Gynecological face diagnosis multi-modal feature fusion analysis system and method thereof

Abstract

The invention relates to the field of artificial intelligence medical treatment, in particular to a multi-modal feature fusion analysis system and method for gynecological face diagnosis, and aims to improve the early diagnosis precision of gynecological diseases, the system fuses multispectral images and symptom text data, and captures capillary morphology and hemodynamic parameters through a microcirculation feature extraction module; the region mapping analysis module is combined with the face region-gynecological viscera knowledge graph to generate viscera abnormality scores; the endocrine cycle modeling module constructs a dynamic model based on historical data and monitors cycle abnormity; the semantic attention fusion module fuses the symptom text and the microcirculation features by using a bidirectional cross attention mechanism to form a fusion feature vector; the knowledge reasoning feedback module is used for generating a diagnosis result and a personalized conditioning scheme based on dynamic knowledge graph multi-path reasoning, and optimizing a diagnosis and treatment path. The system overcomes the limitation of a traditional two-dimensional RGB image, effectively captures the subepidermal microcirculation state, and provides an important basis for early diagnosis of gynecological diseases.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence medicine, and in particular to a multimodal feature fusion analysis system and method for gynecological facial diagnosis. Background Art

[0002] Facial diagnosis in Traditional Chinese Medicine (TCM) is a crucial component of TCM diagnosis. By observing changes in facial color, shape, spots, and texture, combined with the corresponding facial zones of internal organs, facial diagnosis can infer the functional status of internal organs and the nature of the disease. In gynecological diseases, facial diagnosis particularly emphasizes the connection between qi and blood, and the liver, spleen, and kidney. The "Lingshu: Five Colors" book proposes that "each of the five colors indicates its own region," dividing the face into several regions, known as the "Mingtang Fanbi Tu," corresponding to the five internal organs. For example, the cheekbones are associated with the kidneys, the bridge of the nose (shangen) with the liver and gallbladder, the tip of the nose (zhuntou) with the spleen and stomach, and the chin (dige) with the reproductive area.

[0003] Existing technologies, such as publication number CN116130088A, disclose a multimodal facial diagnosis method. This method integrates facial image data and condition description text, determines facial features using a target detection network, obtains classification results using a classification network, and performs matching queries based on a facial diagnosis database. However, this existing technology has the following shortcomings:

[0004] 1. Analyzing facial features solely based on 2D RGB images fails to capture the subcutaneous microcirculatory state, changes in which are an important indicator of early manifestations of gynecological diseases.

[0005] 2. A unified network is used to process the entire face, without establishing a specialized analysis model for the gynecological facial area, which reduces the specificity of the diagnosis;

[0006] 3. Only static facial analysis is performed, lacking the ability to dynamically model how facial features change with the endocrine cycle, resulting in insufficient diagnostic sensitivity for cyclical gynecological diseases;

[0007] 4. The multimodal fusion method is simple and does not establish a precise correspondence between symptom descriptions and facial regions, making it impossible to achieve symptom-guided targeted analysis.

[0008] 5. The system uses a static knowledge base matching method to provide diagnosis and treatment recommendations, which lacks personalized adaptability and cannot continuously optimize the diagnosis and treatment pathway based on patient feedback. Summary of the Invention

[0009] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a multimodal feature fusion analysis system and method for gynecological facial diagnosis, which can achieve precision and personalization of gynecological facial diagnosis through microcirculation feature extraction, regional analysis, cycle dynamic modeling, semantically guided attention fusion and adaptive knowledge reasoning.

[0010] The present invention proposes a multimodal feature fusion analysis system for gynecological facial diagnosis, including:

[0011] Multispectral acquisition module, used to obtain facial multispectral image data and symptom description text data;

[0012] a microcirculation feature extraction module, connected to the multispectral acquisition module, for extracting microvascular morphological features, hemodynamic parameters and regional microcirculation index from the multispectral image data;

[0013] A regional mapping analysis module, connected to the microcirculation feature extraction module, is used to perform regional analysis on the microvascular morphological features based on a predefined facial region-gynecological organ association knowledge graph to generate an organ abnormality score;

[0014] An endocrine cycle modeling module, connected to the regional mapping analysis module, is used to construct an endocrine cycle dynamic change model based on historically collected microcirculatory characteristic time series data and generate cycle abnormality indicators;

[0015] a semantic attention fusion module, connected to the symptom description text data, the region mapping analysis module, and the endocrine cycle modeling module, respectively, for fusing the semantic features of the symptom text with the microcirculation abnormality features through a bidirectional cross-attention mechanism to generate a fused feature vector;

[0016] The knowledge reasoning feedback module is connected to the semantic attention fusion module, and is used to receive the fused feature vector, perform multi-path reasoning based on the dynamic knowledge graph, generate diagnosis results and personalized conditioning plans, and optimize the diagnosis and treatment path according to the feedback data.

[0017] Preferably, the multispectral acquisition module includes:

[0018] Narrowband spectroscopy unit, used to collect hemoglobin absorption characteristics in the 590nm-610nm band;

[0019] Near-infrared acquisition unit, used to capture the microvascular network at a depth of 2-3mm under the skin in the 750nm-850nm band;

[0020] Polarization filter unit, used to eliminate skin surface reflection interference;

[0021] The spectral fusion processing unit is used to fuse images of different bands to generate enhanced microcirculation visualization images.

[0022] Preferably, the microcirculation feature extraction module is configured to calculate the microcirculation index of 12 key areas of the face ( ):

[0023] = ( Concentration coefficient × microvascular density × blood flow velocity index) / regional baseline value,

[0024] in, The concentration coefficient indicates the ratio of hemoglobin concentration to the standard value, the microvessel density indicates the number of microvessels per unit area, the blood flow velocity index indicates the relative value of blood flow velocity in the microvessels, and the regional benchmark value indicates the standard reference value of the area under healthy conditions.

[0025] Preferably, the region mapping analysis module includes:

[0026] A region segmentation unit for segmenting the facial image into 13 regions;

[0027] The regional feature extraction network consists of 13 parallel sub-networks, each of which includes 3 layers of dynamic convolutional layers and a residual connection structure;

[0028] The organ association scoring unit is used to calculate the organ abnormality score based on the predefined facial region-gynecological organ association knowledge graph.

[0029] Preferably, the viscera-related scoring unit is configured to calculate the viscera-abnormality score:

[0030] Zang-fu abnormality score = Σ(regional deviation × ) / number of regions,

[0031] Among them, the regional deviation indicates the degree of difference between the microcirculation characteristics of each region and the standard model. Indicates the association weight coefficient of each region corresponding to the internal organs.

[0032] Preferably, the endocrine cycle modeling module includes:

[0033] Time series data acquisition unit, used to collect facial microcirculation data according to four key time points of the menstrual cycle;

[0034] 3D convolutional feature extraction network, used to extract the spatiotemporal changes of microcirculatory features;

[0035] Long short-term memory network, used to model the changing patterns of features at different cycle stages;

[0036] The periodic anomaly detection unit is used to calculate the deviation between the actual change trajectory and the benchmark model.

[0037] Preferably, the semantic attention fusion module includes:

[0038] Symptom semantic extraction unit, used to extract key semantic information from symptom description text;

[0039] Symptom-region mapping matrix, which stores the correlation weights of 350 standard symptoms and 13 facial regions;

[0040] A bidirectional cross-attention network is used to achieve bidirectional enhancement of symptom semantics to text-guided regional feature extraction and regional abnormality feedback to enhance symptom attention.

[0041] Preferably, the bidirectional crisscross attention network is configured to calculate:

[0042] Regional attention weight = Softmax (symptom semantic vector regional feature matrix),

[0043] Symptom attention = Softmax (regional anomaly vector symptom vector matrix),

[0044] Among them, the regional attention weight represents the importance of each region in the analysis, and the symptom attention represents the attention weight of each symptom description.

[0045] Preferably, the knowledge reasoning feedback module includes:

[0046] Dynamic knowledge graph, including a 3,500-node gynecological consultation knowledge network;

[0047] Graph attention network, used to achieve multi-path reasoning;

[0048] A reinforcement learning optimizer for optimizing diagnosis and treatment pathways based on patient feedback;

[0049] The Bayesian updating unit is used to update the individual sensitivity model based on the feedback data. Compute the posterior probability of the parameters, where are model parameters, Provide patient feedback data.

[0050] Multimodal feature fusion analysis method for gynecological facial diagnosis, including:

[0051] Collect facial multispectral image data and symptom description text data;

[0052] extracting microvascular morphological characteristics, hemodynamic parameters, and regional microcirculatory indexes from the multispectral image data;

[0053] Based on a predefined facial region-gynecological organ association knowledge graph, the microvascular morphological characteristics are regionalized and analyzed to generate an organ abnormality score;

[0054] Based on the historically collected microcirculatory characteristic time series data, a dynamic change model of the endocrine cycle is constructed to generate cycle abnormality indicators;

[0055] Through the bidirectional cross-attention mechanism, the semantic features of symptom text and microcirculatory abnormality features are fused to generate a fusion feature vector;

[0056] Multi-path reasoning is performed based on dynamic knowledge graphs to generate diagnostic results and personalized treatment plans, and the diagnosis and treatment path is optimized based on feedback data.

[0057] The beneficial effects of the present invention are:

[0058] 1. Through multispectral microcirculation imaging technology, it can penetrate the epidermis to capture the microvascular morphology and blood flow status at a depth of 2-3mm, detecting microcirculatory abnormalities 4-6 weeks in advance, greatly improving early warning capabilities;

[0059] 2. Based on the facial region-gynecological organ mapping model, it achieves regionalized and precise analysis, improving diagnostic positioning accuracy by 68%, and can accurately distinguish similar facial manifestations caused by different organs;

[0060] 3. Establish a dynamic face-to-face diagnosis model for endocrine cycles, which increases the diagnostic sensitivity of cyclical gynecological diseases by 75%, effectively identifying diseases such as premenstrual syndrome and luteal phase insufficiency that are difficult to diagnose through a single examination;

[0061] 4. An innovative two-way cross-attention fusion mechanism establishes a dynamic association between symptoms and facial regions, improving diagnostic targeting by 70% and consultation efficiency by 52%;

[0062] 5. Reinforcement learning is used to continuously optimize the diagnosis and treatment pathway. The system's accuracy continues to improve with increasing usage time. After 6 months, it has increased by 18% compared to the initial state, demonstrating its continuous learning ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Schematic diagram of the overall architecture of the multimodal feature fusion analysis system for gynecological facial diagnosis of the present invention;

[0064] Figure 2 This is a schematic diagram of the structure of the multi-spectral acquisition module of the present invention;

[0065] Figure 3 Schematic diagram of the microcirculation feature extraction process of the present invention;

[0066] Figure 4 This is a schematic diagram of the division of the face 13 regions according to the present invention;

[0067] Figure 5 Schematic diagram of the endocrine cycle dynamic model of the present invention;

[0068] Figure 6 Schematic diagram of the bidirectional cross-attention fusion mechanism of the present invention;

[0069] Figure 7 This is a schematic diagram of the dynamic knowledge graph structure of the present invention;

[0070] Figure 8 This is a flow chart of the multimodal feature fusion analysis method for gynecological facial diagnosis of the present invention;

[0071] Figure 9 This is the corresponding relationship diagram of the gynecological facial diagnosis mapping of the present invention. DETAILED DESCRIPTION

[0072] Please refer to Figure 1-9 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0073] Example 1: Overall architecture of the multimodal feature fusion analysis system for gynecological facial diagnosis

[0074] like Figure 1 As shown, the multimodal feature fusion analysis system for gynecological facial diagnosis provided by the present invention includes a multispectral acquisition module 1, a microcirculation feature extraction module 2, a regional mapping analysis module 3, an endocrine cycle modeling module 4, a semantic attention fusion module 5 and a knowledge reasoning feedback module 6.

[0075] Multispectral acquisition module 1 is used to acquire facial multispectral image data and symptom description text data. Multispectral image data contains spectral information of different bands, which can penetrate the epidermis and capture deep microcirculatory characteristics. Symptom description text data contains the patient's written or spoken description of their symptoms, providing the subjective information required for diagnosis.

[0076] Microcirculatory feature extraction module 2 is connected to multispectral acquisition module 1 and is used to extract microvascular morphological characteristics, hemodynamic parameters, and regional microcirculatory index from multispectral image data. This module processes multispectral images using a specific algorithm to identify the subcutaneous microvascular network structure of the face and quantify the microcirculatory status.

[0077] Regional Mapping Analysis Module 3, connected to Microcirculatory Feature Extraction Module 2, performs regionalized analysis of microvascular morphology based on a predefined knowledge graph of facial region-gynecological organ associations, generating organ abnormality scores. This module divides the face into 13 regions, establishing corresponding associations with specific gynecological organs, and assesses the functional status of these organs using regional microcirculatory characteristics.

[0078] Endocrine Cycle Modeling Module 4, connected to Regional Mapping Analysis Module 3, constructs a dynamic model of the endocrine cycle based on historically collected microcirculatory feature time series data and generates cycle abnormality indicators. This module continuously collects facial microcirculatory data to establish a personalized baseline model of facial features that change with the endocrine cycle and identify abnormal cycle change patterns.

[0079] The semantic attention fusion module 5 is connected to the symptom description text data provided by the multispectral acquisition module 1, the regional mapping analysis module 3, and the endocrine cycle modeling module 4. It uses a bidirectional cross-attention mechanism to fuse the semantic features of the symptom text with the features of microcirculatory abnormalities to generate a fused feature vector. This module achieves a two-way enhancement of "symptom-guided regional analysis" and "regional abnormalities enhance symptom understanding," improving diagnostic targeting.

[0080] The knowledge reasoning feedback module 6 is connected to the semantic attention fusion module 5. It receives the fused feature vectors, performs multi-path reasoning based on the dynamic knowledge graph, generates diagnostic results and personalized treatment plans, and optimizes the diagnosis and treatment path based on the feedback data. This module upgrades static matching to dynamic reasoning, enabling personalized customization and continuous optimization of diagnosis and treatment plans.

[0081] A complete data flow closed loop is formed between the various modules of the system: after the microcirculation features of the multispectral image are extracted, it enters the regional mapping analysis module to generate regionalized abnormality indicators; at the same time, historical data is used by the endocrine cycle modeling module to build a dynamic model, including data such as menstrual cycle and hormone levels, and outputs cycle deviation and abnormal fluctuation warning signals; regional abnormality indicators, cycle change characteristics and symptom text are integrated into a fused feature vector through the semantic attention fusion module; finally, the knowledge reasoning feedback module generates diagnostic results and conditioning plans based on the fused feature vector, and uses feedback data to continuously optimize system parameters.

[0082] Example 2: Structure and Function of Multispectral Acquisition Module

[0083] like Figure 2 As shown, the multi-spectral acquisition module 1 includes a narrow-band spectrum unit 11 , a near-infrared acquisition unit 12 , a polarization filter unit 13 and a spectrum fusion processing unit 14 .

[0084] The narrowband spectral unit 11 is used to collect hemoglobin absorption characteristics in the 590nm-610nm band. This band is the region where the hemoglobin absorption peak occurs and is extremely sensitive to the hemoglobin content in blood vessels. Experiments have shown that after penetrating the skin's surface, light waves in the 590nm-610nm band are selectively absorbed by hemoglobin in microvessels, forming a distinct absorption signature. Preferably, the narrowband spectral unit 11 is implemented using a bandpass filter in conjunction with a broadband light source, with a central wavelength set at 600nm and a full width at half maximum of 10nm.

[0085] The near-infrared acquisition unit 12 is used to capture the subcutaneous microvascular network at a depth of 2-3 mm in the 750nm-850nm wavelength range. Near-infrared light has strong penetrating power, capable of penetrating the epidermis and reaching the dermis, enabling imaging of deep microvascular networks. Preferably, the near-infrared acquisition unit 12 utilizes an 850nm near-infrared light-emitting diode array for illumination, coupled with an image sensor with a corresponding filter, to capture images of the deep subcutaneous microvascular distribution.

[0086] The polarization filter unit 13 is used to eliminate interference from skin surface reflections. These reflections can cause highlights in the image, affecting the accuracy of microvascular feature extraction. Using cross-polarization technology, the polarization filter unit 13 filters out surface-reflected light while retaining deep-lying scattered light, enhancing microvascular contrast. Preferably, this unit includes a polarizer perpendicular to the polarization direction of the light source, allowing only light with a changed polarization direction after scattering to pass through.

[0087] The spectral fusion processing unit 14 is used to fuse images of different wavelengths to generate an enhanced microcirculatory visualization image. This unit receives multi-band image data collected by the narrowband spectral unit 11 and the near-infrared acquisition unit 12, and fuses them using a specific algorithm to form a composite image containing rich microcirculatory information. Preferably, this unit uses a weighted fusion algorithm:

[0088] ,

[0089] in, is the fused image, is a narrowband spectral image, is a near-infrared image, 、 is the weight coefficient of each band, the value range is 0-1, and the preferred values ​​are 0.4 and 0.5 respectively. is the cross enhancement coefficient, with an optimal value of 0.1, which is used to enhance the areas where both bands are significant.

[0090] Multispectral acquisition module 1 also includes an image preprocessing subunit (not shown) that normalizes, suppresses noise, and enhances the fused image to ensure stability during subsequent processing. This subunit uses adaptive histogram equalization to enhance image contrast and a non-local means filtering algorithm to suppress noise while preserving microvascular edge detail.

[0091] Example 3: Structure and Function of Microcirculation Feature Extraction Module

[0092] like Figure 3 As shown, the microcirculation feature extraction module 2 is configured to calculate the microcirculation index (MCI) of 12 key areas of the face. The microcirculation index is a comprehensive quantitative indicator reflecting the regional microcirculation status, and the calculation formula is:

[0093] ,

[0094] in, It represents the hemoglobin concentration coefficient, which is calculated based on the spectral absorption characteristics of the 590nm-610nm band. The value range is 0.5-1.5, and the normal value is 1.0; It represents the microvascular density, which is defined as the number of microvascular pixels per unit area (mm2) and is statistically calculated after extracting microvascular pixels using an image segmentation algorithm; Represents the blood flow velocity index, which is estimated by the temporal variation characteristics of the pixel grayscale value in the microvessels and reflects the blood flow velocity in the microvessels; It represents the regional benchmark value, which is the standard reference value of the area under healthy conditions and is obtained through statistics of a large number of normal samples.

[0095] In practical applications, the microcirculation feature extraction module 2 performs the following steps:

[0096] First, a modified Frangi filter is applied to the multispectral fusion image to enhance microvascular structure. The Frangi filter, based on Hessian matrix eigenvalue analysis, effectively enhances tubular structures. This embodiment improves on the traditional Frangi filter by introducing multiscale analysis and directional enhancement, increasing the sensitivity of microvascular detection.

[0097] Next, a U-Net convolutional neural network was used for microvascular segmentation. The network input was an enhanced multispectral fusion image, and the output was a microvascular segmentation mask. The network consisted of a five-layer encoder and a five-layer decoder, using a residual connection architecture to ensure accurate microvascular segmentation. The training data consisted of 2,000 manually labeled microvascular segmentation samples, and transfer learning was used to adapt the model to gynecological facial diagnosis applications.

[0098] Next, microvascular morphological features are extracted based on the segmentation results, including microvascular length, diameter, tortuosity, number of branches, and number of endpoints. These features are calculated using a thinning algorithm and connected domain analysis to effectively characterize the morphological properties of the microvascular network.

[0099] The microcirculatory index was then calculated for 12 key areas: the glabella, the eye area, the zygomatic area, the nasal alar area, the philtrum, the lip area, and the mandibular area. The baseline value for each area was determined based on age groups, taking into account the impact of age on microcirculatory status.

[0100] Finally, a matrix of inter-regional MCI ratios was constructed to quantify microcirculatory heterogeneity. Studies have shown that under healthy conditions, the relative ratios of MCI across different regions remain relatively stable, and abnormal inter-regional ratios often indicate dysfunction of specific zang-fu organs.

[0101] Microcirculation feature extraction module 2 also extracts microvascular abnormality morphology indices to identify specific microcirculatory abnormalities, such as moles, bloodshot eyes, and telangiectasias. These abnormalities are often highly correlated with specific gynecological conditions. For example, telangiectasias in the nose are common in patients with abnormal hormone levels.

[0102] Example 4: Structure and Function of the Regional Mapping Analysis Module

[0103] like Figure 4 As shown, the region mapping analysis module 3 includes a region segmentation unit 31, a region feature extraction network 32 and an organ association scoring unit 33.

[0104] The region segmentation unit 31 is used to segment the facial image into 13 regions. These 13 regions are determined based on traditional Chinese medicine face diagnosis theory and modern medical research, and each region corresponds to a specific gynecological organ. Figure 4 As shown in the figure, these 13 regions include: 1-left glabella, 2-right glabella, 3-left periocular, 4-right periocular, 5-left zygomatic, 6-right zygomatic, 7-nasal tip, 8-left ala, 9-right ala, 10-philtrum, 11-left lip, 12-right lip, and 13-mandibular. Regional segmentation uses predefined anatomical landmarks as a basis and combines them with an active contour model to achieve accurate segmentation.

[0105] The regional feature extraction network 32 consists of 13 parallel sub-networks, each of which is responsible for processing the microcirculatory features of a specific region. Each sub-network includes three layers of dynamic convolutional layers and a residual connection structure. The specific structure is as follows:

[0106] The first dynamic convolution layer uses kernel size adaptively adjusted based on regional characteristics, ranging from 3×3 to 7×7. Smaller kernels are used for smaller regions, such as the nose, while larger kernels are used for larger regions, such as the cheekbones. This layer outputs 32 feature maps.

[0107] In the second dynamic convolution layer, the convolution kernel size is also adaptively adjusted, and this layer outputs 64 feature maps.

[0108] The third dynamic convolution layer has a fixed 5×5 convolution kernel and outputs 128 feature maps.

[0109] Each convolutional layer is followed by batch normalization and ReLU activation, and a residual connection structure is used to preserve the original microcirculatory morphology. The network also integrates an attention mechanism to enhance key point features within the region. The 13 sub-networks process in parallel, and the final output forms a regionalized feature tensor with a size of [13, 128].

[0110] The organ association scoring unit 33 is used to calculate the organ abnormality score based on the predefined facial region-gynecological organ association knowledge map. This unit first calculates the deviation of each region's microcirculation characteristics from the standard model, and then applies the weight coefficient to calculate the organ abnormality score:

[0111] ,

[0112] in, The score for organ abnormality ranges from 0 to 100, with higher values ​​indicating more severe abnormalities. is the deviation of the i-th region, calculated by the Euclidean distance between the current microcirculation characteristics and the standard model; is the correlation weight coefficient of the viscera corresponding to the i-th region, reflecting the importance of the region in judging the state of the viscera; It is the number of regions related to the organ.

[0113] The facial region-gynecological organ association knowledge graph is constructed based on traditional Chinese medicine theory combined with modern clinical research. According to the classic Chinese medicine book "Lingshu·Five Colors", the Mingtang Fanbi diagram divides the face into different regions, which correspond to the human organs as follows:

[0114] Mingtang (nose); the tip of the nose (zhuntou) corresponds to the spleen; the bridge of the nose (nian shou) corresponds to the liver; the root of the nose (shangen) corresponds to the heart; the sides of the nose correspond to the stomach; other facial areas, the forehead (ting) corresponds to the head and face; the space between the eyebrows (quezhong) corresponds to the lungs; the inner sides of the two cheekbones correspond to the small intestine; the area near the philtrum corresponds to the bladder and uterus (zichu); the area in front of the ears (bi) corresponds to the kidneys;

[0115] The viscera-correlation scoring unit 33 also generates a facial region abnormality heat map, visually displaying problem areas. The heat map uses a color gradient to indicate the degree of abnormality in each region, with red indicating severe abnormality and blue indicating normal, providing clinicians with an intuitive diagnostic reference.

[0116] Example 5: Detailed Explanation of the Scoring Mechanism of the Zang-Fu Association Scoring Unit

[0117] According to claim 5, the organ correlation scoring unit 33 is configured to calculate an organ abnormality score. The organ abnormality score is a core indicator for quantifying the functional status of each gynecological organ, and is calculated based on the deviation degree and correlation weight of regional microcirculatory characteristics.

[0118] Regional deviation The calculation of is based on the improved Mahalanobis distance formula:

[0119] .

[0120] in, is the microcirculation characteristic vector of the current i-th region; is the characteristic mean vector of the region in a healthy state; is the covariance matrix of the regional features, reflecting the correlation of each dimension of the features; It is the age correction coefficient, which increases with age and reflects the impact of age on the microcirculation baseline.

[0121] The value (facial region-gynecological organ association weight coefficient) is determined based on the following three factors:

[0122] 1. Theoretical basis of Traditional Chinese Medicine: the strength of the description of the relationship between regions and internal organs in traditional Chinese medicine facial diagnosis theory;

[0123] 2. Clinical statistical correlation: The correlation coefficient between regional abnormalities and organ diseases based on statistical analysis of 5,000 clinical data;

[0124] 3. Expert consensus weight: The consensus weight assessed by 30 senior gynecologists and traditional Chinese medicine experts.

[0125] The three factors are combined to determine the final value:

[0126] ,

[0127] in, The weight of traditional Chinese medicine theory, is the statistical correlation weight, is the expert consensus weight.

[0128] After the organ abnormality score calculation is completed, the system determines the abnormality level based on the score value:

[0129] 0-20 points: normal range;

[0130] 21-40 points: mild abnormality, observation is recommended;

[0131] 41-60 points: moderately abnormal, further examination is recommended;

[0132] 61-80 points: significant abnormality, timely medical treatment is recommended;

[0133] 81-100 points: Severe abnormality, immediate intervention is recommended.

[0134] The scoring results also generate a heat map of facial region abnormalities. This uses a dual-color scheme: blue (cold) indicates normality and red (hot) indicates abnormality. The color depth corresponds to the degree of abnormality. The heat map is overlaid on the original facial image to visually demonstrate the distribution of abnormal regions.

[0135] In addition, the organ-related scoring unit 33 also calculates the matching degree of organ abnormality patterns and assists in the diagnosis of specific gynecological diseases by comparing the similarity between the current scoring pattern and the typical disease pattern. The system has built-in typical scoring patterns for 50 common gynecological diseases and uses cosine similarity to calculate the matching degree:

[0136] ,

[0137] Where A is the current score vector and B is the typical disease pattern vector. Disease patterns with a matching degree exceeding 0.85 will be output as diagnostic references.

[0138] Example 6: Structure and Function of the Endocrine Cycle Modeling Module

[0139] like Figure 5 As shown, the endocrine cycle modeling module 4 includes a time series data acquisition unit 41, a 3D convolutional feature extraction network 42, a long short-term memory network 43 and a cycle anomaly detection unit 44.

[0140] The time series data acquisition unit 41 is used to collect facial microcirculation data at four key time points during the menstrual cycle. These four key time points are: the menstrual period (approximately 3-7 days), the follicular phase (approximately 7-14 days, from the end of menstruation to ovulation), the ovulatory phase (approximately 1-2 days, usually occurring around 14 days before the next menstrual period), and the luteal phase (approximately 14 days, from ovulation to the next menstrual period). These time points correspond to key changes in hormone levels during the endocrine cycle, and facial microcirculation characteristics often exhibit regular changes at these points. The unit uses intelligent reminders to ensure that users perform tests at key time points and records any deviations between the actual test date and the ideal time point for subsequent data calibration.

[0141] The 3D convolutional feature extraction network 42 is used to extract the spatiotemporal variations of microcirculatory features. This network utilizes a 3D convolutional architecture, taking multiple consecutively acquired facial microcirculatory feature maps as input and outputting spatiotemporal feature vectors. The network comprises four 3D convolutional layers with a kernel size of 3×3×3, outputting 32, 64, 128, and 256 feature maps, respectively. The convolution operation operates simultaneously in both spatial and temporal dimensions, effectively capturing the temporal variations of microcirculatory features.

[0142] The Long Short-Term Memory Network 43 is used to model the variation of features across different cyclical stages. This network receives the spatiotemporal feature vectors output by the 3D convolutional feature extraction network and builds a sequence model using LSTM (Long Short-Term Memory) units. The LSTM network consists of two layers, each with 128 hidden units, effectively modeling long-term dependencies and capturing variations across multiple cycles. A periodic feature decay function is designed within the network to balance the influence of historical data:

[0143] ,

[0144] in, is the weight at time point t, is the current time point, is the attenuation coefficient, and the preferred value is 0.1 / cycle, which means that the weight of historical data decreases by about 10% after each complete cycle.

[0145] The cycle anomaly detection unit 44 is used to calculate the deviation between the actual change trajectory and the baseline model. This unit first constructs a personalized normal cycle change baseline model based on the user's historical data, then calculates the deviation between the microcirculatory characteristics at each time point in the current cycle and the baseline model to generate a cycle anomaly index. The anomaly index calculation formula is:

[0146] ,

[0147] in, It is a period anomaly index with a value range of 0-1. The larger the value, the higher the degree of anomaly. is the weight coefficient of the i-th key time point, reflecting the importance of this time point; is the feature vector of the i-th time point in the current cycle; is the feature vector of the i-th time point in the baseline model; represents the L2 norm.

[0148] When a cycle anomaly indicator exceeds a preset threshold (default 0.3), the system triggers an abnormal cycle warning and recommends appropriate diagnosis and treatment based on the abnormal pattern. Simultaneously, the system dynamically updates individual baseline models to adapt to long-term trends and avoid misjudgments caused by model rigidity.

[0149] The endocrine cycle modeling module 4 also provides a cycle prediction function. Based on historical data of more than three consecutive cycles, it predicts the start time of the next cycle and possible abnormal risks, and provides users with personalized health management suggestions.

[0150] Example 7: Structure and Function of Semantic Attention Fusion Module

[0151] like Figure 6 As shown, the semantic attention fusion module 5 includes a symptom semantic extraction unit 51, a symptom-region mapping matrix 52 and a bidirectional cross-attention network 53.

[0152] The symptom semantic extraction unit 51 is used to extract key semantic information from symptom description text. This unit uses the BERT (Bidirectional Encoder Representations from Transformers) pre-trained language model, which has been fine-tuned with medical-domain text to accurately understand medical terminology and colloquial expressions. The unit first performs pre-processing on the input text, such as word segmentation and stop word removal. It then uses the BERT model to extract text feature vectors, which are then mapped to fixed-dimensional symptom semantic vectors through a fully connected layer. To improve the ability to recognize specific terms in the gynecology field, the unit integrates a domain dictionary containing more than 2,000 gynecological terms, giving these terms higher weights during the word segmentation and feature extraction stages.

[0153] The symptom-region mapping matrix 52 stores the correlation weights between 350 standard symptoms and 13 facial regions. These 350 standard symptoms cover common gynecological symptoms, such as "heavy menstrual flow," "dysmenorrhea," and "abnormal vaginal discharge." Each element in the matrix represents the strength of the correlation between a specific symptom and a specific facial region, ranging from 0 to 1. This matrix is ​​constructed based on Traditional Chinese Medicine theory and modern medical research, and the weights are optimized using machine learning methods based on 3,000 clinical cases. The symptom-region mapping matrix provides prior knowledge of the association between textual symptoms and facial regions, guiding the initial weight assignment of the attention mechanism.

[0154] The bidirectional cross-attention network 53 is used to achieve bidirectional enhancement of symptom attention by guiding regional feature extraction with symptom semantic vectors and regional abnormality feedback. The network mainly realizes bidirectional information flow through two attention calculations:

[0155] Symptom-guided regional attention calculation:

[0156] ,

[0157] ,

[0158] in, is the symptom semantic vector, is the regional feature matrix, and is the learnable parameter matrix, α is the attention weight matrix, and R' is the weighted regional feature matrix.

[0159] Regional abnormalities guide symptom attention calculation:

[0160] ,

[0161] ,

[0162] in, is the regional anomaly vector, indicating the degree of anomaly in each region, and is the learnable parameter matrix, is the symptom attention weight vector, is the weighted symptom semantic vector.

[0163] The bidirectional cross-attention network53 uses a multi-head attention mechanism to simultaneously focus on multiple sets of feature relationships. It employs eight attention heads, each of which independently calculates attention weights to capture associations across different semantic dimensions. The network also employs an iterative optimization strategy, performing three rounds of attention calculations to enable regional and symptom attention to mutually reinforce each other.

[0164] The semantic attention fusion module 5 also includes a multimodal feature adaptive fusion unit (not shown in the figure) for the final fusion of the attention-weighted regional features and symptom features. This unit uses a variable-scale fusion layer to dynamically adjust the modal weights based on the information quality:

[0165] ,

[0166] in, is the fusion feature vector, is the symptom status weight, dynamically calculated:

[0167] ,

[0168] For the sigmoid function, ensure The value is between 0-1. and are learnable parameters.

[0169] In addition, this module introduces a contrastive learning framework to enhance the representation of inter-modal correlation features and uses the InfoNCE (Information Noise-Contrastive Estimation) loss function for training to maximize the mutual information of positive sample pairs and improve the quality of multimodal fusion.

[0170] Example 8: Attention calculation mechanism of bidirectional cross-attention network

[0171] According to claim 8, the bidirectional cross-attention network 53 is configured to calculate the regional attention weight and the symptom attention. These two calculation processes together constitute the core of the bidirectional cross-attention mechanism, achieving mutual enhancement of symptom and regional features.

[0172] The detailed process of regional attention weight calculation is as follows:

[0173] First, the symptom semantic vector S is mapped to the query space through linear transformation:

[0174] ,

[0175] in, is the learnable query projection matrix, is the symptom semantic vector dimension, is the internal dimension of the attention mechanism.

[0176] Next, the regional feature matrix R is mapped to the key space through linear transformation:

[0177] ,

[0178] in, is the learnable key projection matrix, is the regional characteristic dimension.

[0179] Then, the dot product of the query vector and the key matrix is ​​calculated and scaled to stabilize the gradient:

[0180] ,

[0181] Finally, the Softmax function is applied to normalize the attention weights:

[0182] ,

[0183] Obtained This is the regional attention weight matrix, which indicates the importance of each region under the current symptom semantics. The dimension is [1, 13], corresponding to 13 facial regions.

[0184] Symptom concern calculations are similar, but in the opposite direction:

[0185] First, the regional anomaly vector Mapping to the query space through linear transformation:

[0186] .

[0187] in is the learnable query projection matrix, is the regional anomaly vector dimension.

[0188] Next, the symptom semantic vector matrix S is mapped to the key space through linear transformation:

[0189] ,

[0190] in, is the learnable key projection matrix.

[0191] Then, we compute the dot product of the query vector and the key matrix and scale it:

[0192] ,

[0193] Finally, the Softmax function is applied to normalize the attention weights:

[0194] ,

[0195] Obtained This is the symptom attention vector, which represents the attention weight of each symptom description in the abnormal situation of the current area. The dimension is [1, number of symptom words], corresponding to each word in the symptom description.

[0196] In practice, the bidirectional cross-attention network 53 uses a multi-head attention mechanism to perform the above calculation process eight times in parallel, using a different projection matrix each time, which can capture attention relationships from different semantic subspaces:

[0197] ,

[0198] ,

[0199] in, express An attention head, is the number of attention heads, here it is 8. The final attention weight is obtained by concatenating the results of each head and passing it through a linear layer:

[0200] ,

[0201] ,

[0202] in, and is the output projection matrix.

[0203] Example 9: Structure and Function of Knowledge Reasoning Feedback Module

[0204] like Figure 7 As shown, the knowledge reasoning feedback module 6 includes a dynamic knowledge graph 61, a graph attention network 62, a reinforcement learning optimizer 63 and a Bayesian update unit 64.

[0205] Dynamic Knowledge Graph 61 contains a 3,500-node gynecological facial diagnosis knowledge network. These nodes include facial diagnosis feature nodes (such as "abnormal zygomatic microcirculation"), symptom nodes (such as "dysmenorrhea"), disease nodes (such as "endometriosis"), constitution nodes (such as "Qi and blood deficiency"), and conditioning measure nodes (such as "blood circulation promotion and stasis removal"). Nodes are connected through different types of edges, such as "manifested as," "cause of," and "applicable to." The knowledge graph is constructed based on Traditional Chinese Medicine (TCM) theory and modern medical knowledge, encompassing two major knowledge systems: TCM theory (1,500 nodes) and modern medical knowledge (2,000 nodes), with a mapping relationship established between the two.

[0206] The Graph Attention Network 62 is used to implement multi-path reasoning. Based on the Graph Attention Network (GAT) architecture, this network enables message passing and information aggregation on a knowledge graph. The network comprises three graph attention layers, each with eight attention heads, adaptively learning the important relationships between nodes. Multi-path reasoning refers to the network's ability to simultaneously consider multiple possible reasoning paths and dynamically adjust path weights based on context. This mechanism supports reasoning under uncertainty and can handle the ambiguity and uncertainty common in medical diagnosis. Each reasoning path is associated with a confidence score, and the final diagnosis results provide both the predicted disease and the confidence level.

[0207] The reinforcement learning optimizer 63 is used to optimize the diagnosis and treatment pathway based on patient feedback. The optimizer models diagnosis and treatment decisions as a Markov decision process (MDP), where the states are the patient's symptoms and facial features, the actions are the diagnosis and treatment recommendations, and the rewards are the degree of improvement in the patient's condition. The optimizer uses the deep Q-network (DQN) algorithm to learn the optimal strategy. The network structure is a fully connected neural network with 3 hidden layers and 128 neurons in each layer. In order to balance exploration and utilization, the ε-greedy strategy is adopted, with an initial ε value of 0.3, which gradually decreases to 0.05 as the system usage time increases. The optimal diagnosis and treatment pathway is explored through Monte Carlo tree search, with a tree depth of 5 and 1000 simulations, which can effectively balance short-term and long-term treatment effects.

[0208] The Bayesian update unit 64 is used to update the personal sensitivity model based on the feedback data. This unit applies the Bayesian inference framework and continuously updates the posterior distribution of the model parameters by observing the data:

[0209] ,

[0210] in, are model parameters, including individual sensitivity coefficients of microcirculatory characteristics in each region and symptom expression intensity coefficients; Provide patients with feedback data, including subjective symptom improvement scores, changes in objective facial diagnosis indicators, etc. is the parameter prior distribution, which is initially set based on the statistical characteristics of the population; is the likelihood function, which represents the probability of observing feedback data under given parameters; is the parameter posterior distribution, which is continuously updated as feedback data accumulates.

[0211] The Knowledge Reasoning Feedback Module 6 also includes a progressive questioning strategy generation unit (not shown), which dynamically adjusts the depth and direction of questions based on the consultation results. This unit utilizes a decision tree structure to select questions with the highest information gain based on currently known information and the areas of abnormality diagnosed, maximizing diagnostic efficiency. The system includes 200 standard question templates covering common gynecological symptoms and signs, which are combined and transformed to generate a personalized consultation path.

[0212] Finally, the knowledge reasoning feedback module 6 generates a diagnosis and treatment report containing the following contents:

[0213] 1. Abnormal facial areas and corresponding organ scores;

[0214] 2. Abnormal indicators and analysis of periodic changes;

[0215] 3. Recommended diagnostic results (Traditional Chinese Medicine and Western Medicine) and their credibility;

[0216] 4. Personalized conditioning plan, including dietary advice, lifestyle adjustments, Chinese medicine recommendations, etc.;

[0217] 5. Follow-up plan and early warning indicators.

[0218] Example 10: Multimodal Feature Fusion Analysis Method for Gynecological Facial Diagnosis

[0219] like Figure 8 As shown, the multimodal feature fusion analysis method for gynecological facial diagnosis provided by the present invention includes the following steps:

[0220] Step S1: Collect facial multispectral image data and symptom description text data.

[0221] Specifically, it includes: using a multispectral acquisition module to obtain 590nm-610nm narrowband spectral images and 750nm-850nm near-infrared images; collecting patient symptom description text, which can be text input or voice input.

[0222] Step S2: Extracting microvascular morphological features, hemodynamic parameters and regional microcirculation index from multispectral image data.

[0223] Specifically, it includes: applying the improved Frangi filter to enhance the microvascular structure; using the U-Net convolutional neural network to perform microvascular segmentation; extracting microvascular morphological features; calculating the microcirculation index of 12 key areas; and constructing the MCI ratio matrix between regions.

[0224] Step S3: Based on the predefined facial region-gynecological organ association knowledge graph, regional analysis of microvascular morphological characteristics is performed to generate organ abnormality scores.

[0225] Specifically, the method includes: dividing the facial image into 13 regions; extracting the features of each region through a regional feature extraction network; calculating the deviation between the microcirculation features of each region and the standard model; applying the associated weight coefficient to calculate the organ abnormality score; and generating a thermal map of facial region abnormalities.

[0226] Step S4: Based on the historically collected microcirculatory characteristic time series data, a dynamic change model of the endocrine cycle is constructed to generate cycle abnormality indicators.

[0227] Specifically, it includes: collecting facial microcirculation data at four key time points of the menstrual cycle; extracting spatiotemporal features through a 3D convolutional network; using an LSTM network to model the changing patterns of features in different cycle stages; calculating the deviation between the actual change trajectory and the baseline model; and generating cycle anomaly indicators and warning information.

[0228] Step S5: Through the bidirectional cross-attention mechanism, the semantic features of the symptom text and the microcirculation abnormality features are fused to generate a fused feature vector.

[0229] Specifically, it includes: extracting key semantic information from symptom description text; calculating the initial association weight based on the symptom-region mapping matrix; realizing bidirectional enhancement of symptom attention by symptom semantic-guided regional feature extraction and regional abnormality feedback through a bidirectional cross-attention network; and using a variable-scale fusion layer for multimodal feature fusion.

[0230] Step S6: Perform multi-path reasoning based on the dynamic knowledge graph to generate diagnostic results and personalized conditioning plans, and optimize the diagnosis and treatment path based on feedback data.

[0231] Specifically, it includes: performing path search on a gynecological facial diagnosis knowledge graph containing 3,500 nodes; implementing multi-path reasoning through a graph attention network to calculate diagnostic results and credibility; optimizing diagnosis and treatment decisions based on a reinforcement learning framework; continuously optimizing individual parameters through a Bayesian update mechanism; and generating personalized diagnosis and treatment reports and conditioning recommendations.

[0232] The following are examples of the system of the present invention in practical application:

[0233] Case 1: Detection of early ovarian insufficiency

[0234] A 32-year-old woman has had decreased menstrual flow in the past six months, but all routine examinations have shown no obvious abnormalities. Using this system for face-to-face analysis, we found:

[0235] 1. The microcirculatory index in the zygomatic region of the face was significantly reduced, with an MCI value of 0.72 (normal reference range 0.85-1.15);

[0236] 2. The microvascular morphology of the zygomatic area is typically "loose", with microvascular density reduced by 25% compared to normal.

[0237] 3. The endocrine cycle dynamic model showed insufficient zygomatic microcirculatory activity during the luteal phase, with a cycle abnormality index of 0.38;

[0238] 4. A bidirectional cross-attention mechanism links the symptom of "low menstrual volume" with abnormal zygomatic microcirculation, and the fused feature vector indicates potential problems with ovarian function.

[0239] 5. The knowledge reasoning system gives the diagnosis of "early ovarian dysfunction" with a credibility of 87%.

[0240] The patient was recommended to undergo an ovarian reserve assessment, and the test results showed a critically low anti-Müllerian hormone (AMH) level, confirming the accuracy of the system's diagnosis. This case demonstrates the system's advantage in detecting early abnormalities in ovarian function, identifying potential problems 3-6 months earlier than conventional testing.

[0241] Case 2: Accurate Analysis of Cyclic Symptoms

[0242] A 28-year-old woman complained of premenstrual breast tenderness and mood swings, suspecting PMS. She used this system to monitor her menstrual cycle for three consecutive months and found the following:

[0243] 1. The facial microcirculation index is generally normal, but the microcirculation in the nose and philtrum areas shows characteristic changes during the premenstrual period, with the blood flow velocity index increasing by 35%;

[0244] 2. Regional analysis showed that the liver-related regional abnormality score increased during the premenstrual period, reaching 62 points (mild to moderate abnormality);

[0245] 3. The endocrine cycle dynamic model captured the dramatic fluctuations in microcirculatory characteristics during the premenstrual period, with a cycle coefficient of variation of 0.41, significantly higher than the normal reference range (0.1-0.25);

[0246] 4. Symptom text analysis identifies keywords such as "emotionally irritable" and "breast tenderness," and a bidirectional attention mechanism enhances the weight of facial features related to the liver area during the premenstrual period.

[0247] 5. The knowledge reasoning system comprehensively analyzes and gives the diagnosis of "liver qi stagnation type premenstrual syndrome" with a credibility of 92%.

[0248] Based on the inference results, the system generated a personalized treatment plan, including dietary recommendations (avoiding spicy foods), emotion management techniques, and Traditional Chinese Medicine (TCM) treatment recommendations. After two months of adherence, the patient experienced significant symptom improvement, with fluctuations in premenstrual facial features reduced by 60%, validating the effectiveness of the system's diagnosis and treatment plan.

[0249] The system adopts a modular design, making it easy to implement. The multispectral acquisition module can be implemented using a commercially available RGB-NIR dual-mode camera, achieving multi-band imaging through an optical filter bank. Each analysis module is developed based on a mature deep learning framework and can be deployed in a hybrid architecture of edge AI processors and cloud servers. The user terminal is a tablet computer, ensuring portability and ease of operation.

[0250] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. Gynecological facial diagnosis multimodal feature fusion analysis system, characterized by: include: Multispectral acquisition module, used to obtain facial multispectral image data and symptom description text data; a microcirculation feature extraction module, connected to the multispectral acquisition module, for extracting microvascular morphological features, hemodynamic parameters and regional microcirculation index from the multispectral image data; A regional mapping analysis module, connected to the microcirculation feature extraction module, is used to perform regional analysis on the microvascular morphological features based on a predefined facial region-gynecological organ association knowledge graph to generate an organ abnormality score; An endocrine cycle modeling module, connected to the regional mapping analysis module, is used to construct an endocrine cycle dynamic change model based on historically collected microcirculatory characteristic time series data and generate cycle abnormality indicators; a semantic attention fusion module, connected to the symptom description text data, the region mapping analysis module, and the endocrine cycle modeling module, respectively, for fusing the semantic features of the symptom text with the microcirculation abnormality features through a bidirectional cross-attention mechanism to generate a fused feature vector; The knowledge reasoning feedback module is connected to the semantic attention fusion module, and is used to receive the fused feature vector, perform multi-path reasoning based on the dynamic knowledge graph, generate diagnosis results and personalized conditioning plans, and optimize the diagnosis and treatment path according to the feedback data.

2. The system according to claim 1, wherein: The multispectral acquisition module includes: Narrowband spectroscopy unit, used to collect hemoglobin absorption characteristics in the 590nm-610nm band; Near-infrared acquisition unit, used to capture the microvascular network at a depth of 2-3mm under the skin in the 750nm-850nm band; Polarization filter unit, used to eliminate skin surface reflection interference; The spectral fusion processing unit is used to fuse images of different bands to generate enhanced microcirculation visualization images.

3. The system according to claim 1, wherein: The microcirculation feature extraction module is configured to calculate the microcirculation index of 12 key areas of the face ( ): = ( Concentration coefficient × microvascular density × blood flow velocity index) / regional baseline value, in, The concentration coefficient indicates the ratio of hemoglobin concentration to the standard value, the microvessel density indicates the number of microvessels per unit area, the blood flow velocity index indicates the relative value of blood flow velocity in the microvessels, and the regional benchmark value indicates the standard reference value of the area under healthy conditions.

4. The system according to claim 1, wherein: The region mapping analysis module includes: A region segmentation unit for segmenting the facial image into 13 regions; The regional feature extraction network consists of 13 parallel sub-networks, each of which includes 3 layers of dynamic convolutional layers and a residual connection structure; The organ association scoring unit is used to calculate the organ abnormality score based on the predefined facial region-gynecological organ association knowledge graph.

5. The system according to claim 4, characterized in that The viscera-related scoring unit is configured to calculate viscera-abnormality scores: Zang-fu abnormality score = Σ(regional deviation × ) / number of regions, Among them, the regional deviation indicates the degree of difference between the microcirculation characteristics of each region and the standard model. Indicates the association weight coefficient of each region corresponding to the internal organs.

6. The system according to claim 1, wherein: The endocrine cycle modeling module includes: Time series data acquisition unit, used to collect facial microcirculation data according to four key time points of the menstrual cycle; 3D convolutional feature extraction network, used to extract the spatiotemporal changes of microcirculatory features; Long short-term memory network, used to model the changing patterns of features at different cycle stages; The periodic anomaly detection unit is used to calculate the deviation between the actual change trajectory and the benchmark model.

7. The system according to claim 1, wherein: The semantic attention fusion module includes: Symptom semantic extraction unit, used to extract key semantic information from symptom description text; Symptom-region mapping matrix, which stores the correlation weights of 350 standard symptoms and 13 facial regions; A bidirectional cross-attention network is used to achieve bidirectional enhancement of symptom semantics to text-guided regional feature extraction and regional abnormality feedback to enhance symptom attention.

8. The system according to claim 7, characterized in that The bidirectional crisscross attention network is configured to compute: Regional attention weight = Softmax (symptom semantic vector regional feature matrix), Symptom attention = Softmax (regional anomaly vector symptom vector matrix), Among them, the regional attention weight represents the importance of each region in the analysis, and the symptom attention represents the attention weight of each symptom description.

9. The system according to claim 1, wherein: The knowledge reasoning feedback module includes: Dynamic knowledge graph, including a 3,500-node gynecological consultation knowledge network; Graph attention network, used to achieve multi-path reasoning; A reinforcement learning optimizer for optimizing diagnosis and treatment pathways based on patient feedback; The Bayesian updating unit is used to update the individual sensitivity model based on the feedback data. Compute the posterior probability of the parameters, where are model parameters, Provide patient feedback data.

10. A multimodal feature fusion analysis method for gynecological facial diagnosis, using the system according to any one of claims 1 to 9, characterized in that: include: Collect facial multispectral image data and symptom description text data; extracting microvascular morphological characteristics, hemodynamic parameters, and regional microcirculatory indexes from the multispectral image data; Based on a predefined facial region-gynecological organ association knowledge graph, the microvascular morphological characteristics are regionalized and analyzed to generate an organ abnormality score; Based on the historically collected microcirculatory characteristic time series data, a dynamic change model of the endocrine cycle is constructed to generate cycle abnormality indicators; Through the bidirectional cross-attention mechanism, the semantic features of symptom text and microcirculatory abnormality features are fused to generate a fusion feature vector; Multi-path reasoning is performed based on dynamic knowledge graphs to generate diagnostic results and personalized treatment plans, and the diagnosis and treatment path is optimized based on feedback data.

Images