Throat abnormity real-time detection method for continuous image flow

Abstract

The invention relates to the technical field of throat anomaly detection, and discloses a throat anomaly real-time detection method for continuous image streams. The method comprises the following steps: performing multi-scale feature extraction on a continuously input throat image stream, separating dynamic texture features from static structure features through a space-time convolutional network, and generating a space-time separated feature tensor; a three-dimensional semantic map of the throat tissue is constructed based on the feature tensor, map nodes represent local tissue areas, edges represent anatomical association or pathological association between the areas, and map weights are dynamically updated through an online learning mechanism; receiving an anomaly detection request input by a user, and analyzing an anomaly type priority in combination with a historical diagnosis record and a medical knowledge base to generate a multi-dimensional detection semantic vector; in an initial detection focus range, a cross-modal feature alignment strategy is adopted, the matching degree of local region features and detection semantic vectors is calculated, and a candidate abnormal region set is screened.

Description

Technical Field

[0001] This invention relates to the field of throat abnormality detection technology, specifically a real-time throat abnormality detection method for continuous image streams. Background Technology

[0002] As a common passageway for the respiratory and digestive systems, the health of the pharynx directly affects the normal functioning of basic human physiology. Early detection and diagnosis of pharyngeal abnormalities, including inflammation, polyps, nodules, and tumors, are crucial for the prevention and treatment of respiratory diseases and early screening for pharyngeal cancer. Traditional pharyngeal examinations primarily rely on visual observation using a laryngoscopy. This method is highly dependent on the doctor's clinical experience and has limitations such as strong subjectivity, difficulty in quantifying and recording data, and inability to perform detailed analysis of dynamic processes. With the development of imaging technologies such as electronic laryngoscopy and high-speed photography, continuous, high-definition dynamic image sequences of the pharynx can be acquired, providing a rich data source for computer-aided diagnosis. However, most existing computer vision-aided diagnostic methods analyze single-frame static images, using algorithms such as image classification, target detection, or segmentation to identify abnormal regions. These methods face significant challenges when processing continuous streams of pharyngeal images.

[0003] The pharynx is a dynamically changing organ; its morphology, texture, and movement patterns undergo regular changes during physiological activities such as breathing, phonation, and swallowing. Many abnormalities are not statically visible but manifest as alterations in dynamic characteristics, such as disruption of vocal cord vibration symmetry and abnormal propagation of local mucosal waves. Static image analysis methods cannot capture these crucial diagnostic information related to movement. Furthermore, the pharyngeal region has a complex anatomy, containing multiple substructures such as the epiglottis, arytenoid cartilage, vocal cords, ventricular folds, and pyriform fossa. These structures are functionally interconnected and may also influence each other pathologically. Existing methods typically analyze the pharynx as a homogeneous region or perform simple region segmentation, lacking in-depth modeling of the interrelationships between anatomical structures. Different abnormality types may have similar image features, but their clinical significance and urgency differ. For example, inflammation and early-stage tumors may be similar in color and texture but require different clinical management strategies. Existing systems struggle to effectively incorporate prior medical knowledge to differentiate these clinical priorities. Simultaneously, fixed detection models cannot adapt to individual patient differences and specific diagnostic needs of physicians, lacking personalized analysis capabilities. Therefore, there is an urgent need for a method for detecting throat abnormalities that can process continuous dynamic image streams, integrate spatiotemporal information and medical knowledge, and perform personalized real-time analysis, in order to improve the objectivity, accuracy and efficiency of throat disease diagnosis. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time detection method for throat abnormalities in continuous image streams, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a real-time detection method for pharyngeal abnormalities in continuous image streams, the method comprising:

[0006] Multi-scale feature extraction is performed on a continuously input stream of throat images. Dynamic texture features and static structural features are separated by a spatiotemporal convolutional network to generate a spatiotemporally separated feature tensor.

[0007] A three-dimensional semantic map of pharyngeal tissue is constructed based on feature tensors, where map nodes represent local tissue regions and edges represent anatomical or pathological associations between regions, and the map weights are dynamically updated through an online learning mechanism.

[0008] It receives anomaly detection requests from users, combines historical diagnostic records and a medical knowledge base, analyzes the priority of anomaly types, and generates multi-dimensional detection semantic vectors.

[0009] The detection semantic vector is mapped to the embedding space of the three-dimensional semantic graph, and potential abnormal regions are located through a progressive neighborhood diffusion algorithm to form the initial detection focus;

[0010] Within the initial detection focus area, a cross-modal feature alignment strategy is adopted to calculate the matching degree between local region features and detection semantic vectors, and to screen the candidate abnormal region set;

[0011] Based on the spatiotemporal distribution characteristics of the candidate anomaly region set, regional clustering analysis is performed to merge anomaly regions with continuity, generating the final anomaly detection result.

[0012] Preferably, the step of performing multi-scale feature extraction on the continuously input throat image stream, separating dynamic texture features from static structural features through a spatiotemporal convolutional network, and generating a spatiotemporally separated feature tensor includes:

[0013] A pyramid-shaped convolutional structure is used to extract multi-scale static structural features from a single frame image, preserving tissue contour and morphological information at different resolutions.

[0014] By capturing dynamic texture changes between consecutive frames through temporal differential convolutional layers, dynamic feature sequences reflecting mucosal motion characteristics are generated.

[0015] Static structural features and dynamic texture features are concatenated along the channel dimension, and after feature normalization, a spatiotemporally separated joint feature tensor is output.

[0016] Preferably, the construction of a three-dimensional semantic atlas of pharyngeal tissue based on feature tensors, wherein atlas nodes represent local tissue regions, edges represent anatomical or pathological associations between regions, and the atlas weights are dynamically updated through an online learning mechanism, includes:

[0017] The joint feature tensor is spatially partitioned, and each partitioned region is mapped to a graph node and assigned an initial semantic label.

[0018] Calculate the feature similarity and motion consistency between adjacent nodes, establish anatomical association edges, and establish cross-regional pathological association edges based on the pathological rules annotated by the medical knowledge base;

[0019] Real-time monitoring of feature changes in newly added image streams; when node features deviate from the historical average by more than a threshold, triggering local graph reweighting and adjusting the confidence weight of associated edges.

[0020] Preferably, the step of receiving the anomaly detection request input by the user, combining historical diagnostic records and a medical knowledge base, parses the priority of the anomaly type, and generates a multi-dimensional detection semantic vector, including:

[0021] The text description entered by the user is converted into a standardized set of abnormality types using a medical entity recognition model;

[0022] Retrieve historical abnormal distribution data of the same patients in the historical diagnosis records and calculate the recurrence probability weight of various abnormalities;

[0023] By integrating the severity level of abnormalities and the recurrence probability weights defined in the medical knowledge base, semantic vector components reflecting the detection priority are generated.

[0024] Preferably, the step of mapping the detected semantic vector to the embedding space of the three-dimensional semantic graph and locating potential anomaly regions through a progressive neighborhood diffusion algorithm to form an initial detection focus includes:

[0025] Graph attention networks are used to embed graph nodes into a low-dimensional space, preserving the topological relationships and semantic associations between nodes.

[0026] The detected semantic vectors are projected onto the same embedding space, and their cosine similarity with each node is calculated. The top K nodes with high similarity are selected as seed nodes.

[0027] Centered on the seed node, the neighboring nodes are iteratively expanded along the associated edge direction. When the cumulative similarity growth rate of the neighboring nodes is lower than the preset threshold, the diffusion is terminated, forming the initial detection focus.

[0028] Preferably, within the initial detection focus area, a cross-modal feature alignment strategy is employed to calculate the matching degree between local region features and the detection semantic vector, and to filter the candidate abnormal region set, including:

[0029] Modal decoupling is performed on the node features within the initial detection focus to separate static morphological features from dynamic motion features;

[0030] Calculate the structural matching degree between static features and the detection semantic vector, and the motion matching degree between dynamic features and the detection semantic vector, respectively.

[0031] The two matching scores are weighted and fused, and nodes whose fusion scores exceed the dynamic threshold are selected to form a candidate abnormal region set.

[0032] Preferably, the step of performing region clustering analysis based on the spatiotemporal distribution characteristics of the candidate anomaly region set, merging continuous anomaly regions, and generating the final anomaly detection result includes:

[0033] Extract the spatial coordinates and temporal occurrence locations of candidate anomaly regions to construct a spatiotemporal distribution matrix;

[0034] High-density connected regions in the spatiotemporal matrix are identified based on density clustering algorithm, and spatially adjacent and temporally continuous abnormal regions are merged.

[0035] Calculate the anomaly confidence score for each merged region, remove isolated regions with confidence scores below the validation threshold, and output the results with labeled anomaly locations and types.

[0036] Preferably, the calculation of anomaly confidence for each merged region includes:

[0037] Statistical analysis of the frequency and duration of this region in historical image streams;

[0038] Extract the highest outlier match degree among all nodes in the region as the baseline value;

[0039] Based on the typical manifestations of this abnormality type in the medical knowledge base, the coverage overlap rate between regional features and typical features is calculated.

[0040] Preferably, the method further includes:

[0041] An incremental learning framework for an anomaly detection model is established, and the final detection results are compared with the physician's corrected annotations.

[0042] When the difference exceeds the acceptable range, the model parameters are fine-tuned, and the map node features related to the false detection or false negative regions are updated first.

[0043] Preferably, the fine-tuning of the trigger model parameters includes:

[0044] Locate the key graph nodes that cause the differences and trace back the feature extraction paths associated with them;

[0045] Adjust the weights of the convolution kernels at the corresponding levels in the spatiotemporal convolutional network to enhance feature suppression of false detection regions or feature enhancement of false negative regions;

[0046] Synchronously update the semantic labels and associated edge weights of relevant nodes in the 3D semantic graph.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] This invention achieves accurate and real-time detection of throat abnormalities by constructing a dynamic semantic atlas and introducing personalized detection semantic vectors. Multi-scale feature extraction is performed on continuous image streams, and a spatiotemporal convolutional network is used to separate dynamic texture features from static structural features. This simultaneously captures morphological and functional information of the tissue, providing a rich feature foundation for comprehensive analysis. The 3D semantic atlas constructed based on these features represents throat tissue as a network structure of nodes and edges, encoding not only local regional features but also reflecting the anatomical and pathological relationships between regions, enabling the system to understand the throat structure from a holistic perspective. An online learning mechanism allows the atlas weights to be dynamically updated with new data input, enhancing the model's adaptability.

[0049] This method receives user anomaly detection requests and generates multi-dimensional detection semantic vectors, transforming abstract detection requirements into concrete, computable feature representations. By combining historical diagnostic records and a medical knowledge base, the system can understand the clinical priority of different anomaly types, enabling personalized and targeted detection. The detection semantic vectors are mapped to a graph embedding space, and a progressive neighborhood diffusion algorithm is used to locate potential anomaly regions. This method intelligently focuses on relevant regions based on user needs, improving detection efficiency and targeting. A cross-modal feature alignment strategy is employed within the initial focal range to further refine the candidate region selection process, ensuring the accuracy of the detection results. Finally, cluster analysis is performed on candidate regions based on spatiotemporal distribution characteristics, merging continuous anomaly regions to generate structured final detection results, facilitating understanding and verification by clinicians. This method elevates throat anomaly detection from traditional single-frame image analysis to the spatiotemporal analysis level of continuous dynamic sequences, and from general detection to personalized, demand-driven precision detection, significantly improving the practicality and clinical value of computer-aided diagnosis. Attached Figure Description

[0050] Figure 1 A comparison chart of the accuracy of multi-scale feature extraction in throat abnormality detection;

[0051] Figure 2 A flowchart for multi-scale feature extraction and spatiotemporal separation of pharyngeal image streams;

[0052] Figure 3 A flowchart for constructing and updating a three-dimensional semantic map of pharyngeal tissue;

[0053] Figure 4 A graph showing the relationship between threshold and performance indicators for detecting throat abnormalities. Detailed Implementation

[0054] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 This invention provides a real-time detection method for throat abnormalities in continuous image streams. The method includes multi-scale feature extraction from the continuously input throat image stream, effectively separating dynamic texture features from static structural features using a spatiotemporal convolutional network to generate a feature tensor containing spatiotemporal information. Based on this feature tensor, a three-dimensional semantic map of the throat tissue is further constructed. Nodes in the map correspond to local tissue regions of the throat, and edges represent anatomical or pathological associations between regions. The entire map dynamically updates its weights through an online learning mechanism. Upon receiving an anomaly detection request from a user, the system combines the patient's historical diagnostic records and a pre-defined medical knowledge base to parse the detection priorities for different anomaly types and generate a multi-dimensional detection semantic vector. This detection semantic vector is mapped into the embedding space of the three-dimensional semantic map. A progressive neighborhood diffusion algorithm is used to locate potential abnormal regions in the image stream, forming an initial detection focus. Within the area defined by the initial detection focus, a cross-modal feature alignment strategy is employed to calculate the matching degree between the features of each local region and the detection semantic vector, thus selecting a set of candidate abnormal regions. Finally, based on the spatiotemporal distribution characteristics of the candidate anomaly region set, a region clustering analysis is performed to merge those anomaly regions that are continuous in space and time, generating and outputting the final anomaly detection results.

[0056] Example 1: See Figure 2This study extracts multi-scale features from a continuously input stream of pharyngeal images. A spatiotemporal convolutional network separates dynamic texture features from static structural features, generating spatiotemporally separated feature tensors. This process employs a pyramidal convolutional structure to extract multi-scale static structural features from a single frame. The pyramidal convolutional structure consists of multiple parallel convolutional branches, each configured with different scale kernels and pooling strategies. The largest-scale convolutional branch uses a large kernel and a large stride to capture the overall contour of the pharyngeal cavity and the positional relationships of key anatomical landmarks in a single frame. Medium-scale convolutional branches use standard-sized kernels to focus on the surface morphology and boundary features of local organs such as the vocal cords and arytenoid cartilage. The smallest-scale convolutional branch uses a small kernel and high-resolution feature maps to analyze the fine textures, capillary distribution, and minute bulges or depressions of the mucosa. The extraction process of multi-scale static structural features maintains the spatial correspondence of the feature maps in each branch. Upsampling operations restore the low-resolution feature maps to their original size, achieving layer-by-layer fusion of features at different scales. The output of static structural features contains a complete spatial information hierarchy, from macroscopic anatomical structure to microscopic tissue morphology.

[0057] Dynamic texture features are captured through temporal differential convolutional layers, designed as a three-dimensional convolutional architecture that performs sliding computation across the spatiotemporal dimensions of consecutive multi-frame image sequences. The weight initialization strategy of the convolutional kernels makes them sensitive to inter-frame differences, suppressing static background information while enhancing the response of moving regions. When processing the image stream, the temporal differential convolutional layer calculates the intensity change gradient of corresponding pixels between adjacent frames and accumulates motion patterns along the time axis. The dynamic texture feature sequence pays particular attention to the wave characteristics of the mucosal surface, the flow trajectory of secretions, and the periodic texture changes caused by vocal cord vibration. The output feature map of the temporal differential convolutional layer contains dual-channel encoding of motion amplitude and direction, forming a vector field describing the dynamic characteristics of the tissue. Dynamic texture features and static structural features have different update frequencies on the time scale; static structural features are updated only when the scene content changes significantly, while dynamic texture features are calculated in real time with each new image input frame.

[0058] Static structural features and dynamic texture features are concatenated along the channel dimension. The feature tensor before concatenation undergoes dimension alignment. Multi-scale static structural features are unified to the same feature dimension through cross-scale convolutional layers, while the dynamic features output by the temporal difference convolutional layer have their resolution adjusted through spatial pooling. The concatenated joint feature tensor then enters the feature normalization stage, employing a grouped normalization method to standardize the static and dynamic feature groups separately. Grouped normalization preserves the distribution characteristics of different modalities, preventing the strong response of static structural features from drowning out subtle changes in dynamic texture features. The tensor after feature normalization undergoes channel compression through a 1x1 convolutional layer, reducing feature redundancy and highlighting highly discriminative feature combinations. The spatiotemporally separated joint feature tensor is ultimately organized as a four-dimensional tensor, with the four dimensions corresponding to batch size, time step, spatial height, spatial width, and feature channels, respectively. Static structural features and dynamic texture features are arranged in segments along the feature channel dimension for easy indexing by subsequent processing modules.

[0059] The parameter configuration of the pyramid-shaped convolutional structure follows a progressive refinement principle. Large-scale branches use 7x7 convolutional kernels with a stride of 2, producing feature maps downsampled by 4 times. Medium-scale branches use 5x5 convolutional kernels with a stride of 1 to maintain the spatial resolution of the feature maps. Small-scale branches use 3x3 convolutional kernels with dilated convolutions to reduce the number of parameters while maintaining the receptive field. Each convolutional branch is followed by a batch normalization layer and a ReLU activation function to enhance the stability and non-linear representation capability of the features. Before fusion, the multi-branch features undergo attention weight adjustment; the attention mechanism calculates the importance weights of features at each scale based on the global context of the current frame. Feature fusion uses a weighted summation method, with scale features with high attention weights dominating the fusion result. The fused static structural feature tensor possesses scale invariance, adapting to throat image data with different shooting distances and resolutions. The temporal difference convolutional layer is designed to balance computational efficiency and feature discriminative power; the convolutional kernel spans 3 frames in the time dimension, balancing short-term motion capture and long-term motion trend analysis. The convolution operation uses separable convolution decomposition in the spatial dimension, decomposing the standard 3D convolution into a concatenation of spatial 2D convolution and temporal 1D convolution. Separable convolution reduces the number of parameters and the risk of overfitting, while maintaining the ability to jointly model spatiotemporal features. The activation function of the temporal difference convolutional layer is ReLU with a leakage parameter, avoiding the gradient vanishing problem and preserving negative motion information. The extraction process of dynamic texture features introduces a motion compensation mechanism, which eliminates global motion artifacts caused by endoscope lens movement by performing global registration on consecutive frames. Pure local tissue motion information is extracted through difference operations to form a description of mucosal dynamic characteristics independent of device motion.

[0060] Channel dimension management during the feature concatenation stage is achieved through a feature selection gateway, which evaluates the information entropy of each feature channel and its correlation with other channels. Redundant channels with low entropy or high correlation are suppressed or discarded, retaining the most representative feature combinations. The concatenated high-dimensional features are compressed using a bottleneck structure, which consists of two fully connected layers. The first fully connected layer expands the number of channels to twice the original size, and the second fully connected layer compresses them to the target number of channels. The channel compression process retains most of the feature information while significantly reducing the computational load, enabling subsequent processing modules to run in real time. The spatiotemporally separated joint feature tensor is finally stored as a multi-dimensional array. The first dimension of the array indexes the time step, the second and third dimensions index the spatial location, and the fourth dimension indexes the feature type. The feature type dimension clearly distinguishes between static and dynamic feature segments. Static feature segments contain multi-scale spatial information, while dynamic feature segments encode temporal motion patterns. This tensor organization structure provides standardized input for the construction of a 3D semantic graph. Static features are used for node attribute initialization, and dynamic features are used for calculating the weights of associated edges.

[0061] The training of the pyramid-shaped convolutional structure employs a multi-task learning strategy. The primary task is pharyngeal tissue region segmentation, while auxiliary tasks include anatomical landmark localization and image quality assessment. Multi-task learning shares convolutional backbone network parameters, forcing the network to learn more generalizable feature representations. Training data includes pharyngeal images under different lighting conditions, shooting angles, and pathological stages, enhancing the robustness of the feature extraction module. Data augmentation techniques are applied during training, including random rotation, color jitter, motion blur simulation, and occlusion simulation. Expanding the diversity of training data improves the adaptability of the feature extraction module to various interference factors in real-world clinical environments. The loss function is designed as a combination of weighted cross-entropy and the Dice coefficient to balance the uneven distribution of different tissue regions. Parameter optimization of the temporal difference convolutional layers focuses on the discriminative power of motion features; the training objective is to minimize the intra-class distance between normal mucosal motion and abnormal motion patterns. A contrastive learning strategy is applied to dynamic feature learning, bringing feature representations of adjacent frames within the same video sequence closer together and pushing feature representations of different cases or anatomical sites further apart. The representation learning of dynamic texture features introduces a cycle consistency constraint to ensure the stability of the feature extraction process against changes in motion speed. The motion feature descriptor is robust to slight changes in playback speed, avoiding feature distribution shifts caused by differences in endoscope advancement speed. The gradient calculation of temporal convolutions employs a truncated backpropagation algorithm to balance the requirements of long-range dependency modeling and training stability.

[0062] Example 2: See Figure 3A three-dimensional semantic atlas of pharyngeal tissue is constructed based on feature tensors, where nodes represent local tissue regions and edges represent anatomical or pathological associations between regions. The atlas weights are dynamically updated through an online learning mechanism. The construction process begins with spatial block processing of the joint feature tensor, dividing the feature map of each frame into 8×8 grid cells, each corresponding to a 4mm×4mm local region of the pharyngeal mucosa. Each block region is mapped to a node in the atlas. The initialization of the node embedding vector comes from the multi-scale feature aggregation of each block region, using a hybrid operation of max pooling and average pooling. The initial semantic label of the node is assigned based on the matching degree calculation between the region feature vector and the standard anatomical atlas. The matching degree calculation uses cosine similarity to measure the distance between the region features and the standard anatomical template features. When the similarity exceeds 0.85, the node is assigned the corresponding anatomical structure label, such as "anterior commissure of the vocal cords," "interarytenoid region," and "epiglottic laryngeal surface," among other standard anatomical terms.

[0063] The similarity of features between adjacent nodes is evaluated by calculating the Euclidean distance between their embedding vectors. Nodes with a distance less than a threshold of 0.3 are considered to have high feature similarity. Motion consistency is evaluated by comparing the change patterns of dynamic features of adjacent nodes over five consecutive frames, calculating the cross-correlation coefficient of the dynamic feature sequences. When the cross-correlation coefficient is greater than 0.7 and the feature Euclidean distance is less than 0.3, anatomical association edges are established between nodes. The initial weight of the anatomical association edges is set to the weighted average of the cross-correlation coefficient and feature similarity. The establishment of pathological association edges relies on pathological rules annotated in a medical knowledge base containing clinical records of over 2000 throat disease transmission paths. Pathological rules define the possible associations between different regions for specific abnormal morphologies. For example, the association strength between "vocal cord leukoplakia" and "laryngeal ventricle invasion" is 0.9, and the association strength between "vocal cord polyp" and "contralateral reactive hyperplasia" is 0.6. Based on these rules, edges are established between cross-regional nodes that may have pathological associations. The weights of the pathological association edges directly adopt the association strength values ​​defined in the knowledge base.

[0064] The online learning mechanism monitors feature changes in newly added image streams in real time. The system maintains a historical mean vector for each node's features, which is updated using a sliding window method, with the window size set to the feature data of the most recent 100 frames. When the current feature vector of a node deviates from its historical mean by more than 2.5 Mahalanobis distance thresholds during real-time processing, a local graph reweighting process is triggered. This process adjusts the confidence weights of all edges connected to a node, using an exponential decay function with the decay coefficient dynamically calculated based on the degree of feature deviation. For nodes with significant feature changes, the weights of their anatomical association edges with neighboring nodes are recalculated based on the current feature similarity and motion consistency assessment results. The weight updates for pathological association edges refer to the latest clinical guidelines on disease progression stages in the medical knowledge base to ensure that the graph weights are consistent with clinical practice. The system receives anomaly detection requests from users, combines historical diagnostic records and the medical knowledge base to parse the anomaly type priority, and generates multi-dimensional detection semantic vectors. User input is typically described in natural language text, which is parsed using a BERT-based medical entity recognition model. The medical entity recognition model was pre-trained on over 500,000 pieces of medical text data and is able to recognize 15 categories of throat-related medical terms. The model outputs a standardized set of abnormality types; for example, the input "check if there are polyps or leukoplakia on the vocal cords" is converted into a standardized set of {"vocal cord polyps", "vocal cord leukoplakia"}.

[0065] The system then retrieves the patient's historical diagnostic records, stored in an HL7-compliant electronic health record containing laryngoscopy results from the past five years. The frequency and spatiotemporal distribution of the same abnormality type in the historical records are analyzed to calculate the recurrence probability weights for each type of abnormality. The recurrence probability weights are calculated using a time decay function, assigning higher weights to more recently occurring abnormalities, with a decay coefficient set to decrease by 0.1 per month. The medical knowledge base predefines severity levels for various throat abnormalities, ranging from 1 to 5, with higher levels indicating greater clinical importance. For example, "invasive cancer" is marked as level 5, "precancerous lesions" as level 4, and "benign polyps" as level 2. The generation of the multi-dimensional detection semantic vector requires fusing the abnormality severity levels defined in the medical knowledge base with the calculated recurrence probability weights. The fusion formula uses a weighted geometric mean. Abnormal types with higher severity levels and higher recurrence probability weights have larger component values ​​in the semantic vector, and the number of dimensions in the semantic vector is consistent with the size of the standardized abnormality type set. Each dimension corresponds to a detection priority score for one abnormality type, with the score range normalized to 0-1. The resulting multi-dimensional detection semantic vector contains anomaly type priority information, providing a quantitative basis for subsequent graph retrieval and anomaly localization.

[0066] The grid size for spatial partitioning is optimized based on the anatomical characteristics of the pharynx. A finer 6×6 grid is used in the glottic region, while a sparser 4×4 grid is used in the supraglottic region. This adaptive grid partitioning strategy ensures the resolution of key functional areas while reducing redundant computation in non-critical areas. The physical size of each grid cell is dynamically adjusted according to the actual distance between the endoscope and the mucosal surface, with distance detection achieved through a stereo vision algorithm. When the distance between the endoscope and the mucosa is less than 2 cm, the grid cell size is automatically adjusted to 2 mm × 2 mm; when the distance is greater than 3 cm, the grid cell size is expanded to 6 mm × 6 mm. The initialization process of the node embedding vectors employs feature dimensionality reduction technology, compressing high-dimensional feature vectors into a 64-dimensional latent space through an autoencoder. The encoder part of the autoencoder consists of three fully connected layers with hidden layer dimensions of 256, 128, and 64, respectively, while the decoder part is symmetrically designed. The autoencoder is pre-trained on a large number of normal pharyngeal image features to learn discriminative low-dimensional representations. The dimensionality-reduced node embedding vectors retain the semantic information of the original features while improving the efficiency of subsequent graph computation.

[0067] The anatomical association edges are established considering the physiological characteristics of the pharyngeal tissues, and a dynamic weight adjustment mechanism is applied to the association edges between nodes in the vocal cord vibration region. The dynamic weights are adjusted based on the vibration phase synchronization, which is calculated using the instantaneous phase difference extracted by the Hilbert transform. When the phase difference is less than π / 6, the association edge weight is increased by 1.5 times; when the phase difference is greater than π / 2, the association edge weight is reduced to 0.8 times. This dynamic adjustment mechanism enables the atlas to accurately reflect the functional connectivity of the pharyngeal tissues. The knowledge base update mechanism for pathological association edges is linked to medical literature databases, automatically retrieving the latest clinical research from databases such as PubMed weekly. Newly discovered disease association paths are added to the knowledge base after review by clinical experts, with a review pass rate of approximately 85%. Knowledge base version control uses semantic version number management, incrementing the minor version number with each update to ensure the repeatability of the atlas construction process.

[0068] The feature change detection of the online learning mechanism employs a multivariate control chart method, simultaneously monitoring the mean shift and covariance changes of the feature vectors. The Hotelling T² statistic is used to detect multivariate anomalies; reweighting is triggered when the T² value exceeds the control limit based on the F-distribution. The confidence level of the control limit is set to 99% to avoid false triggers caused by normal physiological changes. Domain adaptation training of the medical entity recognition model uses pharyngeal specialty medical record data, incrementally training it based on the general medical BERT model. Incremental training employs an anti-forgetting learning strategy, preserving the model's memory of general medical knowledge while learning pharyngeal domain-specific expressions. The model achieves an F1 score of 0.92 on the pharyngeal domain entity recognition task, accurately recognizing various professional terms related to pharyngeal abnormalities.

[0069] The historical diagnostic record retrieval interface conforms to the FHIR standard and supports flexible GraphQL-based queries. The retrieval process considers the structure of the diagnostic reports; structured reports directly extract anomaly type fields, while unstructured reports utilize text mining techniques to extract key information. Text mining combines rule-based regular expression matching with deep learning-based named entity recognition to ensure the completeness of retrieval results. The generation process of multi-dimensional detection semantic vectors incorporates patient-specific factors, calculated based on the patient's age, gender, and occupational risk factors. For patients over 60 years of age, the weight of malignant tumor-related anomalies is automatically increased by 20%; for patients with a smoking history, the weight of precancerous lesions is increased by 15%. These personalized factors make the semantic vectors more reflective of individualized risk characteristics, improving the targeting of detection. The 3D semantic graph is stored using the graph database Neo4j. Node attributes include spatial coordinates, feature vectors, semantic labels, and timestamps. Edge attributes include association type, weight value, and confidence interval. The graph database supports efficient neighborhood queries and path analysis, providing underlying support for subsequent progressive diffusion algorithms. The serialized storage of the map uses the Apache Parquet format, which supports fast loading and incremental updates, meeting the time constraints of real-time processing.

[0070] The computational optimization of the local graph reweighting process employs sparse matrix operations, updating only the adjacency matrix elements related to changed nodes. A parallel computing architecture distributes the reweighting task across multiple GPU threads, with each thread processing one connected component. This distributed computing strategy ensures that the reweighting process can be completed within 16ms, even in large-scale graph scenarios, meeting real-time requirements. The severity level definitions in the medical knowledge base reference the clinical guidelines of the American Academy of Otolaryngology-Head and Neck Surgery and are updated every two years. Level adjustments have been validated through multi-center clinical studies, ensuring consistency between the level classification and clinical practice. Anomaly type priority calculation also considers the availability of medical resources; in resource-constrained situations, the priority of highly treatable anomalies is appropriately increased. The visualization interface for multi-dimensional detection semantic vectors provides interactive adjustment functionality, allowing clinicians to manually adjust the weight coefficients of specific anomaly types. Adjusted weight coefficients are protected using digital signature technology to prevent unauthorized modifications. The version history of semantic vectors fully records the physician ID, timestamp, and reason for each adjustment, meeting medical quality audit requirements.

[0071] Example 3: The detected semantic vector is mapped to the embedding space of a 3D semantic graph. A progressive neighborhood diffusion algorithm is used to locate potential anomaly regions, forming initial detection focal points. The mapping process uses a graph attention network to embed each node in the 3D semantic graph into a low-dimensional vector space. The graph attention network employs a multi-head attention mechanism, with each attention head containing eight parallel-computed attention subspaces. While encoding the node's own features, the graph attention network aggregates information from its first-order neighbor nodes, ensuring that the generated node embedding vector not only contains local features but also preserves the node's topological relationships and semantic associations in the graph. The graph attention network has three layers, and the output of each layer is stabilized through residual connections and layer normalization during the training process. The detected semantic vector is projected onto the same 128-dimensional low-dimensional space as the node embedding through a two-layer fully connected network. The first layer of the fully connected network compresses the 512-dimensional detected semantic vector to 256 dimensions, and the second layer further compresses it to 128 dimensions.

[0072] In the shared embedding space, the cosine similarity between the projected detection semantic vector and the embedding vector of each node in the graph is calculated. Cosine similarity measures the degree of matching between the node's semantics and the detection target. Similarity calculation employs a batch processing approach, calculating the similarity score between the detection semantic vector and all nodes at once. The score range is normalized to 0-1 using the sigmoid function. The top K nodes with the highest cosine similarity scores are selected as seed nodes. The value of K is dynamically adjusted based on the graph size, typically set to 5% of the total number of nodes. Seed nodes represent the initial regions most relevant to the detection request; these nodes may be spatially dispersed but semantically highly correlated with the target anomaly type.

[0073] The progressive neighborhood diffusion algorithm starts with these seed nodes and iteratively incorporates neighboring nodes along the established association edges in the graph. Each iteration of the diffusion process consists of three steps: similarity calculation, threshold determination, and neighborhood expansion. In the similarity calculation step, the average similarity of all nodes in the currently diffused region is calculated; in the threshold determination step, the current average similarity is compared with the value of the previous iteration; in the neighborhood expansion step, neighboring nodes that meet the criteria are added to the diffusion region. The diffusion process terminates when the increase in the average similarity of the entire region caused by the newly added neighboring nodes is lower than a preset growth rate threshold. The growth rate threshold is set to 0.05 to ensure that only regions with high semantic consistency are included. In this way, the algorithm can adaptively expand around the seed nodes to a semantically coherent region, forming an initial detection focus. This focus region covers tissue regions that are highly relevant to the detection target and anatomically or pathologically connected.

[0074] Within the initial detection focus area, a cross-modal feature alignment strategy is employed to calculate the matching degree between local region features and the detection semantic vector, thus filtering the candidate abnormal region set. The cross-modal feature alignment strategy first performs modal decoupling on the features of each node within the initial detection focus area. This decoupling operation separates the joint feature vector of the node into static morphological feature components and dynamic motion feature components. Static morphological feature components mainly describe the tissue's shape, color, texture, and other attributes that do not change rapidly over time; dynamic motion feature components describe the tissue's movement patterns, frequency, amplitude, and other temporal characteristics. Modal decoupling is achieved through two independent fully connected layers. The weight matrices of the fully connected layers are trained with orthogonal constraints to ensure the independence of static and dynamic features after decoupling.

[0075] The structural matching degree between static morphological feature components and the detected semantic vector is calculated separately. The structural matching degree is achieved by measuring the projection distance of the feature vector in the subspace. The projection subspace is constructed using principal component analysis, selecting the first 20 principal component directions to form the feature projection space. The motion matching degree between dynamic motion feature components and the detected semantic vector is calculated using a dynamic time warping algorithm. This algorithm can handle motion sequences of different lengths, aligning the dynamic features of nodes with typical motion patterns of anomaly types. Structural matching degree and motion matching degree are two different dimensions of measurement, requiring weighted fusion. The weights for this fusion are configured according to the characteristics of the anomaly type. For morphological anomalies such as polyps or leukoplakia, the structural matching degree weight is set to 0.7, and the motion matching degree weight is 0.3; for functional anomalies such as vocal cord paralysis, the motion matching degree weight is set to 0.8, and the structural matching degree weight is 0.2.

[0076] The matching degree fusion is calculated using the following formula:

[0077]

[0078] in: This represents the matching score after fusion. This represents the structural matching score. Indicates the motion matching score. The linear weighting coefficients represent the degree of structural matching. The linear weighting coefficients representing the degree of motion matching. This represents the weight coefficient of the interaction term. Interaction terms capture the synergistic effect between structural and motion features, improving the overall score when both are significant. The weight coefficients are determined through grid search optimization based on the anomaly type, with the optimization objective being to maximize the F1 score on the validation set.

[0079] After fusion, a comprehensive matching score is obtained. The system sets a dynamic threshold, which is adjusted based on the image stream feature quality and the statistical characteristics of the detection focus region. The dynamic threshold calculation is based on the statistical distribution of node matching scores within the detection focus region, using the mean plus 1.5 times the standard deviation as the benchmark. Nodes exceeding the threshold constitute a candidate anomaly region set. These nodes may be spatially discontinuous, but all exhibit feature patterns consistent with the target anomaly type. The graph attention network training adopts a contrastive learning strategy, using the triplet loss function to cluster similar nodes and separate dissimilar nodes. The training data contains 10,000 annotated throat image sequences, and the network convergence criterion is that the validation set loss changes by less than 0.001 for 10 consecutive epochs. The progressive neighborhood diffusion algorithm simultaneously monitors the average similarity growth rate and the contour coefficient, stopping diffusion when the contour coefficient is below 0.3. The algorithm uses a breadth-first search strategy, managing the node expansion order through a priority queue.

[0080] The modal decoupling fully connected layer achieves feature separation through orthogonal constraints, with a constraint strength coefficient set to 0.01. Decoupled features undergo independent batch normalization to eliminate distributional differences. Structural matching is calculated using incremental principal component analysis (PCA) subspaces, updating principal component directions every 100 new cases. Mahalanobis distance is used for projection distance to better reflect semantic relationships between features. The dynamic time warping algorithm employs Itakura parallelogram constraints on path slopes, with the cost function combining feature distance and first-order derivative distance. Weighted fusion coefficients are determined through Bayesian optimization, finding the optimal parameters after 100 iterations. The dynamic threshold is adaptively adjusted based on image sharpness, decreasing by 0.1 when sharpness is below 0.8. Sharpness is calculated using the variance of the Laplacian operator; images with a variance greater than 1000 are considered high-quality. Candidate anomaly regions are smoothed using morphological closing operations and are required to appear continuously in three consecutive frames. The graph attention network inference employs a node sampling strategy with a sampling ratio of 0.3, approximating attention weights through Monte Carlo sampling. The progressive neighborhood diffusion algorithm is implemented in parallel using multiple threads, and overlapping regions are assigned affiliations based on semantic similarity. The structural matching degree subspace dimension is determined through a feature value scree map, retaining 90% of feature variation information. The dynamic time warping algorithm employs a multi-scale strategy to accelerate computation, and the warping path is subject to second-order difference smoothing constraints. The matching degree fusion formula includes an interaction term, using a square root form to enhance feature synergy. The dynamic threshold is calculated using the median and absolute median difference when the distribution is non-normal. The candidate anomaly region set is stored using an R-tree spatial index, supporting efficient spatiotemporal queries.

[0081] Example 4: Based on the spatiotemporal distribution characteristics of the candidate anomaly region set, region clustering analysis is performed to merge anomaly regions with continuity. Each candidate anomaly node contains spatial coordinates (x, y) and a time frame number (t). This information is constructed into an N×3 dimensional spatiotemporal distribution matrix, where N is the total number of nodes. The coordinate data is normalized to eliminate the influence of resolution differences. A density clustering algorithm is used to analyze the spatiotemporal distribution matrix, with parameters set as follows: neighborhood radius 0.15, minimum sample size 5. The density clustering algorithm identifies spatially adjacent and temporally continuous node groups, merging anomaly nodes that meet the conditions into the same anomaly region instance. Anomaly confidence is calculated for each merged anomaly region, taking into account four factors: occurrence frequency (time decay weight 0.95 / second), duration ratio, highest matching degree benchmark value, and feature coverage overlap rate. The weights of each indicator are allocated as follows: frequency 0.3, duration 0.2, matching degree 0.3, and overlap rate 0.2. A weighted sum is used to obtain a confidence score between 0 and 1. The system sets verification thresholds based on the type of detection task: 0.6 for screening tasks and 0.8 for diagnostic tasks. Isolated regions below the threshold are removed, and the remaining high-confidence regions are labeled with their spatial location, temporal range, and anomaly type before being output as the final detection result. Confidence scores above 0.75 are marked as highly suspicious, 0.6-0.75 as moderately suspicious, and scores below 0.6 require further verification.

[0082] The calculation of anomaly confidence involves multiple quantitative indicators, the weights of which are determined based on clinical importance. Frequency of occurrence has a weight of 0.3, duration proportion has a weight of 0.2, the highest matching baseline has a weight of 0.3, and feature coverage overlap rate has a weight of 0.2. After weighted summation, a confidence score between 0 and 1 is obtained. Regions with scores above 0.75 are marked as highly suspicious, those between 0.6 and 0.75 are marked as moderately suspicious, and regions below 0.6 require further verification.

[0083] Table 1: Weight Allocation Table for Anomaly Confidence Calculation Indicators

[0084] Label Name Calculation method Weighting coefficient Standardization methods Frequency of occurrence Time decay weighted count 0.3 Divide by the maximum possible frequency Duration ratio Abnormal frame count / Total frame count 0.2 Original scale value Highest matching degree Maximum node matching degree within the region 0.3 sigmoid function normalization Feature overlap rate Intersection of actual features and typical features 0.2 Jaccard similarity coefficient

[0085] The parameters of the density clustering algorithm are optimized through grid search, and the clustering quality of different parameter combinations is evaluated on the validation set. Clustering quality is jointly evaluated using the silhouette coefficient and the Calinski-Harabasz index. The silhouette coefficient measures the compactness within clusters, while the Calinski-Harabasz index assesses the separation between clusters. The optimal parameter combination maximizes both indices while ensuring the number of clusters remains within a reasonable range, avoiding excessive fragmentation or insufficient coarse-grained clusters. Preprocessing of the spatiotemporal distribution matrix includes outlier detection and missing value imputation. Outliers are identified using the Isolation Forest algorithm, and missing values ​​are imputed using multiple imputation. The preprocessed matrix undergoes principal component analysis for dimensionality reduction, retaining 95% of the variance explained, projecting the three-dimensional spatiotemporal data into a two-dimensional space for easier visualization and analysis. Dimensionality reduction does not alter the fundamental structure of the data but significantly improves the computational efficiency of the clustering algorithm. Merged outlier regions undergo shape regularization using a combination of opening and closing operations in morphological operations. Opening eliminates small spikes, and closing fills internal holes, resulting in smoother and more natural region boundaries. The regularized area is labeled with the smallest bounding rectangle, and the coordinates and dimensions of the rectangle are recorded in the inspection results.

[0086] Typical features in the medical knowledge base are stored as feature vectors, with each anomaly type corresponding to a standard feature template. These feature templates are summarized by domain experts based on numerous clinical cases and are regularly updated to reflect the latest medical understanding. The overlap rate between actual and typical features is calculated using cosine similarity, with a similarity greater than 0.7 considered significantly consistent. A dynamic adjustment mechanism for the validation threshold considers image quality and pathological complexity; the threshold automatically decreases by 0.1 when the image quality index is below 0.6, and increases by 0.05 for complex cases (with multiple anomalies simultaneously). This adaptive mechanism balances detection sensitivity and specificity, maintaining stable performance across different scenarios. The final output format conforms to the DICOM standard, with anomaly region annotation information embedded in the annotation layer of the original image. The output includes the boundary coordinates of the anomaly region, confidence score, anomaly type code, and time range information, supporting seamless integration with hospital information systems. The detection report automatically generates a natural language description, describing the location, size, morphological features, and confidence level of the anomaly.

[0087] The parallel implementation of the density clustering algorithm employs a spatial partitioning strategy, dividing the spatiotemporal distribution matrix into multiple subspaces for parallel processing. The size of the subspaces is optimized using kd-trees, ensuring that each subspace contains a sufficient number of data points and that the computational load is balanced. Parallel processing reduces the clustering time from O(N^2) to O(NlogN), enabling the algorithm to process large-scale candidate region sets in real time. The calculation of anomaly confidence introduces a temporal consistency constraint, requiring anomaly regions to be continuous in the temporal dimension. Isolated regions that appear briefly, even with high matching degrees, are downweighted, while regions appearing consecutively for more than 10 frames receive a 20% increase in confidence. Temporal continuity checks utilize a run-length encoding algorithm to detect the duration of anomaly regions on the time axis. The region merging process considers anatomical constraints, limiting merging to adjacent regions within the same anatomical structure. Anomalies in the vocal cord region will not be merged with the epiglottis region, even if they are spatially close. Anatomical constraints are implemented through a predefined anatomical partition map, which divides the pharynx into 16 functional regions, each with independent merging rules. The feature coverage overlap rate is calculated using a multi-scale matching strategy, calculating similarity at both the global and local feature levels. Global feature similarity measures overall morphological consistency, while local feature similarity focuses on detailed texture matching. Multi-scale similarity is fused using a weighted average, with a global weight of 0.4 and a local weight of 0.6, highlighting the importance of detailed features. The validation threshold is learned through reinforcement learning, with the agent adjusting the threshold strategy based on feedback from historical detection results. The reward function comprehensively considers recall, precision, and clinical utility, finding the optimal threshold setting through long-term optimization. The threshold is updated every 24 hours to gradually adapt to the department's diagnostic habits and case characteristics.

[0088] The final visualization of the test results utilizes augmented reality technology, with abnormal regions overlaid as highlighted outlines on the original video stream. Outline colors are graded according to confidence level: red for high confidence, yellow for medium confidence, and green for low confidence. Doctors can interactively adjust display parameters to focus on abnormal regions at specific confidence levels. The test results are stored in a hierarchical structure: the raw data layer stores complete spatiotemporal distribution information, the summary layer contains key statistical indicators, and the report layer generates test conclusions that conform to clinical standards. This hierarchical storage supports data retrieval at different granularities, meeting diverse needs in clinical practice, research, and management. The region clustering algorithm is evaluated using the silhouette coefficient and the Davies-Bouldin index. The silhouette coefficient measures cluster cohesion, while the Davies-Bouldin index assesses inter-cluster separation. These two indicators work together to ensure that the clustering results are neither overly dispersed nor excessively merged, maintaining the integrity of abnormal regions while avoiding erroneous merging of different abnormal regions. Clustering quality is monitored in real time; when evaluation indicators fall below a threshold, clustering parameters are automatically adjusted and recalculated.

[0089] See Figure 4This figure illustrates a key analytical tool used in generating final detection results through cluster analysis based on the spatiotemporal distribution of candidate anomaly regions. The horizontal axis represents the validation threshold, indicating the confidence level for anomaly regions, while the vertical axis quantifies the accuracy and recall of the detection. The black solid dotted line represents the accuracy curve, which initially increases and then decreases with rising thresholds, reflecting the precision of anomaly detection at high thresholds. The gray square dotted line represents the recall curve, which continuously decreases with rising thresholds, reflecting the impact of high thresholds on missed anomaly detections.

[0090] The screening and diagnostic thresholds, marked with dashed lines, have clear clinical implications: in screening scenarios, they balance the breadth of anomaly detection with basic accuracy; in diagnostic scenarios, the focus is on reducing false positives. This figure clearly illustrates the trade-off logic between validation thresholds and detection performance, providing quantitative support for clinicians to select screening or diagnostic strategies as needed, and demonstrating the dynamic adaptability of this invention between anomaly detection accuracy and coverage.

[0091] Example 5: An incremental learning framework for the anomaly detection model is established. The final detection results are compared with the physician's corrected annotations. The comparison process is achieved by calculating the difference between the system output and the physician's annotations in terms of the location, range, and type of the abnormal region. The difference calculation adopts a multi-indicator comprehensive evaluation, including spatial overlap, type consistency, and boundary fit. Spatial overlap uses the Dice coefficient to measure the degree of overlap between the detected and labeled regions. Type consistency calculates the accuracy of the classification results, and boundary fit assesses the matching degree of the contour boundaries. When the comprehensive difference exceeds the preset acceptable range, the system triggers a model parameter fine-tuning process. The acceptable range is dynamically adjusted according to clinical needs, set to 0.15 for screening scenarios and 0.08 for diagnostic scenarios. Model parameter fine-tuning does not retrain the entire model, but rather makes targeted local adjustments, prioritizing the updating of atlas node features related to false positives or false negatives. The system will locate the key atlas nodes that led to the misjudgment based on the difference. These nodes correspond to normal regions that were misjudged as abnormal or real abnormal regions that were not detected. After locating the key atlas nodes causing the discrepancies, the system traces back the origins of these nodes during the feature extraction stage, identifying which layers and convolutional kernels in the spatiotemporal convolutional network computed their features. This backtracking mechanism is achieved by maintaining the feature propagation path; each atlas node records its corresponding original image region and the information of the convolutional layers it passed through.

[0092] The weights of convolutional kernels at corresponding layers in the spatiotemporal convolutional network are adjusted. For false positive regions, the weights of kernels that overly respond to their features are reduced to suppress the features of false positive regions. For false negative regions, the weights of kernels that capture their unique features are enhanced to improve the detection of potential abnormal features. Gradient descent is used to adjust the kernel weights to reduce the loss function between the model output and the physician's annotations. The loss function includes cross-entropy loss and boundary regression loss; cross-entropy loss handles classification errors, and boundary regression loss handles localization biases. The semantic labels and edge weights of relevant nodes in the 3D semantic graph are updated synchronously. Based on the physician's corrected annotations, the semantic labels of relevant nodes are corrected. The weights of the edges connected to these nodes, especially pathological edges, are re-evaluated to better reflect actual pathological associations. Through this incremental learning, the model can continuously learn from physician feedback and gradually adapt to the anatomical characteristics of different patients and new abnormal patterns.

[0093] Taking a specific case as an example, the system detected an abnormal region of 4mm × 3mm in the anterior middle part of the left vocal cord during a video sequence of a patient with vocal cord polyps. After review by the physician, it was found that the system misidentified a normal vocal cord vascular plexus as a polyp, while simultaneously missing a tiny 2mm polyp in the posterior part of the right vocal cord. The incremental learning framework first calculated the dissimilarity. The misdetection in the left vocal cord region led to a decrease in spatial overlap, while the missed detection in the right vocal cord region reduced type consistency. The combined dissimilarity exceeded the acceptable range, triggering fine-tuning of the model parameters. The system located the atlas node corresponding to the left vocal cord region. During feature extraction, this atlas node primarily underwent the second and third layers of a spatiotemporal convolutional network. Retrospective analysis revealed that certain convolutional kernels over-responded to vascular textures, incorrectly amplifying vascular patterns as polyp features. To address the misdetection issue, the fine-tuning process reduced the weights of the relevant convolutional kernels. Simultaneously, the atlas node corresponding to the posterior region of the right vocal cord was located, indicating that this atlas node was not sufficiently activated in the original detection. Analysis revealed that some convolutional kernels did not respond adequately to the features of small polyps, failing to effectively aggregate local features. To address the issue of missed detections, the fine-tuning process increased the weights of the corresponding convolutional kernels.

[0094] At the 3D semantic map level, the semantic labels of misjudged nodes were updated, correcting "vocal cord polyp" to "normal vascular plexus," and the semantic labels of missed nodes were updated from "normal mucosa" to "vocal cord polyp." The weights of the association edges between misjudged nodes and surrounding nodes were adjusted, reducing the weight of pathological association edges with typical polyp areas and increasing the weight of anatomical association edges with normal vascular areas. Missed nodes had new pathological association edges added with polyp characteristics, with reasonable initial weights. After the fine-tuning process, the system reprocessed the video sequences of specific cases. The second detection correctly identified the left vocal cord vascular plexus as a normal structure, while also detecting a small polyp area in the posterior part of the right vocal cord. The physician confirmed that the corrected results matched the actual situation, and the difference was reduced to an acceptable range. This learning experience was recorded in the incremental learning knowledge base, which accumulates and stores similar amendment examples, forming clinical experience assets.

[0095] The incremental learning framework employs a snapshot mechanism for version management, saving the complete model state before each fine-tuning. Snapshots include model parameters, graph structure, and hyperparameter configurations, allowing rollback to previous versions when needed. Version numbers are timestamped for easy tracking of model evolution. Snapshot data is compressed and stored, supporting historical data preservation. The learning rate setting for model parameter fine-tuning uses an adaptive strategy, dynamically adjusting the initial learning rate based on the rate of loss reduction. The learning rate is halved when loss decreases slowly and appropriately reduced when loss oscillates. This adaptive learning rate balances convergence speed and stability. An early stopping mechanism for fine-tuning training monitors validation set performance, terminating training when the validation set loss no longer improves. Quality management of physician-corrected annotations includes consistency checks, with multiple physicians cross-validating annotations for the same case. Annotation consistency is measured using statistical coefficients; annotations with excessively low coefficients require re-discussion and confirmation. High-quality labeled data is used for model fine-tuning to ensure the correctness of the learning direction. The annotation platform provides standardized tools and guidelines to reduce subjective differences among physicians.

[0096] The incremental learning framework's disaster recovery mechanism regularly backs up critical data, including model parameters, graph structure, and training logs, stored on servers in different physical locations. In the event of a system failure, it can quickly recover from the backup point, minimizing service interruption time. Long-term monitoring of model performance is achieved through a quality indicator dashboard, displaying trends in key metrics such as accuracy and recall. Metric data is updated regularly and supports filtering by time range, case type, and other dimensions. When a metric shows a sustained downward trend, the system automatically issues an alert, prompting a reassessment of the model's status. Integration of the incremental learning framework with the hospital information system is achieved through a standardized interface. The system automatically acquires new physician-annotated data, which is then anonymized and used for model optimization. Patient privacy is protected during the integration process; all personally identifiable information is encrypted during transmission and storage. Model updates are deployed using a blue-green deployment strategy. The new model version runs in shadow mode, processing the same data in parallel with the production environment without generating formal reports. Only after performance stability is confirmed in shadow mode is the new version officially switched to production mode. Uninterrupted service is ensured during deployment.

[0097] 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.

[0098] 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. A method for real-time detection of throat abnormalities oriented to a continuous stream of images, characterized by, The method comprises the following steps: Multi-scale feature extraction is performed on a continuously input throat image stream, dynamic texture features and static structure features are separated through a space-time convolution network, and a space-time separated feature tensor is generated; A three-dimensional semantic graph of the throat tissue is constructed based on the feature tensor, wherein a graph node represents a local tissue region, an edge represents an anatomical or pathological correlation between regions, and the graph weight is dynamically updated through an online learning mechanism; An abnormality detection request input by a user is received, historical diagnosis records and a medical knowledge base are combined, an abnormality type priority is analyzed, and a multi-dimensional detection semantic vector is generated; The detection semantic vector is mapped to an embedding space of the three-dimensional semantic graph, a potential abnormal region is located through a progressive neighborhood diffusion algorithm, and an initial detection focus is formed; Within the range of the initial detection focus, a cross-modal feature alignment strategy is adopted, the matching degree of local region features and the detection semantic vector is calculated, and a candidate abnormal region set is screened; According to the space-time distribution characteristics of the candidate abnormal region set, region clustering analysis is performed, and the abnormal regions with continuity are merged to generate a final abnormality detection result.

2. The method of claim 1, wherein, The multi-scale feature extraction on the continuously input throat image stream, the separation of dynamic texture features and static structure features through the space-time convolution network, and the generation of the space-time separated feature tensor comprise: Multi-scale static structure features of a single frame image are extracted by adopting a pyramid convolution structure, and tissue contour and morphological information at different resolutions are retained; Dynamic texture changes between consecutive frames are captured through a time difference convolution layer to generate a dynamic feature sequence reflecting mucosa movement characteristics; The static structure features and the dynamic texture features are spliced along a channel dimension, feature normalization processing is performed, and a space-time separated joint feature tensor is output.

3. The method of claim 2, wherein the method comprises: The three-dimensional semantic graph of the throat tissue is constructed based on the feature tensor, wherein a graph node represents a local tissue region, an edge represents an anatomical or pathological correlation between regions, and the graph weight is dynamically updated through an online learning mechanism, which comprises: The joint feature tensor is subjected to spatial block processing, each block region is mapped to a graph node, and an initial semantic label is assigned; Feature similarity and motion consistency between adjacent nodes are calculated, anatomical correlation edges are established, and pathological correlation edges between regions are established according to pathological rules labeled in the medical knowledge base; Feature changes in newly added image streams are monitored in real time, local graph reweighting is triggered when the node features deviate from the historical mean value by more than a threshold value, and the confidence weight of the correlation edge is adjusted.

4. The method of claim 3, wherein the method comprises: The abnormality detection request input by the user is received, the historical diagnosis records and the medical knowledge base are combined, the abnormality type priority is analyzed, and the multi-dimensional detection semantic vector is generated, which comprises: Text descriptions input by the user are converted into a standardized abnormality type set through a medical entity recognition model; Past abnormality distribution data of the same patient in the historical diagnosis records are retrieved, and recurrence probability weights of various abnormalities are calculated; The semantic vector components reflecting the detection priority are generated by fusing the abnormality severity levels defined in the medical knowledge base and the recurrence probability weights.

5. The method of claim 4, wherein the method comprises: The detection semantic vector is mapped to the embedding space of the three-dimensional semantic graph, the potential abnormal region is located through the progressive neighborhood diffusion algorithm, and the initial detection focus is formed, which comprises: The graph nodes are embedded into a low-dimensional space using a graph attention network, so as to reserve the topological relationship and semantic association between the nodes; The detection semantic vector is projected into the same embedding space, and the cosine similarity between the detection semantic vector and each node is calculated, and the top K nodes with high similarity are selected as seed nodes; The seed nodes are taken as centers, and neighbor nodes are iteratively expanded along the associated edges, and the diffusion is terminated when the growth rate of the cumulative similarity of the neighbor nodes is lower than a preset threshold, so as to form an initial detection focus.

6. The method of claim 5, wherein the method comprises: Within the range of the initial detection focus, a cross-modal feature alignment strategy is adopted to calculate the matching degree of the local region feature and the detection semantic vector, and a candidate abnormal region set is screened, including: The node features in the initial detection focus are modality decoupled, and the static morphological features and the dynamic motion features are separated; The structural matching degree of the static features and the detection semantic vector and the motion matching degree of the dynamic features and the detection semantic vector are respectively calculated; The two kinds of matching degrees are weighted and fused, and nodes with a fusion score higher than a dynamic threshold are screened to form the candidate abnormal region set.

7. The method of claim 6, wherein the method comprises: According to the spatiotemporal distribution characteristics of the candidate abnormal region set, a region clustering analysis is performed to merge the abnormal regions with continuity, and a final abnormal detection result is generated, including: The spatial coordinates and the time sequence appearance positions of the candidate abnormal regions are extracted to construct a spatiotemporal distribution matrix; Based on a density clustering algorithm, high-density connected regions in the spatiotemporal matrix are identified, and abnormal regions that are spatially adjacent and temporally continuous are merged; The abnormal confidence of each merged region is calculated, and isolated regions with a confidence lower than a verification threshold are removed, and a result of labeling the abnormal positions and types is output.

8. The method of claim 7, wherein the method comprises: The calculation of the abnormal confidence of each merged region includes: The frequency and the duration ratio of the region in a historical image stream are counted; The highest abnormal matching degree of all nodes in the region is extracted as a reference value; In combination with typical performance characteristics of the abnormal type in a medical knowledge base, an overlap rate of the region features and the typical features is calculated.

9. The method of claim 1, wherein, The method further includes: An incremental learning framework of the abnormal detection model is established, and the final detection result is compared with the physician correction label; When the difference exceeds an acceptable range, model parameter fine-tuning is triggered, and the graph node features related to the misdetected or missed regions are preferentially updated.

10. The method of claim 9, wherein the method comprises: The triggering of the model parameter fine-tuning includes: Key graph nodes causing the difference are located, and the feature extraction paths associated with the key graph nodes are traced back; The convolution kernel weights of the corresponding layers in the spatiotemporal convolution network are adjusted, so as to enhance the feature suppression of the misdetected regions or the feature enhancement of the missed regions; The semantic labels and the associated edge weights of the related nodes in the three-dimensional semantic graph are synchronously updated.

Images