Oral cavity potential malignant disease-oriented whole-process auxiliary decision method and system

Abstract

This invention discloses a comprehensive decision support method and system for potential oral malignant diseases, comprising: constructing a pre-diagnosis risk prediction agent that integrates patient behavioral risk factors, genetic susceptibility markers, and multispectral initial screening images; generating a malignancy probability trajectory and triggering tiered screening through a time-series prediction model; establishing an in-diagnosis precision assessment agent that performs differential homeomorphic registration of autofluorescence and narrowband imaging images to generate fused data, and constructing an interpretable reasoning network to locate critical areas of malignancy in conjunction with pathological grading; constructing a post-diagnosis dynamic monitoring agent that establishes a digital twin model based on lesion baseline data, extracts mucosal texture and vascular morphological evolution features from each follow-up visit to generate malignancy trend curves and individualized intervention strategies; and constructing a multi-agent co-evolutionary framework that uses desensitized full-cycle data to perform federated closed-loop iterative optimization of the parameters of each agent. This aims to improve the early identification accuracy of precancerous lesion malignancy risk and optimize clinical intervention pathways.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to a whole-process auxiliary decision-making method and system for potential oral malignant diseases. Background Technology

[0002] Oral potential malignant diseases are a group of clinical conditions with malignant potential that lie between normal oral mucosa and oral squamous cell carcinoma. They include various subtypes such as oral leukoplakia, erythroplakia, oral lichen planus, and oral submucosal fibrosis. Accurate identification and timely intervention are key to preventing the occurrence of oral cancer.

[0003] Currently, the management of potential oral malignancies in clinical practice mainly relies on the physician's experience in visual and palpation examinations and morphological interpretation under single-lens white light endoscopy, combined with necessary biopsy and pathological examinations. However, there are significant subjective differences among physicians at different levels in assessing the risk of malignancy for the same lesion, resulting in a lack of unified quantitative basis for biopsy decisions and the determination of follow-up intervals. Existing computer-aided diagnostic technologies mostly focus on the binary classification of benign and malignant at a single moment, lacking continuous decision support capabilities throughout the entire cycle of pre-diagnosis screening, accurate assessment during diagnosis, and dynamic monitoring after diagnosis. In addition, in clinical practice, multimodal imaging data such as autofluorescence and narrow-band imaging are often analyzed separately from pathological grading information, and a collaborative reasoning mechanism integrating morphological features, functional metabolic information, and histological evidence has not yet been formed. For patients who have been diagnosed but have not yet undergone surgery, current follow-up strategies are mostly empirical fixed-period re-examinations, failing to provide individualized predictions of malignant trends and early warnings of intervention timing based on the temporal evolution of lesion texture and dynamic changes in vascular morphology. Therefore, there is an urgent need for an auxiliary decision-making method that can integrate multimodal time-series data and cover the entire process of managing potential oral malignant diseases, in order to improve the early identification accuracy of the risk of malignant transformation of precancerous lesions and optimize clinical intervention pathways. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides a whole-process auxiliary decision-making method and system for potential oral malignant diseases. Its important purpose is to improve the early identification accuracy of the risk of malignant transformation of precancerous lesions and optimize the clinical intervention path.

[0005] To achieve the above objectives, the first aspect of this invention provides a full-process auxiliary decision-making method for potential oral malignant diseases, comprising: Construct a pre-diagnosis risk prediction intelligent agent, collect patient behavioral risk factor scores, expression abundance of genetic susceptibility biomarkers and multispectral oral screening images, calculate the long-term cumulative effect of risk factors and perform malignant transformation probability trajectory prediction and graded screening of abnormal hyperplasia, and generate graded screening information; Establish a precise assessment agent during diagnosis. Based on the results of graded screening, obtain autofluorescence images and narrowband imaging images. Use the differential homeomorphic registration algorithm to perform pixel-level alignment and generate a fused data cube. Combine the abnormal proliferation graded labels of pathological sections to construct an interpretable reasoning network to locate the critical area of ​​malignancy and generate a risk stratification report. A post-diagnosis dynamic monitoring intelligent agent is constructed. The baseline status data of the lesion is used as the initial input to build a digital twin model of the patient. Images from each follow-up examination are obtained and the differences in mucosal texture evolution and the drift of vascular morphology index are extracted. A malignancy trend prediction curve is constructed and a personalized prevention and intervention strategy is generated. A multi-agent co-evolutionary framework was constructed to obtain actual post-diagnosis malignancy outcome data and post-operative panoramic pathological results. The parameters of each agent were optimized by combining the desensitized full-cycle case data in the central shared memory bank, and the multi-agent closed-loop iterative optimization was carried out using a federated learning mechanism.

[0006] In this solution, the construction of a pre-diagnosis risk prediction intelligent agent involves collecting patient behavioral risk factor scores, the expression abundance of genetic susceptibility biomarkers, and multispectral oral screening images. It calculates the long-term cumulative effect of risk factors and performs prediction of the malignant transformation probability trajectory of abnormal hyperplasia and graded screening, generating graded screening information. Specifically, this includes: Create an independent operating container for the pre-diagnosis risk prediction intelligent agent, establish a standardized data communication interface with the electronic medical record system, laboratory information system and multispectral oral screening equipment, and deploy data preprocessing pipeline components, time series risk prediction model and hierarchical decision rule engine within the operating environment; The standardized data communication interface is used to collect patient behavioral risk factor scores, genetic susceptibility biomarker expression abundance values, and multispectral oral screening images. These data are then input into the data preprocessing pipeline component to perform missing value imputation and maximum / minimum value normalization on the structured data, and to perform dark channel prior illumination non-uniformity correction and adaptive histogram equalization enhancement on the image data, thereby generating standardized input data. The normalized behavioral risk factor vector and the abundance vector of genetic susceptibility biomarkers are concatenated in the feature dimension to generate an initial state vector. Based on a sequence encoder consisting of multiple gated loop units stacked sequentially, the retention and discard ratio of historical accumulated information is adaptively controlled through update gate and reset gate mechanisms. After iterative deduction, a high-dimensional hidden state feature vector representing the long-term cumulative effect of risk factors is output. Multispectral oral screening images are input into a branched lightweight convolutional feature extractor. The first branch extracts the granular texture and plaque boundary morphology features of the mucosal surface, the second branch captures the distribution features of fluorescence signal attenuation regions, and the third branch extracts the spatial thermal distribution features of capillary density and oxygen saturation. The three feature maps are then compressed into a multispectral image feature summary vector after channel stitching and global average pooling. The high-dimensional hidden state feature vector and the multispectral image feature summary vector are summed element-wise in the feature fusion layer. The weighting coefficients are adaptively determined by trainable gating attention parameters to generate a joint feature vector that fuses image features and clinical features. The joint feature vector is fed into a time-series risk prediction model consisting of alternating stacks of fully connected layers and nonlinear activation functions. The output layer of the time-series risk prediction model contains twenty-four parallel neurons that correspond to the probability of malignant events occurring in each of the next twenty-four months, generating a continuous malignant probability trajectory prediction curve. Based on the malignancy probability trajectory prediction curve, the hierarchical decision rule engine is imported to calculate the integral value of the area under the curve within a preset time window, the average slope of the three to six months interval, and the detection results of the mutation inflection point. The risk level of the examinee is comprehensively determined. For high-risk levels, a priority biopsy channel instruction is triggered; for medium-risk levels, an image monitoring prompt is generated; and for low-risk levels, health education content and follow-up plans are pushed to generate hierarchical screening information.

[0007] In this solution, the establishment of a precise assessment agent during diagnosis involves acquiring autofluorescence images and narrowband imaging images based on tiered screening results. A differential homeomorphic registration algorithm is used for pixel-level alignment to generate a fused data cube. Combined with pathological slide abnormal proliferation grading labels, an interpretable reasoning network is constructed to locate critical malignant regions and generate a risk stratification report. Specifically, this includes: Create an independent operating container for the intelligent agent for precise assessment during diagnosis, establish a standardized data communication interface with autofluorescence imaging equipment, narrowband imaging equipment and pathology information system, and deploy a differential homeomorphic deformation registration engine, an interpretable reasoning network model and a risk stratification report generation component within the operating environment to complete the construction of the intelligent agent; The system receives tiered screening information output by the pre-diagnosis risk prediction intelligent agent. When the system indicates that it is entering the image evaluation process, it sends a collection trigger signal to the autofluorescence imaging device and the narrowband imaging device through the standardized data communication interface to acquire autofluorescence images and narrowband imaging images and complete the initial synchronization of timestamps and spatial coordinate systems. Using narrowband imaging as a fixed reference image and autofluorescence image as a floating image, a velocity vector field is constructed through a differential homeomorphic deformation registration engine and the normalized mutual information measure is iteratively optimized. Under the condition of satisfying the topology preservation constraint, the autofluorescence image is mapped pixel by pixel to the coordinate space of the narrowband imaging image to generate a spatially aligned dual-channel fusion data cube. Synchronously, the pathological slice abnormal proliferation grade label corresponding to the current lesion is retrieved from the pathological information system through the standardized data communication interface. It is then mapped into a high-dimensional semantic vector through the embedding matrix and injected as a conditional modulation signal into the conditional batch normalization layer of the interpretable inference network. The channel mean and variance of the convolutional feature map are modulated by affine transformation. The dual-channel fused data cube is input into the interpretable reasoning network model. High-level semantic feature maps are extracted through a multi-layer convolutional cascaded feature encoding backbone network and conditional batch normalization modulation. After hierarchical attention aggregation operation, channel attention weights and spatial attention heatmaps are generated. The data is then diverted to the malignant risk classification head to output the malignant risk probability value, and an attribution heatmap is calculated to locate the critical region of malignant transformation. Risk levels are divided based on the numerical range in which the malignancy risk probability value falls. The attribution strength statistics within the critical area of ​​malignancy are extracted from the attribution heatmap and the confidence score is calculated. Targeted biopsy sites are planned for cases with confidence scores below the lower limit of the criticality. For cases with confidence scores that meet the standard, the malignancy risk probability, risk level, attribution information and pathological grade consistency description are assembled, and finally, a risk stratification report with attached evidence chain is generated.

[0008] In this solution, the construction of a post-diagnosis dynamic monitoring intelligent agent uses baseline lesion status data as initial input to build a digital twin model of the patient, acquires images from previous follow-up examinations, extracts the difference in mucosal texture evolution and the drift of vascular morphology index, constructs a malignancy trend prediction curve, and generates personalized prevention and intervention strategies, specifically including: Create an independent operating container for the post-diagnosis dynamic monitoring agent, establish a standardized data retrieval interface with the image archiving and communication system, and deploy a patient digital twin model storage unit, a graph convolution-based registration and comparison module, a Gaussian process regression generator, and a treatment strategy generator within the operating environment; After deployment, the post-diagnosis dynamic monitoring agent sends a data request command to the in-diagnosis precision assessment agent to obtain the risk stratification report of the target patient and extract the baseline status data of the lesion and write it into the digital twin model storage unit to initialize the patient's digital twin model. When the target patient returns to the hospital for a follow-up examination, the current follow-up standardized endoscopic image sequence is obtained through the standardized data retrieval interface. Dark channel prior illumination non-uniformity correction, adaptive histogram equalization, and automatic cropping of the region of interest based on the baseline lesion contour are performed in sequence to obtain the current follow-up standardized image. Using a patient digital twin model to obtain baseline images, and inputting them with current follow-up standardized images, a graph convolution-based registration and comparison module is used to construct a spatial adjacency graph structure and calculate cross-graph attention weights. The output is a dense two-dimensional deformation field and a pixel-by-pixel texture residual map. The mean and standard deviation of the texture residuals are calculated within the contour of the critical region of malignancy as the difference in mucosal texture evolution. Vascular skeleton maps were extracted from baseline images and current follow-up standardized images, and vascular morphology analysis indices including mean branch angle, coefficient of variation of vessel diameter, and distribution density entropy were calculated. The vascular morphology analysis indices at the current time were subtracted from the indices at the baseline time and then weighted and summed to generate the vascular morphology index drift. The abnormal proliferation grading label encoding vector is obtained by using the patient's digital twin model. It is then input into the Gaussian process regression generator along with the difference in mucosal texture evolution and the vascular morphology index drift. All observation points accumulated before the current moment are used as the training set. The hyperparameters of the composite kernel function are updated by maximizing the marginal log-likelihood function to generate a malignant transformation trend prediction curve.

[0009] In this solution, the generation of personalized prevention and intervention strategies specifically includes: A treatment strategy generator is built based on a differentiable decision tree and embedded in the post-diagnosis dynamic monitoring agent. The decision tree depth is set to five layers. Each internal node outputs membership values ​​in multiple branch directions after linearly combining the input feature vector through the Sigmoid function. The leaf nodes store the treatment plan parameter vector including treatment method code, treatment intensity, treatment start time and follow-up interval. The slope value of the malignancy trend prediction curve in the near time window, the predicted risk probability value at the current moment, the difference in mucosal texture evolution, the drift of the vascular morphology index, and the contribution ratio of each sub-indicator are concatenated into an input feature vector, which is then input into the treatment strategy generator for forward propagation calculation to determine the branch membership degree of each internal node layer by layer. Based on the branch membership degree of each internal node, the arrival probability of each leaf node relative to the current input sample is calculated as a soft weight. The treatment plan parameter vectors stored in all leaf nodes are convexly combined according to the soft weight to generate a set of individualized continuous treatment plan parameters for the target patient. The training of the treatment strategy generator adopts joint optimization of multi-objective loss functions. The first objective is to minimize the probability of malignant transformation within the prediction time window, the second objective is to minimize the preset tissue damage index corresponding to the treatment method, and the third objective is to maximize the treatment compliance score based on the prediction of patient behavior logs. The gradient descent algorithm is used to simultaneously optimize the feature linear combination parameters of the internal nodes of the differentiable decision tree and the treatment plan parameter vector stored in the leaf nodes, so that the branching logic and the plan recommendation content converge in a coordinated manner towards maximizing malignant transformation inhibition, minimizing tissue damage, and promoting compliance optimization. The treatment mode coding distribution vector in the parameter set of individualized continuous treatment plan is analyzed to identify the main recommended plan and alternative plans. A structured treatment path diagram with the time axis as the main line is constructed, and a radar comparison chart of the marginal benefits and marginal risks of the main recommended plan and alternative plans is generated. Finally, the personalized prevention and intervention strategy is obtained by integrating the data.

[0010] In this scheme, the construction of a multi-agent co-evolutionary framework involves acquiring actual post-diagnosis malignancy outcome data and post-operative comprehensive pathological results, combining this data with desensitized full-cycle case data in a central shared memory bank to optimize the parameters of each agent, and utilizing a federated learning mechanism for multi-agent closed-loop iterative optimization. Specifically, this includes: Create a co-evolutionary management container, establish a two-way encrypted communication channel between the deployment nodes of the pre-diagnosis risk prediction agent, the in-diagnosis accurate assessment agent, and the post-diagnosis dynamic monitoring agent, deploy a central shared memory bank storage array within the container and divide it into a time-series prediction trajectory storage area, a pathological standard verification storage area, and a follow-up outcome annotation storage area, and configure a data desensitization engine for each partition; When any intelligent agent deployment node completes the full follow-up of the patient and generates malignant outcome annotation data or postoperative panoramic pathology report, a data return request is triggered. After integrity verification and desensitization processing, the data is uploaded to the corresponding logical partition of the central shared memory for persistent storage, and aggregated to form a standardized full-cycle case dataset with globally unified feature mapping. When the accumulated incremental data in the central shared memory reaches a preset threshold, the time-series calibration error is calculated by extracting paired samples of the pre-diagnosis predicted trajectory and follow-up outcome. The parameters of the gated cyclic unit sequence encoder and the weights of the multispectral image encoder convolutional layer in the pre-diagnosis risk prediction agent are updated in reverse. Versioned parameter patches are generated and pushed to each deployment node. Extract case samples from the diagnostic stage that delineate the critical area of ​​malignant transformation and obtain panoramic pathological verification. Compare the Dess similarity coefficient and Hausdorff distance between the attribution heat map area and the pathologically labeled area. Calculate the spatial positioning accuracy and probability calibration error. Update the hierarchical attention weights and conditional batch normalized affine transformation coefficients in the interpretable inference network in reverse. Extract population texture evolution and vascular morphology drift time-series trajectory according to the abnormal proliferation grade label, maximize the marginal log-likelihood function of each level to re-evaluate the radial basis function kernel length scale parameter and period kernel period length parameter of the Gaussian process regression generator, generate level-specific hyperparameters and push them to the post-diagnosis dynamic monitoring agent. Each deployment node calculates the incremental data gradient locally, encrypts it after differential privacy Gaussian noise perturbation, and uploads it to the federated aggregation server agent. It performs federated average aggregation based on sample size weighting to generate global parameter increments, updates the global model parameter snapshot in the central shared memory, and distributes it back to each node for replacement, thereby realizing multi-agent closed-loop iterative optimization.

[0011] A second aspect of the present invention provides a full-process auxiliary decision-making system for potential oral malignant diseases. The system includes: a memory, a processor, and a communication interface. The memory contains a full-process auxiliary decision-making method program for potential oral malignant diseases. When the processor executes the full-process auxiliary decision-making method program for potential oral malignant diseases, it implements the steps of the full-process auxiliary decision-making method for potential oral malignant diseases as described in any of the above claims. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0013] Figure 1 This is a flowchart of the first method of a full-process auxiliary decision-making method for potential oral malignant diseases provided in an embodiment of the present invention; Figure 2 This is a flowchart of a second method for a full-process auxiliary decision-making method for potential oral malignant diseases, provided by an embodiment of the present invention. Figure 3 This is a block diagram of a full-process auxiliary decision-making system for potential oral malignant diseases provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0015] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0016] Figure 1 This is a flowchart of the first method of a full-process auxiliary decision-making method for potential oral malignant diseases provided in an embodiment of the present invention; like Figure 1 As shown, the present invention provides a first method flowchart for a full-process assisted decision-making method for potential oral malignant diseases, comprising: S102, construct a pre-diagnosis risk prediction intelligent agent, collect patient behavioral risk factor scores, expression abundance of genetic susceptibility biomarkers and multispectral oral screening images, calculate the long-term cumulative effect of risk factors and perform malignant transformation probability trajectory prediction and graded screening of abnormal hyperplasia, and generate graded screening information. S104. Establish a precise assessment agent during diagnosis. Based on the results of graded screening, obtain autofluorescence images and narrowband imaging images. Use the differential homeomorphic registration algorithm to perform pixel-level alignment and generate a fused data cube. Combine the abnormal proliferation graded labels of pathological sections to construct an interpretable reasoning network to locate the critical area of ​​malignancy and generate a risk stratification report. S106, construct a post-diagnosis dynamic monitoring intelligent agent, use the baseline status data of the lesion as the initial input to construct a digital twin model of the patient, obtain images of each follow-up examination and extract the difference in mucosal texture evolution and the drift of vascular morphology index, construct a malignant transformation trend prediction curve and generate personalized prevention and intervention strategies. S108 constructs a multi-agent collaborative evolution framework, obtains actual post-diagnosis malignancy outcome data and post-operative panoramic pathology results, combines desensitized full-cycle case data in a central shared memory bank to optimize the parameters of each agent, and uses a federated learning mechanism to perform multi-agent closed-loop iterative optimization.

[0017] Furthermore, in a preferred embodiment of the present invention, the construction of the pre-diagnosis risk prediction intelligent agent, which involves collecting patient behavioral risk factor scores, expression abundance of genetic susceptibility biomarkers, and multispectral oral screening images, calculating the long-term cumulative effect of risk factors, predicting the malignant transformation probability trajectory of abnormal hyperplasia, and performing graded screening to generate graded screening information, specifically includes: Create an independent operating container for the pre-diagnosis risk prediction intelligent agent, establish a standardized data communication interface with the electronic medical record system, laboratory information system and multispectral oral screening equipment, and deploy data preprocessing pipeline components, time series risk prediction model and hierarchical decision rule engine within the operating environment; The standardized data communication interface is used to collect patient behavioral risk factor scores, genetic susceptibility biomarker expression abundance values, and multispectral oral screening images. These data are then input into the data preprocessing pipeline component to perform missing value imputation and maximum / minimum value normalization on the structured data, and to perform dark channel prior illumination non-uniformity correction and adaptive histogram equalization enhancement on the image data, thereby generating standardized input data. The normalized behavioral risk factor vector and the abundance vector of genetic susceptibility biomarkers are concatenated in the feature dimension to generate an initial state vector. Based on a sequence encoder consisting of multiple gated loop units stacked sequentially, the retention and discard ratio of historical accumulated information is adaptively controlled through update gate and reset gate mechanisms. After iterative deduction, a high-dimensional hidden state feature vector representing the long-term cumulative effect of risk factors is output. Multispectral oral screening images are input into a branched lightweight convolutional feature extractor. The first branch extracts the granular texture and plaque boundary morphology features of the mucosal surface, the second branch captures the distribution features of fluorescence signal attenuation regions, and the third branch extracts the spatial thermal distribution features of capillary density and oxygen saturation. The three feature maps are then compressed into a multispectral image feature summary vector after channel stitching and global average pooling. The high-dimensional hidden state feature vector and the multispectral image feature summary vector are summed element-wise in the feature fusion layer. The weighting coefficients are adaptively determined by trainable gating attention parameters to generate a joint feature vector that fuses image features and clinical features. The joint feature vector is fed into a time-series risk prediction model consisting of alternating stacks of fully connected layers and nonlinear activation functions. The output layer of the time-series risk prediction model contains twenty-four parallel neurons that correspond to the probability of malignant events occurring in each of the next twenty-four months, generating a continuous malignant probability trajectory prediction curve. Based on the malignancy probability trajectory prediction curve, the hierarchical decision rule engine is imported to calculate the integral value of the area under the curve within a preset time window, the average slope of the three to six months interval, and the detection results of the mutation inflection point. The risk level of the examinee is comprehensively determined. For high-risk levels, a priority biopsy channel instruction is triggered; for medium-risk levels, an image monitoring prompt is generated; and for low-risk levels, health education content and follow-up plans are pushed to generate hierarchical screening information.

[0018] It should be noted that the pre-diagnosis risk prediction intelligent agent, as an independently running software entity, can be implemented using Docker containerization technology or a lightweight virtual machine to ensure environmental isolation and ease of deployment with the hospital's existing information systems. Standardized data communication interfaces follow the HL7-FHIR medical data exchange standard or a custom protocol based on JSON-RPC. In the data preprocessing pipeline components, missing value imputation can employ mean imputation or a multivariate imputation algorithm based on K-nearest neighbors. Dark channel prior illumination unevenness correction specifically estimates and compensates for specular highlights generated on the moist mucosal surface of the oral cavity. Adaptive histogram equalization enhances mucosal texture contrast while avoiding excessive noise amplification through contrast limiting.

[0019] The standardized behavioral risk factor vector and the genetic biomarker abundance vector are concatenated along the feature dimension to form an initial state vector, which is then fed into a sequence encoder consisting of multiple gated recurrent units stacked sequentially. This encoder utilizes update and reset gate mechanisms to iteratively extrapolate the progressive cumulative effects of tobacco and alcohol exposure on mucosal barrier function over a 24-month time span, ultimately outputting a high-dimensional latent state feature vector characterizing the long-term intensity of the risk factors. Simultaneously, the three channels of the multispectral initial screening image are fed into three parallel branches of a lightweight convolutional feature extractor: the white light branch focuses on mucosal granularity and plaque edge morphology; the autofluorescence branch captures fluorescence attenuation regions caused by collagen breakdown; and the narrow-band blood oxygen branch extracts capillary density and abnormal oxygen saturation distributions. These three features are then concatenated and globally pooled to compress an image feature summary vector. The high-dimensional latent state vector and the image summary vector are then element-wise weighted and summed at the feature fusion layer using trainable gated attention parameters to generate a joint feature vector incorporating multi-source information. The joint vector is fed into a time-series risk prediction model consisting of alternating fully connected layers and nonlinear activation functions. Its output layer, with twenty-four parallel neurons, generates malignancy probability prediction curves for each of the next twenty-four months in a single operation. The hierarchical decision rule engine then calculates the area integral value of the curve within a preset window, the average slope over the first three to six months, and the inflection point of abrupt change, comprehensively determining the risk level of the examinee and triggering corresponding hierarchical screening instructions. The tiered decision-making system automatically classifies examinees into three mutually exclusive risk levels: individuals with an area under the curve (AUC) exceeding the first threshold and a proximal slope exceeding the second threshold, or who have shown a mutation inflection point, are classified as high-risk and automatically trigger a priority biopsy pathway, generating an alert instruction containing recommended anatomical quadrants for biopsy and pushing it to the clinical terminal; individuals with an AUC between the first and third thresholds and a stable slope without mutations are classified as intermediate-risk and receive a suggestion to enter the high-resolution imaging periodic monitoring process; individuals with an AUC below the third threshold and a flat or even declining curve are classified as low-risk and automatically generate patient-oriented oral precancerous lesion health education content and a long-term follow-up plan.

[0020] Furthermore, in a preferred embodiment of the present invention, the establishment of a precise assessment agent during diagnosis involves acquiring autofluorescence images and narrowband imaging images based on tiered screening results, performing pixel-level alignment using a differential homeomorphic registration algorithm to generate a fused data cube, and constructing an interpretable reasoning network by combining pathological slide abnormal proliferation grading labels to locate critical malignant regions and generate a risk stratification report. Specifically, this includes: Create an independent operating container for the intelligent agent for precise assessment during diagnosis, establish a standardized data communication interface with autofluorescence imaging equipment, narrowband imaging equipment and pathology information system, and deploy a differential homeomorphic deformation registration engine, an interpretable reasoning network model and a risk stratification report generation component within the operating environment to complete the construction of the intelligent agent; The system receives tiered screening information output by the pre-diagnosis risk prediction intelligent agent. When the system indicates that it is entering the image evaluation process, it sends a collection trigger signal to the autofluorescence imaging device and the narrowband imaging device through the standardized data communication interface to acquire autofluorescence images and narrowband imaging images and complete the initial synchronization of timestamps and spatial coordinate systems. Using narrowband imaging as a fixed reference image and autofluorescence image as a floating image, a velocity vector field is constructed through a differential homeomorphic deformation registration engine and the normalized mutual information measure is iteratively optimized. Under the condition of satisfying the topology preservation constraint, the autofluorescence image is mapped pixel by pixel to the coordinate space of the narrowband imaging image to generate a spatially aligned dual-channel fusion data cube. Synchronously, the pathological slice abnormal proliferation grade label corresponding to the current lesion is retrieved from the pathological information system through the standardized data communication interface. It is then mapped into a high-dimensional semantic vector through the embedding matrix and injected as a conditional modulation signal into the conditional batch normalization layer of the interpretable inference network. The channel mean and variance of the convolutional feature map are modulated by affine transformation. The dual-channel fused data cube is input into the interpretable reasoning network model. High-level semantic feature maps are extracted through a multi-layer convolutional cascaded feature encoding backbone network and conditional batch normalization modulation. After hierarchical attention aggregation operation, channel attention weights and spatial attention heatmaps are generated. The data is then diverted to the malignant risk classification head to output the malignant risk probability value, and an attribution heatmap is calculated to locate the critical region of malignant transformation. Risk levels are divided based on the numerical range in which the malignancy risk probability value falls. The attribution strength statistics within the critical area of ​​malignancy are extracted from the attribution heatmap and the confidence score is calculated. Targeted biopsy sites are planned for cases with confidence scores below the lower limit of the criticality. For cases with confidence scores that meet the standard, the malignancy risk probability, risk level, attribution information and pathological grade consistency description are assembled, and finally, a risk stratification report with attached evidence chain is generated.

[0021] It should be noted that the in-diagnosis precision assessment agent is activated and runs after receiving the tiered screening information output by the pre-diagnosis risk prediction agent and determining that the examinee needs to enter the high-resolution imaging assessment process. First, a standardized data communication interface with the autofluorescence imaging device, narrowband imaging device, and pathology information system is established in an independent running container, and the pre-trained differential homeomorphic registration engine, interpretable inference network model parameters, and risk stratification report generation component are loaded.

[0022] After the acquisition trigger signal is synchronously sent to both imaging devices, the autofluorescence image and the narrowband imaging image are captured in real time and initially aligned at the timestamp level. Since the autofluorescence signal reflects the submucosal collagen metabolism state while narrowband imaging focuses on the morphology of epithelial capillaries, there are inherent differences in their imaging depth and spatial resolution. Therefore, differential homeomorphic deformation registration is required to map the autofluorescence image pixel-by-pixel to the narrowband imaging coordinate space. This registration process uses the narrowband imaging image as a fixed reference image, constructs a velocity vector field parameterized by a control point grid, and iteratively optimizes the normalized mutual information measure under the topology-preserving constraint of a constant positive Jacobian determinant. In each iteration, the normalized mutual information measure between the floating image and the fixed image after distortion by the current deformation field is calculated, and this measure is used as the objective function to drive the parameters of the velocity vector field to update along the gradient ascent direction. The registration process terminates when the normalized mutual information value converges and the Jacobian determinant of the deformation field is positive at all pixel locations. At this point, each pixel coordinate in the autofluorescence image is precisely mapped to the corresponding anatomical location in the narrowband imaging image. Then, the spatially aligned autofluorescence grayscale matrix and the narrowband imaging grayscale matrix are stacked along the channel dimension to generate a fused data cube containing two channels. The first channel stores the autofluorescence intensity value, and the second channel stores the narrowband imaging reflectance value. Each spatial coordinate point of this data cube simultaneously contains two complementary pathological information types: the submucosal collagen metabolism status and the epithelial microvascular morphology at that location. The abnormal proliferation grading labels, synchronously retrieved from the pathology information system, are mapped into high-dimensional semantic vectors via an embedding matrix and injected into the conditional batch normalization layer of the interpretable inference network. Affine transformation modulates the channel mean and variance of the convolutional feature map, enabling the network to learn differentiated feature response patterns for different pathological manifestations of mild, moderate, and severe dysplasia.

[0023] The fused data cube, after undergoing multi-layer convolutional encoding and hierarchical attention aggregation, is distributed to two parallel downstream processing heads. The first processing head is a malignancy risk classification head, which compresses the spatially attention-weighted feature map into a fixed-length feature vector through global average pooling. This vector is then transformed layer by layer through two fully connected layers and a non-linear activation function, outputting a malignancy risk probability value between zero and one. This value represents the immediate risk estimate of the current lesion progressing to invasive cancer under the prior constraint of pathological grading. The second processing head is an attribution heatmap generation head. While the network performs forward propagation, it calculates the gradient of the malignancy risk probability value relative to the final convolutional feature map. Using the global average pooling result of this gradient along the channel dimension as the importance weight coefficient for each feature channel, it performs a weighted linear summation on the corresponding feature maps, generating a two-dimensional attribution heatmap with the same spatial resolution as the original input image. The numerical value of each pixel in this attribution heatmap reflects the pixel's contribution to the final malignancy risk prediction result; regions with higher values ​​indicate that the network relies more heavily on visual evidence when making a high-risk judgment. Subsequently, a pre-set statistical significance threshold was applied to the attribution heatmap for binarization segmentation, and connected components exceeding the threshold were extracted as spatial masks for the critical malignancy region. Finally, risk levels were defined based on the numerical range within which the malignancy risk probability value fell, and the mean and peak statistics of attribution intensity within the critical malignancy region were extracted from the attribution heatmap to calculate a confidence score. If the confidence score is below a preset lower limit, targeted biopsy sites are planned; if the confidence score meets the standard, a structured report containing malignancy risk probability, risk level labels, an attribution heatmap overlay mask, and a description of the consistency with pathological grading is assembled, providing clinicians with a decision-making reference accompanied by a visualized chain of evidence.

[0024] Furthermore, in a preferred embodiment of the present invention, the construction of the post-diagnosis dynamic monitoring intelligent agent, using the baseline state data of the lesion as the initial input to construct a digital twin model of the patient, acquiring images from each follow-up examination and extracting the difference in mucosal texture evolution and the drift of the vascular morphology index, constructing a malignancy trend prediction curve and generating a personalized prevention and intervention strategy, specifically includes: Create an independent operating container for the post-diagnosis dynamic monitoring agent, establish a standardized data retrieval interface with the image archiving and communication system, and deploy a patient digital twin model storage unit, a graph convolution-based registration and comparison module, a Gaussian process regression generator, and a treatment strategy generator within the operating environment; After deployment, the post-diagnosis dynamic monitoring agent sends a data request command to the in-diagnosis precision assessment agent to obtain the risk stratification report of the target patient and extract the baseline status data of the lesion and write it into the digital twin model storage unit to initialize the patient's digital twin model. When the target patient returns to the hospital for a follow-up examination, the current follow-up standardized endoscopic image sequence is obtained through the standardized data retrieval interface. Dark channel prior illumination non-uniformity correction, adaptive histogram equalization, and automatic cropping of the region of interest based on the baseline lesion contour are performed in sequence to obtain the current follow-up standardized image. Using a patient digital twin model to obtain baseline images, and inputting them with current follow-up standardized images, a graph convolution-based registration and comparison module is used to construct a spatial adjacency graph structure and calculate cross-graph attention weights. The output is a dense two-dimensional deformation field and a pixel-by-pixel texture residual map. The mean and standard deviation of the texture residuals are calculated within the contour of the critical region of malignancy as the difference in mucosal texture evolution. Vascular skeleton maps were extracted from baseline images and current follow-up standardized images, and vascular morphology analysis indices including mean branch angle, coefficient of variation of vessel diameter, and distribution density entropy were calculated. The vascular morphology analysis indices at the current time were subtracted from the indices at the baseline time and then weighted and summed to generate the vascular morphology index drift. The abnormal proliferation grading label encoding vector is obtained by using the patient's digital twin model. It is then input into the Gaussian process regression generator along with the difference in mucosal texture evolution and the vascular morphology index drift. All observation points accumulated before the current moment are used as the training set. The hyperparameters of the composite kernel function are updated by maximizing the marginal log-likelihood function to generate a malignant transformation trend prediction curve.

[0025] It should be noted that, firstly, an independent runtime container instance is created, a standardized data retrieval interface with the hospital's image archiving and communication system is established, and a digital twin model storage unit, a non-rigid registration and comparison module based on graph convolutional networks, a Gaussian process regression generator, and a treatment strategy generator are deployed within the runtime environment. After deployment, the agent sends a data request command to the in-diagnosis precision assessment agent to obtain the registered dual-channel fusion data cube, high-level semantic feature vector, spatial contour coordinate sequence of the malignant border region, and abnormal proliferation grade label contained in the target patient's risk stratification report. The above data is packaged into a lesion baseline state data block and written into the digital twin model storage unit, completing the initial creation of the patient's individualized digital twin model. This enables the continuous generation of a malignant trend prediction curve with confidence intervals by dynamically integrating the lesion baseline state with mucosal texture evolution and vascular morphology drift data from each follow-up visit, providing individualized quantitative basis for adaptive adjustment of follow-up intervals and optimization of treatment strategies.

[0026] When the patient returns for a follow-up examination as scheduled, the agent retrieves the standardized endoscopic image sequence collected during the examination through a standardized data retrieval interface. After preprocessing, the current standardized follow-up image is obtained. Subsequently, the baseline image is retrieved from the digital twin model and input together with the current follow-up image into a non-rigid registration and comparison module based on a graph convolutional network. This module divides the two images into overlapping image blocks and constructs a spatial adjacency graph structure. By calculating cross-graph attention weights, it locates regions corresponding to anatomical positions but whose texture has changed, and outputs a dense two-dimensional deformation field and a pixel-by-pixel texture residual map. Within the contour of the malignant critical region, the mean and standard deviation of the texture residuals are calculated as the mucosal texture evolution difference. At the same time, vascular skeleton maps are extracted from the baseline image and the current follow-up image, respectively. Three vascular morphology analysis indicators, namely the mean of branch angles, the coefficient of variation of vessel diameter, and the distribution density entropy, are calculated. The current indicator value is subtracted from the corresponding indicator value at the baseline time and then weighted and summed to generate the vascular morphology index drift. For example, if a patient's capillary branch angle significantly increases and distribution density entropy rises during a follow-up examination compared to the baseline time, it indicates that the microvascular network is evolving towards a disordered state. Finally, the abnormal proliferation grading label encoding vector is obtained through a digital twin model and input into a Gaussian process regression generator along with the mucosal texture evolution difference and the vascular morphology index drift. This generator uses all accumulated observation points before the current time as the training set, updates the kernel function hyperparameter (composed of a radial basis function kernel and a periodic kernel) by maximizing the marginal log-likelihood function, and generates a malignancy trend prediction curve with a confidence interval envelope, providing a dynamic risk assessment basis for subsequent treatment strategy generation.

[0027] Furthermore, in a preferred embodiment of the present invention, the construction of a multi-agent co-evolutionary framework, obtaining actual post-diagnosis malignancy outcome data and post-operative panoramic pathology results, optimizing the parameters of each agent by combining the desensitized full-cycle case data in the central shared memory bank, and using a federated learning mechanism for multi-agent closed-loop iterative optimization, specifically includes: Create a co-evolutionary management container, establish a two-way encrypted communication channel between the deployment nodes of the pre-diagnosis risk prediction agent, the in-diagnosis accurate assessment agent, and the post-diagnosis dynamic monitoring agent, deploy a central shared memory bank storage array within the container and divide it into a time-series prediction trajectory storage area, a pathological standard verification storage area, and a follow-up outcome annotation storage area, and configure a data desensitization engine for each partition; When any intelligent agent deployment node completes the full follow-up of the patient and generates malignant outcome annotation data or postoperative panoramic pathology report, a data return request is triggered. After integrity verification and desensitization processing, the data is uploaded to the corresponding logical partition of the central shared memory for persistent storage, and aggregated to form a standardized full-cycle case dataset with globally unified feature mapping. When the accumulated incremental data in the central shared memory reaches a preset threshold, the time-series calibration error is calculated by extracting paired samples of the pre-diagnosis predicted trajectory and follow-up outcome. The parameters of the gated cyclic unit sequence encoder and the weights of the multispectral image encoder convolutional layer in the pre-diagnosis risk prediction agent are updated in reverse. Versioned parameter patches are generated and pushed to each deployment node. Extract case samples from the diagnostic stage that delineate the critical area of ​​malignant transformation and obtain panoramic pathological verification. Compare the Dess similarity coefficient and Hausdorff distance between the attribution heat map area and the pathologically labeled area. Calculate the spatial positioning accuracy and probability calibration error. Update the hierarchical attention weights and conditional batch normalized affine transformation coefficients in the interpretable inference network in reverse. Extract population texture evolution and vascular morphology drift time-series trajectory according to the abnormal proliferation grade label, maximize the marginal log-likelihood function of each level to re-evaluate the radial basis function kernel length scale parameter and period kernel period length parameter of the Gaussian process regression generator, generate level-specific hyperparameters and push them to the post-diagnosis dynamic monitoring agent. Each deployment node calculates the incremental data gradient locally, encrypts it after differential privacy Gaussian noise perturbation, and uploads it to the federated aggregation server agent. It performs federated average aggregation based on sample size weighting to generate global parameter increments, updates the global model parameter snapshot in the central shared memory, and distributes it back to each node for replacement, thereby realizing multi-agent closed-loop iterative optimization.

[0028] It should be noted that the multi-agent co-evolutionary framework, acting as the central coordination layer connecting the three agents—pre-diagnosis, diagnosis, and post-diagnosis—creates an independent co-evolutionary management container on the central coordination server. It then establishes a bidirectional encrypted communication channel with each deployment node through a transport layer security protocol, ensuring the confidentiality and integrity of cross-node case data. The central shared memory storage array deployed within the container is divided into three logical partitions based on data type: a time-series prediction trajectory storage area, a pathology gold standard verification storage area, and a follow-up outcome annotation storage area. Each partition is configured with an independent data anonymization engine, automatically removing direct identifiers such as patient names and medical record numbers before data is written and blurring the date field to the month granularity. When any agent deployment node completes a patient's full follow-up cycle and generates malignant transformation outcome annotation data or a postoperative panoramic pathology report, it captures the event and generates a return request message. After hash integrity verification confirms that the data has not been tampered with, the node-side anonymization engine completes local anonymization and uploads the incremental data block to the corresponding logical partition for persistent storage.

[0029] When the cumulative incremental data volume of a certain partition in the central shared memory reaches the preset batch threshold, the co-evolutionary framework sequentially initiates a three-stage calibration process. The first calibration extracts paired samples from the pre-diagnosis stage where predicted trajectories have been generated and follow-up has revealed the actual outcome. It calculates the weighted deviation between the continuous rank probability score and the Blair score as the temporal calibration error, and uses this error to update the update and reset gate parameter matrices of the gated recurrent unit sequence encoder in the pre-diagnosis risk prediction agent, as well as the weights of each convolutional layer in the multispectral image encoder. The second calibration screens cases in the mid-diagnosis stage that have identified the critical region for malignancy and subsequently obtained panoramic pathological slides for verification. It compares the region defined by the attribution heatmap with the highest atypical hyperplasia region marked by the pathology department, calculates the Dessian similarity coefficient and Hausdorff distance to assess spatial positioning accuracy, and simultaneously calculates the ordinal regression deviation between the malignancy risk probability and the actual grading. This process is used to update the affine transformation coefficients of the hierarchical attention weight generation subnetwork and the conditional batch normalization layer in the interpretable inference network. The third calibration extracts the temporal trajectories of population texture evolution and vascular morphology drift in three levels: mild, moderate, and severe abnormal proliferation. It maximizes the marginal log-likelihood function of each level to re-estimate the radial basis function kernel length scale parameter and periodic kernel period length parameter in the Gaussian process regression generator, generating level-specific hyperparameters that are pushed to the post-diagnosis dynamic monitoring agent. Each deployment node calculates the incremental data gradient locally, applies Gaussian noise perturbation satisfying differential privacy constraints to the gradient vector, encrypts it, and uploads it to the federated aggregation server agent. The agent executes a federated average aggregation based on the sample size contributed by each node to generate global parameter increments, updates the global model snapshot in the central shared memory, and distributes it back to each node for atomic replacement, enabling continuous collaborative evolution of the three agents without directly sharing the original patient data.

[0030] Figure 2 This is a flowchart of a second method for a full-process auxiliary decision-making method for potential oral malignant diseases, provided by an embodiment of the present invention. like Figure 2 As shown, the present invention provides a second method flowchart for a full-process assisted decision-making method for potential oral malignant diseases, including: S202 is a treatment strategy generator built on a differentiable decision tree and embedded in the post-diagnosis dynamic monitoring agent. The decision tree is set to a depth of five layers. Each internal node outputs membership values ​​in multiple branch directions after linearly combining the input feature vector through the Sigmoid function. The leaf nodes store the treatment plan parameter vector, which includes the treatment method code, treatment intensity, treatment start time and follow-up interval. S204, the slope value of the malignancy trend prediction curve in the near time window, the predicted risk probability value at the current moment, the difference in mucosal texture evolution, the drift of the vascular morphology index and the contribution ratio of each sub-indicator are concatenated into an input feature vector, which is then input into the treatment strategy generator for forward propagation calculation, and the branch membership degree of each internal node is determined layer by layer. S206, based on the branch membership degree of each internal node, the arrival probability of each leaf node relative to the current input sample is calculated as a soft weight, and the treatment plan parameter vectors stored in all leaf nodes are convexly combined according to the soft weight to generate a set of individualized continuous treatment plan parameters for the target patient. S208, the training of the treatment strategy generator adopts joint optimization of multi-objective loss function. The first objective is to minimize the probability of malignant transformation within the prediction time window, the second objective is to minimize the preset tissue damage index corresponding to the treatment method, and the third objective is to maximize the treatment compliance score based on the prediction of patient behavior logs. S210 utilizes the gradient descent algorithm to simultaneously optimize the feature linear combination parameters of the internal nodes of the differentiable decision tree and the treatment plan parameter vector stored in the leaf nodes, so that the branch logic and the plan recommendation content converge in a coordinated manner towards maximizing malignant transformation inhibition, minimizing tissue damage, and promoting compliance optimization. S212 analyzes the distribution vector of treatment mode codes in the parameter set of individualized continuous treatment plan to identify the main recommended plan and alternative plans, constructs a structured treatment path diagram with the time axis as the main line, and generates a radar comparison chart of the marginal benefits and marginal risks of the main recommended plan and alternative plans. Finally, it integrates these to obtain a personalized prevention and intervention strategy.

[0031] It should be noted that the treatment strategy generator is activated after the post-diagnosis dynamic monitoring agent updates the malignancy trend prediction curve. This generator is built on a five-layer differentiable decision tree. Unlike traditional hard decision trees that execute a definite branch at each node, its internal nodes continuously map the linear combination of input feature vectors using a sigmoid activation function. The output sample is assigned a membership value between its left and right child nodes, with the membership value ranging from zero to one and the two values ​​being complementary. This makes the entire decision path a fully probabilistic weighted propagation process from the root node to all leaf nodes. The leaf nodes do not store a single-class label, but rather a four-dimensional treatment plan parameter vector containing the treatment method code, treatment intensity parameter, treatment start time, and follow-up interval. The treatment method code represents a discrete probability distribution over five options: observation and follow-up, local drug treatment, cryoablation, laser vaporization, and surgical resection. During forward propagation, the slope of the proximal end of the malignancy trend prediction curve, the predicted risk probability at the current moment, the difference in mucosal texture evolution, the vascular morphology index drift, and the contribution ratio of each sub-indicator are concatenated into an input feature vector. The branch membership degree of each internal node is calculated layer by layer and accumulated and multiplied along the tree structure to obtain the arrival probability of each leaf node as a soft weight. Then, the treatment plan parameter vectors of all leaf nodes are convexly combined and weighted according to the soft weight to generate a personalized continuous treatment plan parameter set for the patient. In the inference stage, the treatment mode encoding distribution in the continuous plan parameter set is parsed into the primary recommended plan with the highest probability and the second highest probability alternative plan. At the same time, a structured treatment path diagram with the time axis as the main line is constructed, marking the treatment start node and the follow-up evaluation node. A radar chart is generated to compare the marginal benefits and marginal risks of the primary and alternative plans from three dimensions: malignancy inhibition rate, tissue damage degree, and compliance score, providing clinicians with transparent and traceable treatment decision support. It is worth mentioning that the decision tree training employs end-to-end optimization using a multi-objective loss function. The first optimization objective is to estimate and minimize the probability of malignant transformation in the twelfth month under a given treatment response prediction network. The second optimization objective is to calculate and minimize the inner product of the treatment mode encoding distribution and the preset tissue damage index vector. The third optimization objective is to predict and maximize the treatment adherence score based on the patient's historical behavior logs. During training, gradient backpropagation synchronously updates the feature linear combination parameter matrix of the internal nodes and the treatment mode parameter vector stored in the leaf nodes, ensuring that the tree's branching logic and the recommended treatment content converge collaboratively towards the direction of optimal overall clinical benefit.

[0032] Figure 3An embodiment of the present invention provides a full-process auxiliary decision-making system 3 for potential oral malignant diseases. The system includes: a memory 301, a processor 302, and a communication interface 303. The memory 301 contains a full-process auxiliary decision-making method program for potential oral malignant diseases. When the processor 302 executes the full-process auxiliary decision-making method program for potential oral malignant diseases, it implements the steps of the full-process auxiliary decision-making method for potential oral malignant diseases as described above.

[0033] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A full-process auxiliary decision-making method for potential oral malignant diseases, characterized in that, include: Construct a pre-diagnosis risk prediction intelligent agent, collect patient behavioral risk factor scores, expression abundance of genetic susceptibility biomarkers and multispectral oral screening images, calculate the long-term cumulative effect of risk factors and perform malignant transformation probability trajectory prediction and graded screening of abnormal hyperplasia, and generate graded screening information; Establish a precise assessment agent during diagnosis. Based on the results of graded screening, obtain autofluorescence images and narrowband imaging images. Use the differential homeomorphic registration algorithm to perform pixel-level alignment and generate a fused data cube. Combine the abnormal proliferation graded labels of pathological sections to construct an interpretable reasoning network to locate the critical area of ​​malignancy and generate a risk stratification report. A post-diagnosis dynamic monitoring intelligent agent is constructed. The baseline status data of the lesion is used as the initial input to build a digital twin model of the patient. Images from each follow-up examination are obtained and the differences in mucosal texture evolution and the drift of vascular morphology index are extracted. A malignancy trend prediction curve is constructed and a personalized prevention and intervention strategy is generated. A multi-agent co-evolutionary framework was constructed to obtain actual post-diagnosis malignancy outcome data and post-operative panoramic pathological results. The parameters of each agent were optimized by combining the desensitized full-cycle case data in the central shared memory bank, and the multi-agent closed-loop iterative optimization was carried out using a federated learning mechanism.

2. The full-process auxiliary decision-making method for potential oral malignant diseases according to claim 1, characterized in that, The aforementioned pre-diagnosis risk prediction intelligent agent collects patient behavioral risk factor scores, genetic susceptibility biomarker expression abundance, and multispectral oral screening images. It calculates the long-term cumulative effect of risk factors and performs malignant transformation probability trajectory prediction and graded screening for abnormal hyperplasia, generating graded screening information, specifically including: Create an independent operating container for the pre-diagnosis risk prediction intelligent agent, establish a standardized data communication interface with the electronic medical record system, laboratory information system and multispectral oral screening equipment, and deploy data preprocessing pipeline components, time series risk prediction model and hierarchical decision rule engine within the operating environment; The standardized data communication interface is used to collect patient behavioral risk factor scores, genetic susceptibility biomarker expression abundance values, and multispectral oral screening images. These data are then input into the data preprocessing pipeline component to perform missing value imputation and maximum / minimum value normalization on the structured data, and to perform dark channel prior illumination non-uniformity correction and adaptive histogram equalization enhancement on the image data, thereby generating standardized input data. The normalized behavioral risk factor vector and the abundance vector of genetic susceptibility biomarkers are concatenated in the feature dimension to generate an initial state vector. Based on a sequence encoder consisting of multiple gated loop units stacked sequentially, the retention and discard ratio of historical accumulated information is adaptively controlled through update gate and reset gate mechanisms. After iterative deduction, a high-dimensional hidden state feature vector representing the long-term cumulative effect of risk factors is output. Multispectral oral screening images are input into a branched lightweight convolutional feature extractor. The first branch extracts the granular texture and plaque boundary morphology features of the mucosal surface, the second branch captures the distribution features of fluorescence signal attenuation regions, and the third branch extracts the spatial thermal distribution features of capillary density and oxygen saturation. The three feature maps are then compressed into a multispectral image feature summary vector after channel stitching and global average pooling. The high-dimensional hidden state feature vector and the multispectral image feature summary vector are summed element-wise in the feature fusion layer. The weighting coefficients are adaptively determined by trainable gating attention parameters to generate a joint feature vector that fuses image features and clinical features. The joint feature vector is fed into a time-series risk prediction model consisting of alternating stacks of fully connected layers and nonlinear activation functions. The output layer of the time-series risk prediction model contains twenty-four parallel neurons that correspond to the probability of malignant events occurring in each of the next twenty-four months, generating a continuous malignant probability trajectory prediction curve. Based on the malignancy probability trajectory prediction curve, the hierarchical decision rule engine is imported to calculate the integral value of the area under the curve within a preset time window, the average slope of the three to six months interval, and the detection results of the mutation inflection point. The risk level of the examinee is comprehensively determined. For high-risk levels, a priority biopsy channel instruction is triggered; for medium-risk levels, an image monitoring prompt is generated; and for low-risk levels, health education content and follow-up plans are pushed to generate hierarchical screening information.

3. The full-process auxiliary decision-making method for potential oral malignant diseases according to claim 1, characterized in that, The establishment of a precise assessment agent during diagnosis involves acquiring autofluorescence and narrowband imaging images based on tiered screening results. A differential homeomorphic registration algorithm is used for pixel-level alignment to generate a fused data cube. This cube is then combined with pathological slide abnormality grading labels to construct an interpretable reasoning network to locate critical malignant regions and generate a risk stratification report. Specifically, this includes: Create an independent operating container for the intelligent agent for precise assessment during diagnosis, establish a standardized data communication interface with autofluorescence imaging equipment, narrowband imaging equipment and pathology information system, and deploy a differential homeomorphic deformation registration engine, an interpretable reasoning network model and a risk stratification report generation component within the operating environment to complete the construction of the intelligent agent; The system receives tiered screening information output by the pre-diagnosis risk prediction intelligent agent. When the system indicates that it is entering the image evaluation process, it sends a collection trigger signal to the autofluorescence imaging device and the narrowband imaging device through the standardized data communication interface to acquire autofluorescence images and narrowband imaging images and complete the initial synchronization of timestamps and spatial coordinate systems. Using narrowband imaging as a fixed reference image and autofluorescence image as a floating image, a velocity vector field is constructed through a differential homeomorphic deformation registration engine and the normalized mutual information measure is iteratively optimized. Under the condition of satisfying the topology preservation constraint, the autofluorescence image is mapped pixel by pixel to the coordinate space of the narrowband imaging image to generate a spatially aligned dual-channel fusion data cube. Synchronously, the pathological slice abnormal proliferation grade label corresponding to the current lesion is retrieved from the pathological information system through the standardized data communication interface. It is then mapped into a high-dimensional semantic vector through the embedding matrix and injected as a conditional modulation signal into the conditional batch normalization layer of the interpretable inference network. The channel mean and variance of the convolutional feature map are modulated by affine transformation. The dual-channel fused data cube is input into the interpretable reasoning network model. High-level semantic feature maps are extracted through a multi-layer convolutional cascaded feature encoding backbone network and conditional batch normalization modulation. After hierarchical attention aggregation operation, channel attention weights and spatial attention heatmaps are generated. The data is then diverted to the malignant risk classification head to output the malignant risk probability value, and an attribution heatmap is calculated to locate the critical region of malignant transformation. Risk levels are divided based on the numerical range in which the malignancy risk probability value falls. The attribution strength statistics within the critical area of ​​malignancy are extracted from the attribution heatmap and the confidence score is calculated. Targeted biopsy sites are planned for cases with confidence scores below the lower limit of the criticality. For cases with confidence scores that meet the standard, the malignancy risk probability, risk level, attribution information and pathological grade consistency description are assembled, and finally, a risk stratification report with attached evidence chain is generated.

4. The full-process auxiliary decision-making method for potential oral malignant diseases according to claim 1, characterized in that, The construction of the post-diagnosis dynamic monitoring intelligent agent uses baseline lesion status data as initial input to build a digital twin model of the patient, acquires images from each follow-up examination, extracts the difference in mucosal texture evolution and the drift of vascular morphology index, constructs a malignancy trend prediction curve, and generates personalized prevention and intervention strategies, specifically including: Create an independent operating container for the post-diagnosis dynamic monitoring agent, establish a standardized data retrieval interface with the image archiving and communication system, and deploy a patient digital twin model storage unit, a graph convolution-based registration and comparison module, a Gaussian process regression generator, and a treatment strategy generator within the operating environment; After deployment, the post-diagnosis dynamic monitoring agent sends a data request command to the in-diagnosis precision assessment agent to obtain the risk stratification report of the target patient and extract the baseline status data of the lesion and write it into the digital twin model storage unit to initialize the patient's digital twin model. When the target patient returns to the hospital for a follow-up examination, the current follow-up standardized endoscopic image sequence is obtained through the standardized data retrieval interface. Dark channel prior illumination non-uniformity correction, adaptive histogram equalization, and automatic cropping of the region of interest based on the baseline lesion contour are performed in sequence to obtain the current follow-up standardized image. Using a patient digital twin model to obtain baseline images, and inputting them with current follow-up standardized images, a graph convolution-based registration and comparison module is used to construct a spatial adjacency graph structure and calculate cross-graph attention weights. The output is a dense two-dimensional deformation field and a pixel-by-pixel texture residual map. The mean and standard deviation of the texture residuals are calculated within the contour of the critical region of malignancy as the difference in mucosal texture evolution. Vascular skeleton maps were extracted from baseline images and current follow-up standardized images, and vascular morphology analysis indices including mean branch angle, coefficient of variation of vessel diameter, and distribution density entropy were calculated. The vascular morphology analysis indices at the current time were subtracted from the indices at the baseline time and then weighted and summed to generate the vascular morphology index drift. The abnormal proliferation grading label encoding vector is obtained by using the patient's digital twin model. It is then input into the Gaussian process regression generator along with the difference in mucosal texture evolution and the vascular morphology index drift. All observation points accumulated before the current moment are used as the training set. The hyperparameters of the composite kernel function are updated by maximizing the marginal log-likelihood function to generate a malignant transformation trend prediction curve.

5. The full-process auxiliary decision-making method for potential oral malignant diseases according to claim 1, characterized in that, The generation of personalized prevention and intervention strategies specifically includes: A treatment strategy generator is built based on a differentiable decision tree and embedded in the post-diagnosis dynamic monitoring agent. The decision tree depth is set to five layers. Each internal node outputs membership values ​​in multiple branch directions after linearly combining the input feature vector through the Sigmoid function. The leaf nodes store the treatment plan parameter vector including treatment method code, treatment intensity, treatment start time and follow-up interval. The slope value of the malignancy trend prediction curve in the near time window, the predicted risk probability value at the current moment, the difference in mucosal texture evolution, the drift of the vascular morphology index, and the contribution ratio of each sub-indicator are concatenated into an input feature vector, which is then input into the treatment strategy generator for forward propagation calculation to determine the branch membership degree of each internal node layer by layer. Based on the branch membership degree of each internal node, the arrival probability of each leaf node relative to the current input sample is calculated as a soft weight. The treatment plan parameter vectors stored in all leaf nodes are convexly combined according to the soft weight to generate a set of individualized continuous treatment plan parameters for the target patient. The training of the treatment strategy generator adopts joint optimization of multi-objective loss functions. The first objective is to minimize the probability of malignant transformation within the prediction time window, the second objective is to minimize the preset tissue damage index corresponding to the treatment method, and the third objective is to maximize the treatment compliance score based on the prediction of patient behavior logs. The gradient descent algorithm is used to simultaneously optimize the feature linear combination parameters of the internal nodes of the differentiable decision tree and the treatment plan parameter vector stored in the leaf nodes, so that the branching logic and the plan recommendation content converge in a coordinated manner towards maximizing malignant transformation inhibition, minimizing tissue damage, and promoting compliance optimization. The treatment mode coding distribution vector in the parameter set of individualized continuous treatment plan is analyzed to identify the main recommended plan and alternative plans. A structured treatment path diagram with the time axis as the main line is constructed, and a radar comparison chart of the marginal benefits and marginal risks of the main recommended plan and alternative plans is generated. Finally, the personalized prevention and intervention strategy is obtained by integrating the data.

6. The full-process auxiliary decision-making method for potential oral malignant diseases according to claim 1, characterized in that, The aforementioned multi-agent collaborative evolution framework acquires post-diagnosis data on actual malignant transformation outcomes and post-operative comprehensive pathological results. It then optimizes the parameters of each agent by combining this with desensitized full-cycle case data from a central shared memory bank, and utilizes a federated learning mechanism for closed-loop iterative optimization of the multi-agent system. Specifically, this includes: Create a co-evolutionary management container, establish a two-way encrypted communication channel between the deployment nodes of the pre-diagnosis risk prediction agent, the in-diagnosis accurate assessment agent, and the post-diagnosis dynamic monitoring agent, deploy a central shared memory bank storage array within the container and divide it into a time-series prediction trajectory storage area, a pathological standard verification storage area, and a follow-up outcome annotation storage area, and configure a data desensitization engine for each partition; When any intelligent agent deployment node completes the full follow-up of the patient and generates malignant outcome annotation data or postoperative panoramic pathology report, a data return request is triggered. After integrity verification and desensitization processing, the data is uploaded to the corresponding logical partition of the central shared memory for persistent storage, and aggregated to form a standardized full-cycle case dataset with globally unified feature mapping. When the accumulated incremental data in the central shared memory reaches a preset threshold, the time-series calibration error is calculated by extracting paired samples of the pre-diagnosis predicted trajectory and follow-up outcome. The parameters of the gated cyclic unit sequence encoder and the weights of the multispectral image encoder convolutional layer in the pre-diagnosis risk prediction agent are updated in reverse. Versioned parameter patches are generated and pushed to each deployment node. Extract case samples from the diagnostic stage that delineate the critical area of ​​malignant transformation and obtain panoramic pathological verification. Compare the Dess similarity coefficient and Hausdorff distance between the attribution heat map area and the pathologically labeled area. Calculate the spatial positioning accuracy and probability calibration error. Update the hierarchical attention weights and conditional batch normalized affine transformation coefficients in the interpretable inference network in reverse. Extract population texture evolution and vascular morphology drift time-series trajectory according to the abnormal proliferation grade label, maximize the marginal log-likelihood function of each level to re-evaluate the radial basis function kernel length scale parameter and period kernel period length parameter of the Gaussian process regression generator, generate level-specific hyperparameters and push them to the post-diagnosis dynamic monitoring agent. Each deployment node calculates the incremental data gradient locally, encrypts it after differential privacy Gaussian noise perturbation, and uploads it to the federated aggregation server agent. It performs federated average aggregation based on sample size weighting to generate global parameter increments, updates the global model parameter snapshot in the central shared memory, and distributes it back to each node for replacement, thereby realizing multi-agent closed-loop iterative optimization.

7. A full-process auxiliary decision-making system for potential oral malignant diseases, characterized in that, The system includes: a memory, a processor, and a communication interface. The memory contains a program for a full-process auxiliary decision-making method for potential oral malignant diseases. When the processor executes the program for the full-process auxiliary decision-making method for potential oral malignant diseases, it implements the steps of the full-process auxiliary decision-making method for potential oral malignant diseases as described in any one of claims 1-6.

Images