Early precise diagnosis method and system for deep vein thrombosis based on multi-modal data fusion

Abstract

The present application relates to the technical field of medical artificial intelligence, and is especially a deep vein thrombosis early precise diagnosis system based on multi-modal data fusion, comprising a multi-modal data acquisition module, a multi-source heterogeneous data preprocessing module, a multi-modal data intelligent fusion module, a DVT precise diagnosis and abnormal positioning module, and an early warning feedback module. The present application has the advantages of comprehensive and precise data acquisition, fine and effective data preprocessing, intelligent and flexible data fusion and feature modeling, precise and comprehensive early diagnosis and positioning, and timely and continuous early warning feedback and dynamic iteration optimization. In actual use, in combination with high-frequency ultrasound (7-15 MHz), thin-layer CT (0.625 mm), and MR contrast, not only morphological data are collected, but also hemodynamics and three-dimensional reconstruction models are included, providing a solid anatomical basis for diagnosis. Wearable devices are used for high-frequency continuous monitoring to capture blood flow and vascular compliance changes under static and dynamic pressure.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, specifically to a method and system for early and accurate diagnosis of deep vein thrombosis based on multimodal data fusion. Background Technology

[0002] Medical artificial intelligence is a deep integration of artificial intelligence technology and the medical field. It is not only an application of technology, but also a disruptive revolution in the medical paradigm. Based on massive medical data, it simulates, extends and expands the cognitive abilities of human experts to build an intelligent medical system that ranges from basic genes and intermediate diseases to advanced diagnosis and treatment.

[0003] Current technologies struggle to comprehensively integrate multi-source data from imaging, wearable physiological data, laboratory data, and clinical texts, resulting in a lack of diagnostic evidence for deep vein thrombosis and an inability to form a complete and accurate information support system. During imaging data acquisition, the lack of scientific parameter settings for probes and equipment affects image clarity and accuracy; unreasonable probe layout and sampling frequency planning for wearable devices makes it difficult to guarantee the quality of physiological data; non-standard laboratory data testing procedures and poor correlation between multi-source data make it difficult to construct a time-consistent dataset, impacting overall data quality; and preprocessing methods lack specificity for different data types. Imaging data denoising algorithms are simplistic, vascular region segmentation models are outdated, and rigid registration errors are significant; artifact removal from wearable physiological data is ineffective, blood flow temporal features are insufficiently extracted, and temperature and hemodynamic data processing methods are unscientific; algorithms for missing value completion and erroneous data removal in laboratory data are inaccurate, affecting data integrity; and there are shortcomings in preprocessing optimization, with imaging data preprocessing failing to deeply improve the model, such as the lack of an attention mechanism to optimize U-axis data. The Net model, with its rigid registration and lack of anatomical landmark constraints, suffers from inadequate preprocessing of wearable physiological data. The wavelet basis and decomposition layer are not clearly defined, resulting in unclear details in blood flow temporal feature extraction, impacting preprocessing effectiveness. Reliance on traditional methods for multimodal data fusion hinders independent feature extraction and effective noise isolation, making it difficult to accurately extract clinical semantic features from electronic medical records and limiting in-depth data value mining. Feature extraction methods are limited, failing to comprehensively utilize multiple models such as CNN, LSTM, and multivariate logistic regression to extract image structure, physiological temporal features, and laboratory quantitative features separately, leading to incomplete feature information. Furthermore, the multimodal data fusion weight allocation mechanism lacks flexibility, employing a fixed weight model that fails to consider individual patient differences, such as high-risk groups. The use of imaging features with dynamically adjusted weights reduces the specificity and accuracy of diagnosis. Early thrombosis diagnosis methods have limitations: low overlap of suspected thrombosis area markers; insufficient sensitivity and specificity of deep classification models; weak ability to identify early microthrombosis; difficulty in achieving accurate early diagnosis; an imperfect dynamic risk grading assessment system; lack of effective risk prediction models; unreasonable patient risk level classification; poor consistency of risk assessment; lack of an immediate risk trigger mechanism; inability to promptly initiate risk reassessment when real-time physiological signals are abnormal, delaying diagnosis; unscientific abnormal location methods; unreasonable monitoring area division; inaccurate multi-source data fusion algorithms; large location errors; low consistency with the gold standard; and output accuracy of location results that fails to meet clinical needs. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for early and accurate diagnosis of deep vein thrombosis based on multimodal data fusion, which has the advantages of high-precision imaging data, accurate image processing, ultra-high sensitivity and accurate positioning, and solves the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a deep vein thrombosis early precision diagnosis system based on multimodal data fusion, comprising: a multimodal data acquisition module, a multi-source heterogeneous data preprocessing module, a multimodal data intelligent fusion module, a DVT precision diagnosis and abnormal location module, and an early warning feedback module.

[0006] A method for early and accurate diagnosis of deep vein thrombosis based on multimodal data fusion is characterized by the following steps: S1: Precise Multimodal Data Acquisition: Image Data Acquisition: Using a high-frequency linear probe (7-15MHz), scan the common femoral vein, superficial femoral vein, popliteal vein, and posterior tibial vein of the lower extremities, acquiring two-dimensional grayscale images at a frame rate ≥30fps and color Doppler blood flow images at a velocity range of 0-10cm / s. Simultaneously, record the vascular lumen compression rate data during the compression test; CT Venography: Slice thickness set to 0.625mm, acquire venous phase enhanced images, and reconstruct a three-dimensional vascular model; MR Venography: Using contrast enhancement, acquire vascular morphology and hemodynamic data, highlighting the signal intensity of suspicious vascular segments.

[0007] Wearable physiological data acquisition: Blood flow and temperature monitoring: PPG blood flow probes and thermistor temperature probes are fixed at the upper 1 / 3 of the thigh, the middle 1 / 3 of the calf, and the ankle of the lower limbs. The sampling frequency is set to 100Hz to collect photoplethysmography and local skin temperature under static conditions, with an accuracy of ±0.1℃; Hemodynamic monitoring: The smart compression stockings have built-in miniature pressure sensors and perform dynamic compression tests every 2 hours, with a pressure gradient of 15-30mmHg, to collect venous return rate and vascular compliance data; Data admission: Physiological signals with a motion artifact ratio exceeding 10% are removed, and abnormal temperature data are screened using the 3σ principle.

[0008] Laboratory data acquisition: Coagulation function indicators: D-dimer was detected by collecting venous blood, with a cutoff value set at 0.5 mg / L FEU, fibrinogen, and prothrombin time; Genetic risk indicators: Factor and prothrombin gene mutations were detected by microarray chip, and an individualized genetic risk score was output; Data association: All laboratory data and imaging and physiological data were bound to the patient's unique ID and collection timestamp to construct a multi-source dataset with a consistent timeline.

[0009] S2: Multi-source heterogeneous data standardization preprocessing: Image data preprocessing: Ultrasound images are processed using adaptive median filtering to remove speckle noise, and CT / MR images are processed using non-local mean filtering to suppress Gaussian noise; the U-Net model is used to automatically segment the venous vessel region, and images from different time points and different modalities are aligned to the same anatomical coordinate system through a rigid registration algorithm, with a registration error ≤1mm.

[0010] Wearable physiological data preprocessing: Adaptive wavelet transform was used to remove motion artifacts and extract 12 blood flow temporal features such as interpeak duration and diastolic attenuation slope of PPG signal; temperature data was mapped to the 0-1 interval using Min-Max normalization and hemodynamic data were standardized using Z-score to eliminate individual differences.

[0011] Laboratory data preprocessing: K-nearest neighbor algorithm was used to supplement missing coagulation indicators, and multiple imputation method was used to complete genetic mutation data; indicators that exceeded the medically reasonable range were combined with clinical medical records to determine whether they were pathological abnormalities, and erroneous data caused by laboratory operation errors were removed.

[0012] S3: Intelligent Fusion and Feature Modeling of Multimodal Data: Clinical Text Features: Meta-learning and differential topological constraints are introduced to separate image noise, physiological signal artifacts, and laboratory data biases, breaking through the dependence of traditional fusion on explicit signal alignment. This enables independent feature extraction and noise isolation of different modal data. A large language model is used to extract clinical semantic features of age, surgical history, and pregnancy status from electronic medical records. Image Structural Features: Deep visual features of vascular texture and thrombus boundaries are extracted through a CNN model. Physiological Temporal Features: A LSTM model is used to extract the periodic changes of blood flow signals and the temporal dynamic features of temperature fluctuations. Laboratory Quantitative Features: A multi-factor logistic regression model is constructed to transform coagulation indicators and genetic risk scores into quantitative risk features.

[0013] By using a sparse activation gating module, four types of features are mapped to a unified high-dimensional feature space. Based on the diagnostic weights of different data (40% for imaging, 25% for physiology, 20% for laboratory features, and 15% for clinical features), a weighted fusion is performed to generate a fusion vector containing 512-dimensional features.

[0014] S4: Precise Diagnosis and Abnormal Localization of Early DVT: Combining fusion features with clinical gold standards, achieving high-sensitivity identification and precise localization of early thrombi: Early thrombus identification: Using a region growing algorithm, suspected thrombus areas with abnormal gray levels and blood flow filling defects in blood vessels are marked from fusion image features. The overlap between the suspected areas and the gold standard is ≥85%. The fusion features are input into a deep classification model, which outputs the DVT positive probability. With contrast-enhanced ultrasound as the gold standard, the model sensitivity is ≥96% and the specificity is ≥93%. For early microthrombi with a diameter <5mm within 72 hours of onset, the recognition rate is ≥90%.

[0015] Dynamic risk grading assessment: Time series model construction: LSTM risk prediction model is trained using multimodal time series features over 30 consecutive days as input; Risk level output: Patients are divided into 4 risk levels: low risk probability <10%, medium risk probability 10%-30%, high risk probability 30%-70%, and very high risk probability >70%, with a Kappa coefficient of consistency ≥0.85 for risk assessment.

[0016] Precise location of abnormalities: The lower limb is divided longitudinally into five blood flow monitoring zones: upper thigh, lower thigh, upper calf, lower calf, and ankle; and laterally into three temperature monitoring zones: medial, lateral, and posterior. Combining anatomical location from imaging, abnormal areas of physiological signals, and local inflammatory characteristics from laboratory data, a weighted voting algorithm is used to determine the location of the thrombus. The location error is ≤2cm, and the consistency with the gold standard CTV is ≥92%.

[0017] S5: Early Warning Feedback and Dynamic Iterative Optimization: Establish a closed-loop diagnosis-monitoring-update mechanism to achieve individualized and precise management: Tiered early warning push: Push corresponding intervention plans according to risk level: low risk pushes home activity guidance and compression stocking wearing suggestions; medium risk pushes weekly follow-up reminders; high risk pushes immediate ultrasound follow-up warning; extremely high risk triggers emergency medical rescue connection; set monitoring frequency according to patient risk level: low risk collects data once every 7 days, high risk collects physiological data once a day, and performs ultrasound follow-up every 3 days; incorporate multimodal data of 1000 new patients every quarter, use transfer learning to fine-tune the fusion model parameters to ensure the model's adaptability to newly diagnosed DVT subtypes and populations, and improve the model's diagnostic accuracy by ≥2% annually.

[0018] Furthermore, as a preferred embodiment of the present invention, in step S2, the U-Net model for image data preprocessing is an improved U-Net with an added attention mechanism, which strengthens the edge feature weights of veins through the channel attention module; the rigid registration algorithm adopts a multimodal registration method based on mutual information, and introduces vascular anatomical landmarks as constraint points during the registration process, further controlling the registration error to ≤0.8mm.

[0019] Furthermore, as a preferred embodiment of the present invention, in step S2, the adaptive wavelet transform for wearable physiological data preprocessing uses a 5th-order Daubechies wavelet basis, with a decomposition layer of 4 layers, and extracts 12 blood flow temporal features of the PPG signal, including: interpeak duration, diastolic attenuation slope, peak height coefficient of variation, blood perfusion index, pulse rate variability, rise time, fall time, dicrotic wave depth, diastolic area ratio, systolic peak velocity, end-diastolic velocity, and mean velocity.

[0020] Furthermore, as a preferred embodiment of the present invention, in step S3, the weight allocation of the multimodal data intelligent fusion is a dynamically adjustable mechanism: when the patient is a high-risk group such as late pregnancy or within 72 hours after surgery, the weight of physiological data is adjusted to 30% and the weight of clinical features is adjusted to 20%; when there are suspected thrombosis grayscale features in the image data, the weight of the image data is temporarily increased to 50%.

[0021] Furthermore, as a preferred embodiment of the present invention, in step S4, the region growth algorithm for early thrombosis identification is an adaptive threshold region growth algorithm: using pixels whose gray value in the blood vessel is 20% lower than the average value of the surrounding normal blood vessels as initial seed points, and combining the blood flow filling threshold during the growth process, dynamically adjusting the growth stopping condition when the blood flow filling of the region to be grown is <60%, and stopping the growth when the blood flow filling of the region to be grown is ≥60%.

[0022] Furthermore, as a preferred embodiment of the present invention, in step S4, the dynamic risk grading assessment also includes an immediate risk triggering mechanism: when the blood flow rate of the real-time acquired PPG signal drops by ≥30% and the local skin temperature rises by ≥1℃, the risk reassessment process is immediately triggered without waiting for 30 days of complete time-series data.

[0023] Furthermore, as a preferred embodiment of the present invention, in step S4, the weighted voting algorithm for abnormal location localization is set as follows: 50% weight for imaging anatomical localization features, based on the thrombus coordinates of the CTV vascular 3D model; 30% weight for wearable physiological signal abnormality areas, based on monitoring areas with sudden drops in blood flow velocity and abnormal temperatures; and 20% weight for laboratory local inflammation features, supplemented by the local increase trend of D-dimer and the level of inflammatory factors. The localization result uses three-dimensional coordinates, with lower limb longitudinal distance, lateral orientation, and depth output, and the coordinate accuracy is ≤1.5cm.

[0024] Beneficial effects: The technical solution of this application has the following advantages: comprehensive and accurate data acquisition, refined and effective data preprocessing, intelligent and flexible data fusion and feature modeling, accurate and comprehensive early diagnosis and localization, and timely and continuous early warning feedback and dynamic iterative optimization. In actual use, it integrates imaging data, wearable physiological data, laboratory data, and clinical text data to comprehensively acquire various information related to deep vein thrombosis, providing rich evidence for accurate diagnosis; imaging data acquisition uses probes and equipment with specific parameters to ensure image quality; wearable devices are rationally arranged with probes and set with appropriate sampling frequencies to ensure the accuracy of physiological data; laboratory data is rigorously tested and correlated with multi-source data to construct a dataset with a consistent timeline.

[0025] Appropriate methods are adopted for different types of data. Different filtering algorithms are used to denoise image data, and an improved U-Net model is used to segment vascular regions. The rigid registration algorithm has small errors. Adaptive wavelet transform is used to remove artifacts and extract rich blood flow temporal features for wearable physiological data. Temperature and hemodynamic data are normalized and standardized, respectively. Different algorithms are used to supplement missing values ​​and remove erroneous data for laboratory data. For example, an attention mechanism is added to the U-Net model in image data preprocessing, and anatomical landmark constraints are introduced into the rigid registration algorithm. The preprocessing of wearable physiological data specifies the wavelet basis and the number of decomposition layers, and lists the extracted blood flow temporal features in detail.

[0026] By introducing meta-learning and differential topological constraints, we break through traditional dependencies and achieve independent feature extraction and noise isolation for data from different modalities. We can also extract clinical semantic features from electronic medical records. We extract image structure, physiological time series and laboratory quantitative features through models such as CNN, LSTM and multifactor logistic regression. The weight allocation of multimodal data fusion is a dynamically adjustable mechanism. The weights are temporarily adjusted according to different patient conditions, such as high-risk groups and suspected image features, to improve the diagnostic specificity.

[0027] An adaptive threshold region growth algorithm is used to mark suspected thrombus areas, with high overlap. The deep classification model has high sensitivity and specificity, and a high recognition rate for early microthrombi. An LSTM risk prediction model is constructed to classify patients into four risk levels, with high consistency in risk assessment. An instant risk triggering mechanism is set up to immediately trigger risk reassessment when real-time physiological signals are abnormal. The monitoring area is divided vertically and horizontally, and a weighted voting algorithm is used to determine the thrombus location by combining multi-source data. The location error is small, and the consistency with the gold standard is high. The location results are output in three-dimensional coordinates, with high accuracy.

[0028] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below can be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other. Attached Figure Description

[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. To better understand the technical content of the present invention, specific embodiments are provided and described in conjunction with the accompanying drawings. Various aspects of the present invention are described in this disclosure with reference to the accompanying drawings, which illustrate numerous illustrative embodiments. It should be understood that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of many ways. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0031] As attached Figure 1 As shown: This embodiment provides an early and accurate diagnosis system for deep vein thrombosis based on multimodal data fusion, including: a multimodal data acquisition module, a multi-source heterogeneous data preprocessing module, a multimodal data intelligent fusion module, a DVT accurate diagnosis and abnormal location module, and an early warning feedback module.

[0032] A method for early and accurate diagnosis of deep vein thrombosis based on multimodal data fusion is characterized by the following steps: S1: Precise Multimodal Data Acquisition: Image Data Acquisition: Using a high-frequency linear probe (7-15MHz), scan the common femoral vein, superficial femoral vein, popliteal vein, and posterior tibial vein of the lower extremities, acquiring two-dimensional grayscale images at a frame rate ≥30fps and color Doppler blood flow images at a velocity range of 0-10cm / s. Simultaneously, record the vascular lumen compression rate data during the compression test; CT Venography: Slice thickness set to 0.625mm, acquire venous phase enhanced images, and reconstruct a three-dimensional vascular model; MR Venography: Using contrast enhancement, acquire vascular morphology and hemodynamic data, highlighting the signal intensity of suspicious vascular segments.

[0033] Wearable physiological data acquisition: Blood flow and temperature monitoring: PPG blood flow probes and thermistor temperature probes are fixed at the upper 1 / 3 of the thigh, the middle 1 / 3 of the calf, and the ankle of the lower limbs. The sampling frequency is set to 100Hz to collect photoplethysmography and local skin temperature under static conditions, with an accuracy of ±0.1℃; Hemodynamic monitoring: The smart compression stockings have built-in miniature pressure sensors and perform dynamic compression tests every 2 hours, with a pressure gradient of 15-30mmHg, to collect venous return rate and vascular compliance data; Data admission: Physiological signals with a motion artifact ratio exceeding 10% are removed, and abnormal temperature data are screened using the 3σ principle.

[0034] Laboratory data acquisition: Coagulation function indicators: D-dimer was detected by collecting venous blood, with a cutoff value set at 0.5 mg / L FEU, fibrinogen, and prothrombin time; Genetic risk indicators: Factor and prothrombin gene mutations were detected by microarray chip, and an individualized genetic risk score was output; Data association: All laboratory data and imaging and physiological data were bound to the patient's unique ID and collection timestamp to construct a multi-source dataset with a consistent timeline.

[0035] S2: Multi-source heterogeneous data standardization preprocessing: Image data preprocessing: Ultrasound images are processed using adaptive median filtering to remove speckle noise, and CT / MR images are processed using non-local mean filtering to suppress Gaussian noise; the U-Net model is used to automatically segment the venous vessel region, and images from different time points and different modalities are aligned to the same anatomical coordinate system through a rigid registration algorithm, with a registration error ≤1mm.

[0036] Wearable physiological data preprocessing: Adaptive wavelet transform was used to remove motion artifacts and extract 12 blood flow temporal features such as interpeak duration and diastolic attenuation slope of PPG signal; temperature data was mapped to the 0-1 interval using Min-Max normalization and hemodynamic data were standardized using Z-score to eliminate individual differences.

[0037] Laboratory data preprocessing: K-nearest neighbor algorithm was used to supplement missing coagulation indicators, and multiple imputation method was used to complete genetic mutation data; indicators that exceeded the medically reasonable range were combined with clinical medical records to determine whether they were pathological abnormalities, and erroneous data caused by laboratory operation errors were removed.

[0038] S3: Intelligent Fusion and Feature Modeling of Multimodal Data: Clinical Text Features: Meta-learning and differential topological constraints are introduced to separate image noise, physiological signal artifacts, and laboratory data biases, breaking through the dependence of traditional fusion on explicit signal alignment. This enables independent feature extraction and noise isolation of different modal data. A large language model is used to extract clinical semantic features of age, surgical history, and pregnancy status from electronic medical records. Image Structural Features: Deep visual features of vascular texture and thrombus boundaries are extracted through a CNN model. Physiological Temporal Features: A LSTM model is used to extract the periodic changes of blood flow signals and the temporal dynamic features of temperature fluctuations. Laboratory Quantitative Features: A multi-factor logistic regression model is constructed to transform coagulation indicators and genetic risk scores into quantitative risk features.

[0039] By using a sparse activation gating module, four types of features are mapped to a unified high-dimensional feature space. Based on the diagnostic weights of different data (40% for imaging, 25% for physiology, 20% for laboratory features, and 15% for clinical features), a weighted fusion is performed to generate a fusion vector containing 512-dimensional features.

[0040] S4: Precise Diagnosis and Abnormal Localization of Early DVT: Combining fusion features with clinical gold standards, achieving high-sensitivity identification and precise localization of early thrombi: Early thrombus identification: Using a region growing algorithm, suspected thrombus areas with abnormal gray levels and blood flow filling defects in blood vessels are marked from fusion image features. The overlap between the suspected areas and the gold standard is ≥85%. The fusion features are input into a deep classification model, which outputs the DVT positive probability. With contrast-enhanced ultrasound as the gold standard, the model sensitivity is ≥96% and the specificity is ≥93%. For early microthrombi with a diameter <5mm within 72 hours of onset, the recognition rate is ≥90%.

[0041] Dynamic risk grading assessment: Time series model construction: LSTM risk prediction model is trained using multimodal time series features over 30 consecutive days as input; Risk level output: Patients are divided into 4 risk levels: low risk probability <10%, medium risk probability 10%-30%, high risk probability 30%-70%, and very high risk probability >70%, with a Kappa coefficient of consistency ≥0.85 for risk assessment.

[0042] Precise location of abnormalities: The lower limb is divided longitudinally into five blood flow monitoring zones: upper thigh, lower thigh, upper calf, lower calf, and ankle; and laterally into three temperature monitoring zones: medial, lateral, and posterior. Combining anatomical location from imaging, abnormal areas of physiological signals, and local inflammatory characteristics from laboratory data, a weighted voting algorithm is used to determine the location of the thrombus. The location error is ≤2cm, and the consistency with the gold standard CTV is ≥92%.

[0043] S5: Early Warning Feedback and Dynamic Iterative Optimization: Establish a closed-loop diagnosis-monitoring-update mechanism to achieve individualized and precise management: Tiered early warning push: Push corresponding intervention plans according to risk level: low risk pushes home activity guidance and compression stocking wearing suggestions; medium risk pushes weekly follow-up reminders; high risk pushes immediate ultrasound follow-up warning; extremely high risk triggers emergency medical rescue connection; set monitoring frequency according to patient risk level: low risk collects data once every 7 days, high risk collects physiological data once a day, and performs ultrasound follow-up every 3 days; incorporate multimodal data of 1000 new patients every quarter, use transfer learning to fine-tune the fusion model parameters to ensure the model's adaptability to newly diagnosed DVT subtypes and populations, and improve the model's diagnostic accuracy by ≥2% annually.

[0044] Specifically, in step S2, the U-Net model for image data preprocessing is an improved U-Net with an added attention mechanism, which strengthens the edge feature weights of veins through the channel attention module; the rigid registration algorithm adopts a multimodal registration method based on mutual information, and introduces vascular anatomical landmarks as constraint points during the registration process, further controlling the registration error to ≤0.8mm.

[0045] Specifically, in step S2, the adaptive wavelet transform for wearable physiological data preprocessing uses a 5th-order Daubechies wavelet basis, with a decomposition layer of 4 layers, and extracts 12 blood flow temporal features of the PPG signal, including: interpeak duration, diastolic attenuation slope, peak height coefficient of variation, blood perfusion index, pulse rate variability, rise time, fall time, dicrotic wave depth, diastolic area ratio, systolic peak velocity, end-diastolic velocity, and mean velocity.

[0046] Specifically, in step S3, the weight allocation of multimodal data intelligent fusion is a dynamically adjustable mechanism: when the patient is a high-risk group such as late pregnancy or within 72 hours after surgery, the weight of physiological data is adjusted to 30% and the weight of clinical features is adjusted to 20%; when there are suspected thrombosis grayscale features in the image data, the weight of the image data is temporarily increased to 50%.

[0047] Specifically, in step S4, the region growth algorithm for early thrombosis identification is an adaptive threshold region growth algorithm: the pixel point whose gray value in the blood vessel is 20% lower than the average value of the surrounding normal blood vessels is used as the initial seed point. During the growth process, the growth stopping condition is dynamically adjusted by combining the blood flow filling threshold. When the blood flow filling is <60%, the growth stops.

[0048] Specifically, in step S4, the dynamic risk grading assessment also includes an immediate risk triggering mechanism: when the blood flow rate of the real-time collected PPG signal drops by ≥30% and the local skin temperature rises by ≥1℃, the risk reassessment process is immediately triggered without waiting for 30 days of complete time-series data.

[0049] Specifically, in step S4, the weights of the weighted voting algorithm for abnormal location localization are set as follows: 50% weight for imaging anatomical localization features, based on the thrombus coordinates of the CTV vascular 3D model; 30% weight for wearable physiological signal abnormality areas, based on monitoring areas with sudden drops in blood flow rate and abnormal temperature; and 20% weight for laboratory local inflammation features, supplemented by the local increase trend of D-dimer and the level of inflammatory factors. The localization result uses three-dimensional coordinates, with the lower limb longitudinal distance, lateral orientation, and depth output, and the coordinate accuracy is ≤1.5cm.

[0050] It should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0051] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A precise early diagnosis system for deep vein thrombosis based on multimodal data fusion, characterized in that, include: Multimodal data acquisition module, multi-source heterogeneous data preprocessing module, multimodal data intelligent fusion module, DVT precise diagnosis and anomaly localization module, and early warning feedback module; A method for early and accurate diagnosis of deep vein thrombosis based on multimodal data fusion is characterized by the following steps: S1: Precise Multimodal Data Acquisition: Image Data Acquisition: Using a high-frequency linear probe (7-15MHz), scan the common femoral vein, superficial femoral vein, popliteal vein, and posterior tibial vein of the lower extremities, acquiring two-dimensional grayscale images at a frame rate ≥30fps and color Doppler blood flow images at a velocity range of 0-10cm / s. Simultaneously, record the vascular lumen compression rate data during the compression test; CT Venography: Slice thickness set to 0.625mm, acquire venous phase enhanced images, and reconstruct a three-dimensional vascular model; MR Venography: Using contrast enhancement, acquire vascular morphology and hemodynamic data, highlighting the signal intensity of suspicious vascular segments; Wearable physiological data acquisition: Blood flow and temperature monitoring: PPG blood flow probes and thermistor temperature probes are fixed at the upper 1 / 3 of the thigh, the middle 1 / 3 of the calf, and the ankle of the lower limbs. The sampling frequency is set to 100Hz to collect photoplethysmography and local skin temperature under static conditions, with an accuracy of ±0.1℃; Hemodynamic monitoring: The smart compression stockings have built-in miniature pressure sensors and perform dynamic compression tests every 2 hours, with a pressure gradient of 15-30mmHg, to collect venous return rate and vascular compliance data; Data admission: Physiological signals with motion artifacts exceeding 10% are removed, and abnormal temperature data are screened using the 3σ principle; Laboratory data acquisition: Coagulation function indicators: D-dimer was detected by collecting venous blood, with a cutoff value set at 0.5 mg / LFEU, fibrinogen, and prothrombin time; Genetic risk indicators: Factor and prothrombin gene mutations were detected by microarray chip, and an individualized genetic risk score was output; Data association: All laboratory data and imaging and physiological data were bound to the patient's unique ID and collection timestamp to construct a multi-source dataset with a consistent timeline; S2: Multi-source heterogeneous data standardization preprocessing: Image data preprocessing: Ultrasound images are processed using adaptive median filtering to remove speckle noise, and CT / MR images are processed using non-local mean filtering to suppress Gaussian noise; the U-Net model is used to automatically segment the venous vessel region, and images from different time points and modalities are aligned to the same anatomical coordinate system using a rigid registration algorithm, with a registration error ≤1mm; Wearable physiological data preprocessing: Adaptive wavelet transform was used to remove motion artifacts and extract 12 blood flow temporal features of PPG signals, such as interpeak duration and diastolic attenuation slope; temperature data was normalized to the 0-1 interval using Min-Max normalization and hemodynamic data was normalized using Z-score to eliminate individual differences. Laboratory data preprocessing: K-nearest neighbor algorithm was used to supplement missing coagulation indicators, and multiple imputation method was used to complete genetic mutation data; indicators that exceeded the medically reasonable range were combined with clinical medical records to determine whether they were pathological abnormalities, and erroneous data caused by laboratory operation errors were removed. S3: Intelligent Fusion and Feature Modeling of Multimodal Data: Clinical Text Features: Introducing meta-learning and differential topological constraints to separate image noise, physiological signal artifacts, and laboratory data bias, breaking through the dependence of traditional fusion on explicit signal alignment, achieving independent feature extraction and noise isolation of different modal data, and using a large language model to extract clinical semantic features of age, surgical history, and pregnancy status from electronic medical records; Image Structural Features: Extracting deep visual features of vascular texture and thrombus boundaries through a CNN model; Physiological Temporal Features: Using an LSTM model to extract the periodic changes of blood flow signals and the temporal dynamic features of temperature fluctuations; Laboratory Quantitative Features: Constructing a multi-factor logistic regression model to transform coagulation indicators and genetic risk scores into quantitative risk features; By using a sparse activation gating module, four types of features are mapped to a unified high-dimensional feature space. Based on the diagnostic weights of different data: 40% for imaging, 25% for physiological features, 20% for laboratory features, and 15% for clinical features, a weighted fusion is performed to generate a fusion vector containing 512-dimensional features. S4: Precise Diagnosis and Abnormal Localization of Early DVT: Combining fusion features with clinical gold standards to achieve highly sensitive identification and precise localization of early thrombi: Early thrombus identification: Using a region growing algorithm, suspected thrombus areas with abnormal gray levels and blood flow filling defects within the vessel are marked from the fusion image features. The overlap between the suspected areas and the gold standard is ≥85%. The fusion features are input into a deep classification model, which outputs the DVT positive probability. Using contrast-enhanced ultrasound as the gold standard, the model has a sensitivity ≥96% and a specificity ≥93%. For early microthrombi with a diameter <5mm within 72 hours of onset, the recognition rate is ≥90%. Dynamic risk grading assessment: Time series model construction: Using multimodal time series features over 30 consecutive days as input, an LSTM risk prediction model is trained; Risk level output: Patients are divided into 4 risk levels: low risk probability <10%, medium risk probability 10%-30%, high risk probability 30%-70%, and very high risk probability >70%, with a Kappa coefficient of consistency ≥0.85 for risk assessment; Precise location of abnormalities: The lower limb is longitudinally divided into five blood flow monitoring zones: upper thigh, lower thigh, upper calf, lower calf, and ankle; and laterally divided into three temperature monitoring zones: medial, lateral, and posterior. Combining anatomical localization from imaging, abnormal areas of physiological signals, and local inflammatory characteristics from laboratory data, a weighted voting algorithm is used to determine the location of the thrombus. The location error is ≤2cm, and the consistency with the gold standard CTV is ≥92%. S5: Early Warning Feedback and Dynamic Iterative Optimization: Establish a closed-loop diagnosis-monitoring-update mechanism to achieve individualized and precise management: Tiered early warning push: Push corresponding intervention plans according to risk level: low risk pushes home activity guidance and compression stocking wearing suggestions; medium risk pushes weekly follow-up reminders; high risk pushes immediate ultrasound follow-up warning; extremely high risk triggers emergency medical rescue connection; set monitoring frequency according to patient risk level: low risk collects data once every 7 days, high risk collects physiological data once a day, and performs ultrasound follow-up every 3 days; incorporate multimodal data of 1000 new patients every quarter, use transfer learning to fine-tune the fusion model parameters to ensure the model's adaptability to newly diagnosed DVT subtypes and populations, and improve the model's diagnostic accuracy by ≥2% annually.

2. The method for early and accurate diagnosis of deep vein thrombosis based on multimodal data fusion according to claim 1, characterized in that: In step S2, the U-Net model for image data preprocessing is an improved U-Net with an added attention mechanism, which strengthens the edge feature weights of veins through the channel attention module; The rigid registration algorithm adopts a multimodal registration method based on mutual information. During the registration process, vascular anatomical landmarks are introduced as constraint points to further control the registration error to ≤0.8mm.

3. The method for early and accurate diagnosis of deep vein thrombosis based on multimodal data fusion according to claim 1, characterized in that: In step S2, the adaptive wavelet transform for wearable physiological data preprocessing uses a 5th-order Daubechies wavelet basis with 4 decomposition layers to extract 12 blood flow temporal features of the PPG signal, including: interpeak duration, diastolic attenuation slope, peak height coefficient of variation, perfusion index, pulse rate variability, rise time, fall time, dicrotic wave depth, diastolic area ratio, systolic peak velocity, end-diastolic velocity, and mean velocity.

4. The method for early and accurate diagnosis of deep vein thrombosis based on multimodal data fusion according to claim 1, characterized in that: In step S3, the weight allocation of multimodal data intelligent fusion is a dynamically adjustable mechanism: when the patient is a high-risk group such as late pregnancy or within 72 hours after surgery, the weight of physiological data is adjusted to 30% and the weight of clinical features is adjusted to 20%; when there are suspected thrombosis grayscale features in the image data, the weight of the image data is temporarily increased to 50%.

5. The method for early and accurate diagnosis of deep vein thrombosis based on multimodal data fusion according to claim 1, characterized in that: In step S4, the region growth algorithm for early thrombosis identification is an adaptive threshold region growth algorithm: the pixel point whose gray value in the blood vessel is 20% lower than the average value of the surrounding normal blood vessels is used as the initial seed point. During the growth process, the growth stopping condition is dynamically adjusted in combination with the blood flow filling threshold. When the blood flow filling is <60%, the growth stops.

6. The method for early and accurate diagnosis of deep vein thrombosis based on multimodal data fusion according to claim 1, characterized in that: In step S4, the dynamic risk grading assessment also includes an immediate risk triggering mechanism: when the blood flow rate of the real-time collected PPG signal drops by ≥30% and the local skin temperature rises by ≥1℃, the risk reassessment process is immediately triggered without waiting for 30 days of complete time-series data.

7. The method for early and accurate diagnosis of deep vein thrombosis based on multimodal data fusion according to claim 1, characterized in that: In step S4, the weights of the weighted voting algorithm for abnormal location localization are set as follows: 50% weight for imaging anatomical localization features, based on the thrombus coordinates of the CTV vascular 3D model; 30% weight for wearable physiological signal abnormality areas, based on the monitoring areas of sudden drop in blood flow rate and abnormal temperature; and 20% weight for laboratory local inflammation features, supplemented by the local increase trend of D-dimer and the level of inflammatory factors. The localization result uses three-dimensional coordinates, with the lower limb longitudinal distance, lateral orientation, and depth output, and the coordinate accuracy is ≤1.5cm.

Images