A system for identifying lower extremity venous thromboembolism based on AI

Abstract

This invention discloses an AI-based system for identifying lower extremity venous thromboembolism, comprising a data acquisition module for acquiring three-dimensional point cloud sequence data of both lower extremities of a target object within a preset time period, and simultaneously acquiring surface thermal radiation distribution matrix data of the two lower extremities; and a morphological feature extraction module for inputting the three-dimensional point cloud sequence data into a pre-trained three-dimensional point cloud registration network to extract the registration point cloud of both lower extremities of the target object in the same spatial coordinate system. This invention, by constructing a multimodal analysis framework incorporating three-dimensional morphology and surface thermal radiation, abandons the traditional contact ultrasound detection mode that requires manual intervention. The system innovatively maps a two-dimensional temperature matrix to a three-dimensional registration point cloud. This scheme effectively filters out artifact interference caused by large-area skin infections, systemic fever, or changes in ambient temperature, significantly improving the specificity and exclusivity of lower extremity venous abnormality identification.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence and multimodal image processing technology, and in particular to an AI-based system for identifying lower extremity venous thromboembolism. Background Technology

[0002] Deep vein thrombosis (DVT) of the lower extremities is a common and critical complication in bedridden patients in the intensive care unit (ICU). Currently, diagnosis mainly relies on manually operated contact ultrasound. However, for extremely high-risk, critically ill, bedridden patients, traditional ultrasound cannot provide continuous 24-hour monitoring; moreover, frequent contact examinations not only increase the risk of nosocomial infection, but excessive pressure may also cause fragile, fresh thrombi to detach, leading to fatal pulmonary embolism.

[0003] Existing non-contact monitoring systems often only measure a single indicator (such as infrared thermometry or visual contour extraction), failing to detect the insidious early signs of thrombosis (i.e., the coupling between subtle edema and abnormal local temperature rise). Simple edema may be caused by heart failure or sleeping posture, while simple fever may be caused by extensive skin inflammation (such as erysipelas). Current technology cannot understand the spatial pathological logic between local volume expansion and thermal radiation from specific vascular pathways, leading to a high susceptibility to false positives and a severe lack of specificity and exclusivity. Consequently, it is difficult to meet the high-precision clinical requirement for zero-blind-zone early warning of thrombosis. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the above-mentioned technologies.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A system for identifying lower extremity venous thromboembolism based on AI includes: a data acquisition module, used to acquire three-dimensional point cloud sequence data of the bilateral lower extremities of a target object within a preset time period, and simultaneously acquire surface thermal radiation distribution matrix data of the bilateral lower extremities.

[0007] The morphological feature extraction module is used to input the three-dimensional point cloud sequence data into a pre-trained three-dimensional point cloud registration network, extract the bilateral lower limb registration point cloud of the target object in the same spatial coordinate system, and generate a morphological evolution feature vector characterizing the dynamic changes of the lower limb volume based on the asymmetric volume difference of the bilateral lower limb registration point cloud at each time node.

[0008] The multimodal mapping module is used to input the surface thermal radiation distribution matrix data into a pre-trained graph neural network, map the two-dimensional temperature values ​​corresponding to the surface thermal radiation distribution matrix data to the three-dimensional spatial nodes of the bilateral lower limb registration point cloud, and extract the temperature gradient corresponding to the spatial node along the preset deep vein topology path to obtain the local temperature gradient feature vector.

[0009] The feature cross-validation module is used to input the morphological evolution feature vector and the local temperature gradient feature vector into a validation network based on a cross-attention mechanism, calculate the spatial overlap distribution matrix of the asymmetric volume difference and the local temperature gradient feature vector within a preset time window, and output a microscopic thrombosis pathological tendency fusion vector based on the spatial overlap distribution matrix.

[0010] The early warning result generation module is used to input the microscopic thrombosis pathological tendency fusion vector into the risk assessment classification network for classification mapping, generate the lower extremity venous risk index sequence of the target object, and perform pixel rendering on the bilateral lower extremity registration point cloud according to the spatial overlap distribution matrix to output the target risk location heat map.

[0011] Preferably, the data acquisition module includes: an environmental visual stream acquisition unit, used to acquire real-time two-dimensional video stream data covering the target object bed unit area.

[0012] The occlusion state determination unit is used to input the real-time two-dimensional video stream data into a pre-trained semantic segmentation network, extract the skin pixel connected regions that represent the exposed skin areas of the target object's limbs, and calculate the area ratio of the skin pixel connected regions in the current frame of the real-time two-dimensional video stream data.

[0013] The trigger control unit is used to generate a data acquisition trigger command when the area ratio is greater than a preset exposure threshold.

[0014] A multi-source data acquisition unit is used to respond to the data acquisition trigger command, drive a binocular depth camera and an infrared thermal imager to acquire the three-dimensional point cloud sequence data of the bilateral lower limbs and the surface thermal radiation distribution matrix data of the bilateral lower limbs at the corresponding time.

[0015] Preferably, the morphological feature extraction module includes: a foreign object point cloud filtering unit, used to perform instance segmentation on the three-dimensional point cloud sequence data using a pre-trained pipeline recognition model, identify and remove interfering point clouds representing non-limb foreign objects, and obtain a clean limb point cloud.

[0016] The three-dimensional registration unit is used to perform spatial mirror symmetry registration of the left lower limb point cloud and the right lower limb point cloud in the pure limb point cloud using the pelvic feature points of the target object as the origin of the coordinate system and the iterative nearest point algorithm, so as to obtain the bilateral lower limb registration point cloud.

[0017] The volume difference calculation unit is used to divide the limb into multiple columnar slice sections of preset thickness along the longitudinal axis of the registration point cloud of the two lower limbs, calculate the volume of the left slice and the volume of the right slice corresponding to each columnar slice section, and perform a difference operation on the volume of the left slice and the volume of the right slice at the same section height, and splice them to obtain the asymmetric volume difference value.

[0018] Preferably, the preset deep vein topology path is a three-dimensional spatial coordinate sequence representing the anatomical direction of blood vessels that is pre-stored in the system. The multimodal mapping module includes: a node feature initialization unit, which is used to define each three-dimensional coordinate point in the bilateral lower limb registration point cloud as a node of the graph neural network, and to use the temperature value mapped from the surface thermal radiation distribution matrix data to the corresponding three-dimensional coordinate point as the initial attribute feature of the node.

[0019] The edge weight construction unit is used to construct the graph network edge structure between the nodes based on the spatial Euclidean distance between the three-dimensional coordinate points.

[0020] The topology path extraction unit is used to aggregate neighborhood information of the initial attribute features of the nodes through graph convolutional layers, extract the temperature variation rate of the nodes that are close to the preset deep vein topology path in three-dimensional space, and generate the local temperature gradient feature vector.

[0021] Preferably, the feature cross-validation module includes: an environmental baseline acquisition unit, used to acquire real-time room temperature data of the ward where the target object is located, and real-time physiological body temperature data representing the core body temperature of the target object.

[0022] The thermal artifact filtering unit is used to establish a temperature baseline drift model that includes the real-time room temperature data and the real-time physiological body temperature data, and to use the temperature baseline drift model to perform bias subtraction on the local temperature gradient feature vector to generate a calibrated temperature gradient vector.

[0023] The cross-attention calculation unit is used to use the morphological evolution feature vector as a query matrix and the calibration temperature gradient vector as the key matrix and value matrix, respectively, to calculate the attention score matrix of the morphological evolution feature vector sequence and the calibration temperature gradient vector sequence on the time axis, and to perform weighted summation calculation on the attention score matrix to obtain the microscopic thrombosis pathological tendency fusion vector.

[0024] Preferably, the warning result generation module includes: an object identification unit, used to acquire image data of the facial region of the target object, extract facial features from the image data using a face recognition algorithm, and compare the facial features with a preset patient database to obtain the identity code of the target object.

[0025] The risk mapping unit is used to input the microscopic thrombosis pathology tendency fusion vector into the risk assessment classification network containing multiple fully connected layers and activation functions, and output a one-dimensional probability value as the lower extremity venous risk index sequence.

[0026] The rendering output unit is used to extract spatial coordinate points in the spatial overlap distribution matrix whose values ​​are higher than a preset attention threshold, assign pixel values ​​to the spatial coordinate points on the 3D model corresponding to the bilateral lower limb registration point cloud with a preset highlight color gradient, generate the target risk location heat map, and bind the target risk location heat map with the identity identification code for output.

[0027] Preferably, the system further includes a data preprocessing module, which is located between the data acquisition module and the morphological feature extraction module. The data preprocessing module includes a spatiotemporal synchronization unit, which is used to extract the first timestamp of the three-dimensional point cloud sequence data and the second timestamp of the surface thermal radiation distribution matrix data, and to pair and slide the first timestamp and the second timestamp with a preset time alignment tolerance window to remove isolated data frames that have not been successfully paired.

[0028] The data interpolation unit is used to supplement the virtual point cloud frame and virtual temperature distribution matrix of the corresponding time node for data sequences that have been successfully paired but have missing time nodes using a cubic spline interpolation algorithm, generating an aligned point cloud sequence and an aligned temperature matrix with a uniform sampling frequency, and then inputting the aligned point cloud sequence and the aligned temperature matrix into the subsequent processing module for calculation.

[0029] Preferably, the system further includes a classification network training module for training the risk assessment classification network before inference. The classification network training module includes a sample acquisition unit for acquiring fusion vector samples of microscopic thrombotic pathology tendencies of historical patients and real lower extremity venous pathology labels corresponding to the fusion vector samples.

[0030] The loss calculation unit is used to input the microscopic thrombosis pathological tendency fusion vector samples into the initial classification network to obtain the predicted risk probability, and to calculate the network loss value using a penalized cross-entropy loss function, which is expressed as follows: ,in This refers to the numerical value of the actual lower extremity venous pathology label. The predicted risk probability, These are the network weight parameters. The preset positive regularization coefficients are used. This represents the total number of categories.

[0031] The parameter update unit is used to backpropagate the network loss value using the gradient descent algorithm to update the internal parameters of the initial classification network.

[0032] Preferably, the preset exposure threshold used in the occlusion state determination unit is a dynamically updated value. The occlusion state determination unit also includes a threshold adaptive adjustment subunit, which is specifically used to perform the following operations to update the threshold: read the skin pixel connected region area set data recorded in multiple historical data acquisition tasks of the target object in the past preset time period stored in the system.

[0033] Mathematical statistics are performed on the skin pixel connected region area set data to extract the average value feature and standard deviation feature vector of the area values ​​within the skin pixel connected region area set data.

[0034] The updated numerical baseline is calculated by adding the mean feature and the standard deviation feature vector together and multiplying by a preset confidence coefficient.

[0035] The value of the preset exposure threshold recorded in the current system memory is overwritten as the updated numerical baseline, so that when the trigger control unit receives the next frame of the real-time two-dimensional video stream data and performs area comparison, it calls the updated numerical baseline to perform condition determination.

[0036] Preferably, the early warning result generation module further includes an evidence package sealing unit, which is used to perform the following operations: extract the corresponding time period in the lower extremity vein risk index sequence that shows a continuous upward trend.

[0037] From the three-dimensional point cloud sequence data and the surface thermal radiation distribution matrix data, a set of data frames falling within the corresponding time period is extracted as key frames for risk tracing.

[0038] The risk tracing keyframe, the target risk location heatmap, the identity identification code, and the lower extremity vein risk index sequence are spliced ​​together and converted into a standardized digital array structure.

[0039] A hash algorithm is used to perform a one-way operation on the digital array structure to generate a digital digest signature. The digital digest signature and the digital array structure are then encapsulated into an immutable electronic evidence package format file, which is then written into the target address space corresponding to a preset storage server.

[0040] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0041] (1): This invention constructs a multimodal analysis framework that includes three-dimensional morphology and surface thermal radiation, thus abandoning the traditional contact ultrasound detection mode that requires manual intervention. The system innovatively maps a two-dimensional temperature matrix to a three-dimensional registration point cloud and extracts the temperature gradient along the pre-defined deep vein topology path unique to anatomy. It also uses a cross-attention mechanism to verify the pathological resonance of asymmetric swelling and vascular heating in physical space and time. This approach effectively filters out artifact interference caused by large-area skin infection, systemic fever, or changes in ambient temperature, significantly improving the specificity and exclusion of lower extremity venous abnormality identification, and achieving non-invasive, non-contact, and continuous accurate risk monitoring for critically ill bedridden patients.

[0042] (2): The early warning result generation module of this invention creatively introduces object recognition technology based on facial feature extraction. When outputting high-dimensional artificial intelligence risk mapping probability and location heatmap, the system automatically binds the object to the target identity code in the hospital database through object recognition. This mechanism ensures that complex physical monitoring indicators can be accurately attributed to specific patient individuals, realizing closed-loop management from feature calculation to identity verification, and greatly reducing the workload of nurses' manual verification and the potential mismatch risk.

[0043] (3): This invention introduces an intelligent triggering mechanism based on video stream semantic segmentation. The system only wakes up the multi-source hardware for high-precision acquisition when the skin exposure area reaches the threshold, which greatly reduces the computing power loss and hardware lifespan consumption caused by long-term ineffective operation. Furthermore, by learning the exposure history area of ​​patients with different body types through statistical algorithms, the dynamic adaptive update of the judgment threshold is realized, avoiding missed triggers or frequent false triggers caused by the target body type being thin or obese, and enhancing the environmental robustness of the system. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a system module architecture diagram proposed in this invention;

[0046] Figure 2 This is a flowchart of the system operation proposed in this invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] The following description, in conjunction with the accompanying drawings, describes an AI-based system for identifying lower extremity venous thromboembolism according to an embodiment of the present invention.

[0049] like Figure 1 and Figure 2 As shown, the present invention provides an AI-based system for identifying lower extremity venous thromboembolism, comprising the following modules:

[0050] The data acquisition module is used to acquire three-dimensional point cloud sequence data of the target object's two lower limbs within a preset time period, and simultaneously acquire surface thermal radiation distribution matrix data of the two lower limbs.

[0051] In practical applications, this step serves as the system's data input source. Since traditional ultrasound cannot be used for non-contact monitoring, the system acquires basic data from physical space using hardware sensors (such as binocular depth cameras and infrared thermal imagers). The three-dimensional point cloud sequence data is a set of three-dimensional spatial coordinates calculated from the infrared speckle emitted by the depth camera and the reflected data, reflecting the geometric morphology of the lower limbs. The surface thermal radiation distribution matrix data is a two-dimensional temperature array converted from the infrared radiation energy captured by the infrared thermal imaging sensor. These two heterogeneous data sets are recorded synchronously within the same preset time period, providing raw material for subsequent multimodal feature fusion. This enables non-contact, continuous data acquisition of the lower limb morphology and temperature of bedridden individuals.

[0052] For example, a depth camera and an infrared thermal imager continuously scan the lower limbs of a bedridden target to generate a three-dimensional point cloud sequence of data on the surface of the target's legs (including X, Y, and Z coordinate information), and at the same time generate the corresponding surface thermal radiation distribution matrix data (including the Celsius value of each point).

[0053] The system continuously captures a sequence of color images, including the bed and its surrounding environment, using a standard RGB camera in the ward; this is the real-time two-dimensional video stream data. Clothing occlusion can cause the infrared and depth sensors to malfunction. The system inputs the real-time two-dimensional video stream data into a semantic segmentation model (such as U-Net), categorizing the images into skin, clothing, background, etc., and extracting pixel blocks belonging to the patient's lower limb skin, i.e., the "skin pixel connected region." This is expressed by the formula: area percentage. ,in For the number of skin pixels, This represents the total number of pixels in the current frame. Only when the area percentage meets the set value (i.e., the preset exposure threshold) is the lower limb considered to be in an unobstructed, observable state, at which point the data acquisition trigger command is issued. This avoids invalid calculations when clothing obstructs the view, saving computing resources. Upon receiving the data acquisition trigger command, the system hardware awakens from sleep mode and accurately captures the 3D point cloud sequence data and the surface thermal radiation distribution matrix data within the exposure window.

[0054] For example, a regular camera on the hospital ward ceiling captures real-time 2D video stream data including the target patient's bed. When a nurse lifts the blanket to provide care, a semantic segmentation network identifies the exposed skin on the target patient's legs and calculates that the connected components of these skin pixels occupy 15% of the total pixels in the current video frame. The system's preset exposure threshold is 10%. Since the target patient's current skin area is greater than 10%, the control unit generates a data acquisition trigger command. Upon receiving the trigger command, the depth camera and infrared thermal imager immediately activate, capturing the point cloud and temperature matrix data of the target patient's exposed lower limbs at the current moment.

[0055] Different patients have varying baseline areas of exposed skin due to differences in body type. The system queries the database to extract the set of skin pixel connected region areas recorded in the past few days under normal, unshaded conditions for the current patient, as the basis for personalized adjustments. For the previously extracted historical area set, its arithmetic mean (the mean feature) is calculated to determine the center point of the patient's normal exposed area, and the standard deviation (the standard deviation feature vector) is calculated to assess the data fluctuation and dispersion of the patient's daily exposed area. The statistical formula is applied: Baseline = (Mean Feature Vector - ... The standard deviation feature vector is used here in the form of a subtraction variant or an addition multiplication coefficient to obtain a reliable lower boundary value (the updated numerical baseline), where the confidence coefficient is used to control the tolerance. The system replaces the previously fixed absolute area threshold (the preset exposure threshold) in memory with the personalized updated numerical baseline calculated in the previous step. This avoids the acquisition module being falsely triggered or missed due to differences in body size between obese or thin patients, greatly improving the robustness of the system in real and complex medical environments.

[0056] For example, the system retrieved and recorded 10 instances of large-area skin pixel connected region area data belonging to the target object during morning sponge baths over the past 3 days. The system calculated the average feature of the target object's exposed area over the past few days to be 20%, with a standard deviation feature vector of 2% due to fluctuations caused by different postures. The system combined the target object's average feature of 20% with the standard deviation, multiplied it by a confidence coefficient, and obtained a personalized updated baseline value of 16% for the target object. The default exposure threshold of 10% in the system memory was overwritten to the target object's specific 16%. Subsequently, when the trigger control unit processes the target object's real-time 2D video stream data, it will use 16% as a threshold to determine whether to initiate deep acquisition.

[0057] This invention provides a system for identifying lower extremity venous thromboembolism based on AI. The system further includes a data preprocessing module, which is located between the data acquisition module and the morphological feature extraction module.

[0058] In practical applications, depth cameras and infrared cameras have different hardware refresh rates, so the system reads their hardware timestamps (the first timestamp and the second timestamp) separately. A very short time difference range is set (a preset time alignment tolerance window, e.g., 10 milliseconds). Data within this range is considered a valid combination at the same time, and mismatched single-modal data is discarded. On the timeline, the system uses a mathematical smoothing function (cubic spline interpolation algorithm) to estimate and generate virtual point cloud frames and virtual temperature distribution matrices to fill gaps based on known preceding and following data frames. This ensures that the output to the feature extraction module is a continuous, smooth, and uniformly sampled aligned point cloud sequence and aligned temperature matrix. This prevents feature extraction temporal networks (such as LSTM) from reporting errors due to inconsistent input lengths.

[0059] For example, a depth camera at the head of the target object's bed captures 30 frames per second, while an infrared camera captures 20 frames per second. The spatiotemporal synchronization unit pairs the data from both cameras within the same millisecond range based on timestamps, discarding any extra point cloud frames that fail to match. Because the target object rolls over, resulting in a 2-second data gap, the data interpolation unit uses morphological and temperature data before and after the rollover to smoothly extrapolate virtual features within these 2 seconds, filling in the gap and generating a stable, aligned point cloud sequence.

[0060] The morphological feature extraction module is used to input the three-dimensional point cloud sequence data into a pre-trained three-dimensional point cloud registration network, extract the bilateral lower limb registration point cloud of the target object in the same spatial coordinate system, and generate a morphological evolution feature vector characterizing the dynamic changes of the lower limb volume based on the asymmetric volume difference of the bilateral lower limb registration point cloud at each time node.

[0061] In practical applications, this step transforms the raw spatial point cloud into quantitative features that characterize abnormal limb swelling. Following the previous step, the extracted 3D point cloud sequence data exhibits coordinate system discrepancies between the left and right legs due to differences in shooting angle and patient position. A 3D point cloud registration network (such as a variant of PointNet) transforms the feature points of the left and right legs into the same mathematical coordinate system by finding the correspondence, forming the bilateral lower limb registration point cloud. Subsequently, the system calculates the difference between the volumes of the left and right legs, i.e., the asymmetric volume difference. This difference, arranged in a time series, constitutes the morphological evolution feature vector. This overcomes the morphological measurement errors caused by changes in patient posture and accurately quantifies the physical phenomenon of unilateral limb abnormal swelling.

[0062] For example, the system spatially aligns the point clouds of the target object's left and right legs, and calculates that at the current time point, the volume of the target object's left lower leg is 150 cm larger than the volume of its right lower leg. 3 The asymmetric volume difference is used to encode the sequence of this difference over the past few hours into a morphological evolution feature vector.

[0063] In the ICU environment, patients' legs often have IV lines, monitoring leads, etc. The geometric protrusions of these items severely interfere with volume calculations. The pipeline identification model extracts these non-human structure points from the 3D point cloud sequence data, outputting a pure limb point cloud containing only the limb surface. This significantly improves the accuracy and robustness of subsequent volume difference calculations. Because patients' bedridden postures are arbitrary, the spatial positions of the left and right legs are asymmetrical. The system finds the pelvis as a reference anchor point in the pure limb point cloud, uses the Iterative Closest Point (ICP) algorithm to calculate the rigid body transformation matrix (rotation and translation), mirrors one leg and overlaps it in the coordinate space of the other leg, forming the bilateral lower limb registration point cloud. After obtaining the bilateral lower limb registration point cloud, the system uses a calculus-like approach to divide it into multiple layers (e.g., each layer is 1 cm thick) along the leg length direction. The formula is used... Calculate the volume of each layer, and then determine the volume at the same height. All heights This constitutes the asymmetric volume difference vector. It can accurately pinpoint a specific segment of the calf or thigh where occult swelling has occurred.

[0064] For example, the model identifies a urinary catheter draped over a target object's leg, deletes its corresponding point cloud coordinates, and retains only the pure limb point cloud belonging to the skin surface of the target object's leg. The target object originally had a bent left leg and a straight right leg. Using the pelvis as the origin, the system uses an algorithm to mirror and overlap the point cloud of the bent left leg with that of the right leg in digital space. The system slices the aligned point cloud of the target object every 1 cm and finds that at a height of 15 cm from the ankle, the volume of the left slice is 5 cm larger than that of the right slice.3 The differences between each layer are summed to obtain the overall asymmetric volume difference.

[0065] The multimodal mapping module is used to input the surface thermal radiation distribution matrix data into a pre-trained graph neural network, map the two-dimensional temperature values ​​corresponding to the surface thermal radiation distribution matrix data to the three-dimensional spatial nodes of the bilateral lower limb registration point cloud, and extract the temperature gradient corresponding to the spatial node along the preset deep vein topology path to obtain the local temperature gradient feature vector.

[0066] In practical applications, a single temperature map is insufficient to pinpoint specific lesions. This step uses the aforementioned two-dimensional surface thermal radiation distribution matrix data as attribute labels, assigning them one by one to the three-dimensional coordinate points in the bilateral lower limb registration point cloud, achieving multimodal fusion of space and temperature. Graph Neural Networks (GNNs) utilize the spatial connections between nodes for information transmission. The preset deep vein topology path is a built-in three-dimensional vascular model conforming to human anatomy. The system extracts the rate of temperature change (i.e., temperature gradient) along this path, generating the local temperature gradient feature vector. This eliminates interference from large-area environmental heating or systemic heating, accurately locating the local abnormal temperature rise zone along the specific deep vein vascular bed.

[0067] For example, the system attaches the temperature distribution of the target object's leg to its 3D leg model, and extracts the values ​​of abnormally high temperatures at each coordinate point along the path of the femoral vein in medical anatomy, forming a local temperature gradient feature vector.

[0068] The system will use each coordinate point in the previously acquired bilateral lower limb registration point cloud. Consider it as a node in the graph structure. Then, look up the table to match the previously mentioned surface thermal radiation distribution matrix data and obtain the temperature value at the corresponding location. The value is assigned to this node as its initial attribute characteristic. In the physical world, heat is conducted continuously; the system calculates this through nodes. and nodes spatial Euclidean distance formula The system establishes connections (edges) between nearby nodes, constructing a graph network edge structure between them. The closer the nodes, the greater the edge weight. The graph convolutional layer smooths the temperature values ​​of each node based on the graph network edge structure, reducing sensor noise. Subsequently, the system focuses on nodes that are close to the preset deep vein topology path (the prior three-dimensional vascular skeleton), calculates the steepness of their temperature rise relative to surrounding nodes (temperature variability), and finally aggregates them into the local temperature gradient feature vector.

[0069] For example, a 3D point (10, 20, 5) on the target object's lower leg model is defined as a node. The system extracts the corresponding pixel value of 37.5℃ from the temperature matrix as the initial attribute feature of this node. The system calculates that the distance between two adjacent nodes on the target object's leg model is 2 mm, determines that they are spatially close, and assigns them high weights to the connected edges to construct the graph network edge structure. The graph convolutional network smooths the temperature noise on the target object's leg and specifically extracts the rate of temperature increase of nodes closely following the femoral vein, synthesizing them into a local temperature gradient feature vector.

[0070] The feature cross-validation module is used to input the morphological evolution feature vector and the local temperature gradient feature vector into a validation network based on a cross-attention mechanism, calculate the spatial overlap distribution matrix of the asymmetric volume difference and the local temperature gradient feature vector within a preset time window, and output a microscopic thrombosis pathological tendency fusion vector based on the spatial overlap distribution matrix.

[0071] In practical applications, the system receives the morphological evolution feature vector and the local temperature gradient feature vector from the preceding steps, and performs an inner product operation on them using a cross-attention mechanism. When the region of asymmetric volume expansion happens to coincide with the region of abnormal temperature rise on a specific vascular path, the attention score is significantly amplified, thereby generating the spatial overlap distribution matrix. The high-dimensional features extracted based on this matrix are the microscopic thrombosis pathological tendency fusion vector. This significantly improves the exclusivity of risk identification and avoids misjudgments caused by simple edema (without local high fever) or simple skin inflammation (large-area fever but not along the vascular pathway).

[0072] For example, cross-attention network calculations revealed that the area of ​​asymmetric swelling in the left calf of the target object highly overlapped with the area of ​​abnormal temperature rise along the deep vein path of the calf in three-dimensional space. Based on this, the system outputs a fusion vector of microscopic thrombotic pathology tendency with high weight.

[0073] Because the absolute temperature measured by infrared thermal imagers is easily affected by room temperature fluctuations or patient fever, the system obtains real-time room temperature data by connecting to an IoT thermometer and hygrometer in the ward via an external interface, and obtains the patient's real-time physiological body temperature data (such as axillary or oral temperature) by connecting to a bedside monitor. The system uses formulas... (in The real-time room temperature data, The real-time physiological body temperature data, A dynamic reference temperature is calculated using adaptive weighting coefficients and subtracted from the previously extracted local temperature gradient feature vector, thereby eliminating artifacts caused by environmental or systemic heat and outputting the calibrated temperature gradient vector reflecting the true local heat source. This significantly improves the stability of temperature difference features in complex clinical environments. The system uses the morphological evolution feature vector (swelling feature) as... The matrix uses the calibration temperature gradient vector (heating characteristic) as... and Matrix, using the standard attention formula The calculation is performed. This process essentially involves finding the resonance points between morphological and temperature anomalies in the timeline of physical evolution. Finally, a weighted fusion vector containing resonance information is obtained, representing the microscopic thrombotic pathological tendency.

[0074] For example, the system reads the current real-time room temperature data of 24℃ from the ICU ward environmental sensor, and the real-time physiological body temperature data of 38℃ (feverish state) from the target subject's temperature monitoring probe. Considering that the target subject is currently feverish and is being directly exposed to air conditioning, the system uses a temperature baseline drift model to subtract the impact of overall fever, and extracts the heat deviation purely caused by local blood flow obstruction from the local temperature gradient to generate a calibrated temperature gradient vector. The system calculation found that the morphological evolution characteristics of the target subject's left leg (gradually thickening) are highly consistent with the trend of change of the calibrated temperature gradient vector (local continuous abnormal temperature rise) over the past 4 hours. The attention score matrix calculates a high score, and outputs a microscopic thrombotic pathological tendency fusion vector containing the weight of this high score.

[0075] The early warning result generation module is used to input the microscopic thrombosis pathological tendency fusion vector into the risk assessment classification network for classification mapping, generate the lower extremity venous risk index sequence of the target object, and perform pixel rendering on the bilateral lower extremity registration point cloud according to the spatial overlap distribution matrix to output the target risk location heat map.

[0076] In practical applications, the microscopic thrombotic pathological tendency fusion vector is input into a classification network composed of fully connected layers, mapping it to a numerical sequence between 0 and 1, namely the lower extremity venous risk index sequence, representing the severity of the objective physical characteristics deviating from the normal baseline. Simultaneously, the system extracts high-score regions from the previously described spatial overlap distribution matrix and assigns color values ​​(e.g., red represents high overlap) to corresponding points in the bilateral lower extremity registration point cloud, forming the target risk localization heatmap. This transforms complex AI high-dimensional features into intuitive, quantifiable data indicators and visualized images, facilitating the reading and monitoring of the status by non-professionals.

[0077] For example, the system maps a lower extremity venous risk index of 0.82 based on the fusion vector, and renders the abnormal area along the blood vessel direction on the 3D model of the target object's left calf in red, generating an intuitive target risk location heat map for monitoring personnel to view.

[0078] To enable the risk assessment network to make judgments, the system needs to read a large set of labeled offline data, including fusion vector samples of the microscopic thrombotic pathology tendency input features, and final results of 0 or 1 from professional methods (such as ultrasound diagnosis) as the actual lower extremity venous pathology labels. The system calculates the predicted risk probability using the input features through a current, immature model. Compare the artificially labeled actual lower extremity venous pathology tags. The prediction bias is calculated using the following formula: The latter half of the formula includes an L2 regularization penalty term. This is used to limit the excessive expansion of network weights. The resulting value is defined as the network loss value. This effectively prevents the model from overfitting on historical small samples. Based on the previously obtained network loss value, the system calculates its partial derivative matrix with respect to the weight parameters of each layer of the network, and then uses the gradient descent algorithm to inversely adjust the neuron connection weights within the network. After multiple iterations, the model acquires the ability to accurately map to unknown data.

[0079] For example, researchers extracted microscopic thrombotic pathology tendency fusion vector samples from the database of 1000 past ICU patients and attached actual lower extremity venous pathology labels (0 for no risk, 1 for high risk) indicating whether they actually had deep vein abnormalities at the time. The initial classification network predicted a risk probability of 0.3 for historical samples, but the true label was 1 (high risk). The loss calculation unit used the above formula to calculate the larger network loss value caused by the current bias. The parameter update unit used the gradient descent algorithm to backpropagate the calculated loss value and fine-tune the parameter weights of the fully connected layers in the network, so that when the network faces a similar high-risk feature like the target object again, the output probability can be closer to 1.

[0080] Image acquisition data is obtained using a standard 2D camera. Facial vectors (facial features) are extracted using a convolutional neural network and matched against a hospital database to obtain a desensitized unique patient ID, i.e., the identification code. The high-dimensional data output from the preceding step, specifically the microscopic thrombotic pathology tendency fusion vector, is difficult for humans to understand directly. Therefore, it is compressed and reduced in dimensionality using a fully connected layer with a nonlinear activation function, outputting a continuous scalar numerical range (e.g., 0-100%) to form a time-varying sequence of lower limb venous risk indices. The system selects a set of coordinates with extremely high scores from the spatial overlap distribution matrix of weight 1. In the graphics rendering pipeline of the bilateral lower limb registration point cloud, these coordinate systems are covered with a highlight color from yellow to red, i.e., the target risk localization heatmap. Finally, the preceding identification code is added for structured storage.

[0081] For example, the system captures the facial image of the target object, extracts facial features, compares them with the hospital's medical record database, and obtains the target object's hospital number "ID-8866" as its identification code. After inputting the target object's fused vector into the risk assessment classification network, the network's terminal activation function outputs a scalar probability of 0.85, which is used as the lower extremity venous risk index at the current moment. The system colors the coordinates of the target object's leg overlap scores exceeding 0.8 with dark red to generate a target risk localization heatmap, adds "ID-8866" to the image's caption, and sends it to the doctor's workstation.

[0082] The system iterates through the previously generated lower limb vein risk index sequence, uses a differential algorithm to find intervals where the values ​​show a monotonically increasing pattern, and records the start and end times of this deterioration process, i.e., the corresponding time period. Based on the time window of the previous positioning, the system backtracks to the underlying database and extracts the most original morphological and temperature input data (i.e., the 3D point cloud sequence data and the surface thermal radiation distribution matrix data) during the period of risk anomaly, naming it the risk tracing keyframe. This can retain the core objective data with the most medical backtracking value while reducing storage overhead. To adapt to subsequent encryption and storage requirements, the system concatenates the multimodal data, the previously generated rendered heatmap, the identification code, and the risk values ​​together, serializing and encoding them into a one-dimensional machine-readable string (such as JSON or a binary sequence), i.e., the standardized digital array structure. The system uses hash algorithms such as SHA-256 to map the standardized digital array structure to a fixed-length unique random string (the digital digest signature). Any slight tampering with the original data will cause the digest signature to change drastically. Ultimately, these two parts are assembled and compressed into the aforementioned electronic evidence package format file and stored on a disk. This endows the pure algorithm evaluation results with a high level of tamper-proof evidence-gathering attributes, protecting the legitimate data rights of both medical staff and patients.

[0083] For example, the system detected that the target object's lower limb vein risk index sequence continuously increased from 0.3 to 0.85 between 2 PM and 4 PM yesterday. The system extracted this two-hour period as the corresponding time series. From the underlying database, the system extracted the original point cloud and thermal radiation matrix frames of the target object's legs recorded between 2 PM and 4 PM yesterday, using them as key frames for risk tracing. The system then combined the target object's risk tracing key frames, the red-highlighted location heatmap, the ID-8866 identifier, and the risk value of 0.85, converting them into a large, standardized binary code array structure. The system used a hash algorithm to generate a unique scrambled digital digest signature for the target object's data array, packaged it along with the data array into a read-only electronic evidence package file, and permanently archived it in the hospital's central storage server.

[0084] The present invention has the following advantages:

[0085] (1): This invention constructs a multimodal analysis framework that includes three-dimensional morphology and surface thermal radiation, thus abandoning the traditional contact ultrasound detection mode that requires manual intervention. The system innovatively maps a two-dimensional temperature matrix to a three-dimensional registration point cloud and extracts the temperature gradient along the pre-defined deep vein topology path unique to anatomy. It also uses a cross-attention mechanism to verify the pathological resonance of asymmetric swelling and vascular heating in physical space and time. This approach effectively filters out artifact interference caused by large-area skin infection, systemic fever, or changes in ambient temperature, significantly improving the specificity and exclusion of lower extremity venous abnormality identification, and achieving non-invasive, non-contact, and continuous accurate risk monitoring for critically ill bedridden patients.

[0086] (2): The early warning result generation module of this invention creatively introduces object recognition technology based on facial feature extraction. When outputting high-dimensional artificial intelligence risk mapping probability and location heatmap, the system automatically binds the object to the target identity code in the hospital database through object recognition. This mechanism ensures that complex physical monitoring indicators can be accurately attributed to specific patient individuals, realizing closed-loop management from feature calculation to identity verification, and greatly reducing the workload of nurses' manual verification and the potential mismatch risk.

[0087] (3): This invention introduces an intelligent triggering mechanism based on video stream semantic segmentation. The system only wakes up the multi-source hardware for high-precision acquisition when the skin exposure area reaches the threshold, which greatly reduces the computing power loss and hardware lifespan consumption caused by long-term ineffective operation. Furthermore, by learning the exposure history area of ​​patients with different body types through statistical algorithms, the dynamic adaptive update of the judgment threshold is realized, avoiding missed triggers or frequent false triggers caused by the target body type being thin or obese, and enhancing the environmental robustness of the system.

[0088] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0089] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0090] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A system for identifying lower extremity venous thromboembolism based on AI, characterized in that, include: The data acquisition module is used to acquire the three-dimensional point cloud sequence data of the target object's two lower limbs within a preset time period, and simultaneously acquire the surface thermal radiation distribution matrix data of the two lower limbs. The morphological feature extraction module is used to input the three-dimensional point cloud sequence data into a pre-trained three-dimensional point cloud registration network, extract the bilateral lower limb registration point cloud of the target object in the same spatial coordinate system, and generate a morphological evolution feature vector characterizing the dynamic change of the lower limb volume based on the asymmetric volume difference of the bilateral lower limb registration point cloud at each time node. The multimodal mapping module is used to input the surface thermal radiation distribution matrix data into a pre-trained graph neural network, map the two-dimensional temperature values ​​corresponding to the surface thermal radiation distribution matrix data to the three-dimensional spatial nodes of the bilateral lower limb registration point cloud, and extract the temperature gradient corresponding to the spatial node along the preset deep vein topology path to obtain the local temperature gradient feature vector. The feature cross-validation module is used to input the morphological evolution feature vector and the local temperature gradient feature vector into a validation network based on a cross-attention mechanism, calculate the spatial overlap distribution matrix of the asymmetric volume difference and the local temperature gradient feature vector within a preset time window, and output a microscopic thrombosis pathological tendency fusion vector based on the spatial overlap distribution matrix. The early warning result generation module is used to input the microscopic thrombosis pathological tendency fusion vector into the risk assessment classification network for classification mapping, generate the lower extremity venous risk index sequence of the target object, and perform pixel rendering on the bilateral lower extremity registration point cloud according to the spatial overlap distribution matrix to output the target risk location heat map.

2. The system according to claim 1, characterized in that, The data acquisition module includes: An environmental visual stream acquisition unit is used to acquire real-time two-dimensional video stream data covering the target object bed unit area; The occlusion state determination unit is used to input the real-time two-dimensional video stream data into a pre-trained semantic segmentation network, extract the skin pixel connected regions that represent the exposed skin areas of the target object's limbs, and calculate the area ratio of the skin pixel connected regions in the current frame of the real-time two-dimensional video stream data. The trigger control unit is used to generate a data acquisition trigger command when the area ratio is greater than a preset exposure threshold; A multi-source data acquisition unit is used to respond to the data acquisition trigger command, drive a binocular depth camera and an infrared thermal imager to acquire the three-dimensional point cloud sequence data of the bilateral lower limbs and the surface thermal radiation distribution matrix data of the bilateral lower limbs at the corresponding time.

3. The system according to claim 1, characterized in that, The morphological feature extraction module includes: The foreign object point cloud filtering unit is used to perform instance segmentation on the three-dimensional point cloud sequence data using a pre-trained pipeline recognition model, identify and remove interfering point clouds that represent non-limb foreign objects, and obtain a clean limb point cloud. The three-dimensional registration unit is used to perform spatial mirror symmetry registration of the left lower limb point cloud and the right lower limb point cloud in the pure limb point cloud using the pelvic feature points of the target object as the origin of the coordinate system and the iterative nearest point algorithm, so as to obtain the bilateral lower limb registration point cloud. The volume difference calculation unit is used to divide the limb into multiple columnar slice sections of preset thickness along the longitudinal axis of the registration point cloud of the two lower limbs, calculate the volume of the left slice and the volume of the right slice corresponding to each columnar slice section, and perform a difference operation on the volume of the left slice and the volume of the right slice at the same section height, and splice them to obtain the asymmetric volume difference value.

4. The system according to claim 1, characterized in that, The preset deep vein topology path is a three-dimensional spatial coordinate sequence representing the anatomical orientation of blood vessels pre-stored within the system. The multimodal mapping module includes: The node feature initialization unit is used to define each three-dimensional coordinate point in the bilateral lower limb registration point cloud as a node of the graph neural network, and to use the temperature value mapped from the surface thermal radiation distribution matrix data to the corresponding three-dimensional coordinate point as the initial attribute feature of the node. An edge weight construction unit is used to construct the graph network edge structure between the nodes based on the spatial Euclidean distance between the three-dimensional coordinate points. The topology path extraction unit is used to aggregate neighborhood information of the initial attribute features of the nodes through graph convolutional layers, extract the temperature variation rate of the nodes that are close to the preset deep vein topology path in three-dimensional space, and generate the local temperature gradient feature vector.

5. The system according to claim 1, characterized in that, The feature cross-validation module includes: The environmental baseline acquisition unit is used to acquire real-time room temperature data of the ward where the target object is located, as well as real-time physiological body temperature data representing the core body temperature of the target object. The thermal artifact filtering unit is used to establish a temperature baseline drift model that includes the real-time room temperature data and the real-time physiological body temperature data, and to use the temperature baseline drift model to perform bias subtraction on the local temperature gradient feature vector to generate a calibration temperature gradient vector. The cross-attention calculation unit is used to use the morphological evolution feature vector as a query matrix and the calibration temperature gradient vector as the key matrix and value matrix, respectively, to calculate the attention score matrix of the morphological evolution feature vector sequence and the calibration temperature gradient vector sequence on the time axis, and to perform weighted summation calculation on the attention score matrix to obtain the microscopic thrombosis pathological tendency fusion vector.

6. The system according to claim 1, characterized in that, The early warning result generation module includes: The object identification unit is used to acquire image data of the facial region of the target object, extract facial features from the image data using a face recognition algorithm, and compare the facial features with a preset patient database to obtain the identity code of the target object. The risk mapping unit is used to input the microscopic thrombosis pathology tendency fusion vector into the risk assessment classification network containing multiple fully connected layers and activation functions, and output a one-dimensional probability value as the lower extremity venous risk index sequence. The rendering output unit is used to extract spatial coordinate points in the spatial overlap distribution matrix whose values ​​are higher than a preset attention threshold, assign pixel values ​​to the spatial coordinate points on the 3D model corresponding to the bilateral lower limb registration point cloud with a preset highlight color gradient, generate the target risk location heat map, and bind the target risk location heat map with the identity identification code for output.

7. The system according to claim 1, characterized in that, The system further includes a data preprocessing module, disposed between the data acquisition module and the morphological feature extraction module, the data preprocessing module comprising: The spatiotemporal synchronization unit is used to extract the first timestamp of the three-dimensional point cloud sequence data and the second timestamp of the surface thermal radiation distribution matrix data, and to pair and slide the first timestamp and the second timestamp with a preset time alignment tolerance window to remove isolated data frames that have not been successfully paired. The data interpolation unit is used to supplement the virtual point cloud frame and virtual temperature distribution matrix of the corresponding time node for data sequences that have been successfully paired but have missing time nodes using a cubic spline interpolation algorithm, generating an aligned point cloud sequence and an aligned temperature matrix with a uniform sampling frequency, and then inputting the aligned point cloud sequence and the aligned temperature matrix into the subsequent processing module for calculation.

8. The system according to claim 1, characterized in that, The system further includes a classification network training module for training the risk assessment classification network before inference. The classification network training module includes: The sample acquisition unit is used to acquire microscopic thrombotic pathology tendency fusion vector samples of historical patients and real lower extremity venous pathology labels corresponding to the fusion vector samples. The loss calculation unit is used to input the microscopic thrombosis pathological tendency fusion vector samples into the initial classification network to obtain the predicted risk probability, and to calculate the network loss value using a penalized cross-entropy loss function, which is expressed as follows: ,in This refers to the numerical value of the actual lower extremity venous pathology label. The predicted risk probability, These are the network weight parameters. The preset positive regularization coefficients are used. The total number of categories; The parameter update unit is used to backpropagate the network loss value using the gradient descent algorithm to update the internal parameters of the initial classification network.

9. The system according to claim 2, characterized in that, The preset exposure threshold used in the occlusion state determination unit is a dynamically updated value. The occlusion state determination unit also includes a threshold adaptive adjustment subunit, which is specifically used to perform the following operations to update the threshold: Read the skin pixel connected region area set data recorded in multiple historical data acquisition tasks of the target object within a preset time period stored in the system; Mathematical statistics are performed on the skin pixel connected region area set data to extract the average value feature and standard deviation feature vector of the area values ​​within the skin pixel connected region area set data; The updated numerical baseline is calculated by adding the mean feature and the standard deviation feature vector together and multiplying by a preset confidence coefficient. The value of the preset exposure threshold recorded in the current system memory is overwritten as the updated numerical baseline, so that when the trigger control unit receives the next frame of the real-time two-dimensional video stream data and performs area comparison, it calls the updated numerical baseline to perform condition determination.

10. The system according to claim 6, characterized in that, The early warning result generation module also includes an evidence package sealing unit, which is used to perform the following operations: Extract the corresponding time periods in the lower extremity venous risk index sequence that show a continuous upward trend; From the three-dimensional point cloud sequence data and the surface thermal radiation distribution matrix data, a set of data frames falling within the corresponding time period is extracted as key frames for risk tracing. The risk tracing keyframe, the target risk location heatmap, the identity identification code, and the lower extremity vein risk index sequence are spliced ​​together and converted into a standardized digital array structure. A hash algorithm is used to perform a one-way operation on the digital array structure to generate a digital digest signature. The digital digest signature and the digital array structure are then encapsulated into an immutable electronic evidence package format file, which is then written into the target address space corresponding to a preset storage server.

Images