideline recommendation levels, and relevant case reports, and displays the evidence level of supporting literature and reminders of contradictory evidence on the decision-making interface. Risk prediction unit 19 integrates patients' time-series examination data and environmental factors, generates a visual prediction curve of disease progression through spatiotemporal modeling, and marks the prediction time window of key pathological changes.

[0034] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0035] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based multimodal diagnostic system for endometrial cancer, characterized in that: Including a multimodal diagnostic system (1) and used for collaborative analysis and decision support of multi-source medical data in the medical diagnosis of endometrial cancer, the multimodal diagnostic system (1) includes: A multi-source heterogeneous data fusion unit (11) is used to receive and standardize heterogeneous data related to endometrial cancer from medical imaging equipment, electronic medical records and genomics testing equipment in real time; The dynamic weight allocation unit (12) uses an attention mechanism to dynamically adjust the contribution weights of different modal data in the diagnosis of endometrial cancer. The cross-modal feature extraction unit (13) includes parallel 3D-CNN branches and Transformer branches, which respectively handle spatiotemporal features and long-range dependencies; The uncertainty quantification module (14) generates a probability distribution output and calculates the diagnostic confidence level through Monte Carlo Dropout.

2. The AI-based multimodal diagnostic system for endometrial cancer according to claim 1, characterized in that: The multi-source heterogeneous data fusion unit (11) includes: The multimodal spatiotemporal alignment module (111) uses a dynamic time warping algorithm to unify the time reference of each modality of data and performs non-rigid registration on medical images such as endometrial ultrasound and pelvic CT / MRI to achieve spatial alignment. The intelligent missing data processing module (112) constructs a dedicated completion model for different modal characteristics based on generative adversarial network, uses a context-aware PatchGAN generator for missing areas of medical images, and uses an LSTM-attention hybrid interpolator for missing physiological signal periods. The standardization module (113) performs adaptive normalization of each modality of data. Medical images are standardized using Z-score after window width and window level adjustment. Genomic data are converted into a unified code through a position-specific score matrix. The genomic data includes ER, PR, p53 immunohistochemical indicators and related gene detection results.

3. The AI-based multimodal diagnostic system for endometrial cancer according to claim 1, characterized in that: The dynamic weight allocation unit (12) includes: Modal quality assessment module (121) calculates the original weights based on signal-to-noise ratio and the proportion of missing data; The clinical scenario adaptation module (122) automatically adjusts the modal weight coefficients according to different stages of endometrial cancer diagnosis and treatment, including screening, diagnosis and recurrence monitoring. The real-time feedback learning mechanism module (123) continuously optimizes the weight parameters by correcting the doctor's records, thereby strengthening the correlation weight between typical symptom data and lesion feature data.

4. The AI-based multimodal diagnostic system for endometrial cancer according to claim 1, characterized in that: The cross-modal feature extraction unit (13) adopts: A cascaded cross-modal attention fusion layer (131) is used to achieve pixel-level alignment in the feature space; Learnable modal embedding vectors (132) are used to identify different data sources; The feature distillation compression engine (133) compresses cross-modal feature dimensions through a teacher-student network, preserving key diagnostic features such as lesion size, infiltration depth, and abnormal gene expression.

5. The AI-based multimodal diagnostic system for endometrial cancer according to claim 1, characterized in that: The multimodal diagnostic system (1) also includes: The incremental learning engine unit (15) supports new modal data through dynamic architecture expansion without retraining the full model; The federated learning interface unit (16) enables each medical node to participate in model optimization under the protection of data privacy, and realizes collaborative training of multi-center endometrial cancer case data.

6. The AI-based multimodal diagnostic system for endometrial cancer according to claim 5, characterized in that: The incremental learning engine unit (15) includes: Modality-aware neural architecture search module (151) automatically generates network branches that adapt to new modalities; The knowledge solidification module (152) uses flexible weight solidification technology to prevent catastrophic forgetting; The cross-modal knowledge transfer module (153) transfers the feature extraction capabilities of existing modalities to new modalities.

7. The AI-based multimodal diagnostic system for endometrial cancer according to claim 1, characterized in that: The uncertainty quantification module unit (14) includes: A Bayesian confidence calculation engine (141) calculates the confidence of endometrial cancer diagnosis results based on a Bayesian neural network; A multimodal conflict arbitrator (142) triggers an alert when the difference between conclusions from different modalities exceeds a threshold. The interpretability output module (143) generates thermal icon annotations for sources of uncertainty, highlighting blurred lesion areas or contradictory gene detection indicators in images.

8. The AI-based multimodal diagnostic system for endometrial cancer according to claim 1, characterized in that: The multimodal diagnostic system (1) also includes: The real-time decision tree engine unit (17) maps the deep learning output to a traceable clinical decision path and associates it with the standard steps of endometrial cancer diagnosis and treatment. The evidence-based medicine verification unit (18) automatically links to the latest clinical guidelines and literature evidence for endometrial cancer and marks the evidence-based basis for diagnostic conclusions; The risk prediction unit (19) dynamically displays the spatiotemporal prediction results of endometrial cancer progression based on multimodal features, including the time trend of recurrence risk and metastasis probability.

Images