Detection method for deep venous thrombosis of patient with traumatic craniocerebral injury

Abstract

The invention discloses a method for detecting deep venous thrombosis of a patient with traumatic craniocerebral injury, and relates to the technical field of deep venous thrombosis detection. The method comprises the following steps: acquiring deep vein images and physiological parameters, performing noise reduction, normalization and abnormal value elimination, extracting vein structures and suspected thrombus features, marking gray scale and morphological anomaly regions, and forming a feature vector library; analyzing a physiological parameter time sequence trend, dividing high, medium and low risks, and associating image analysis; weighting and fusing the image and the physiological parameters, judging a thrombus area and marking the type, the position and the size; the thrombus change is manually rechecked and dynamically monitored, and clinical intervention is associated to form a closed loop; and generating and filing an encrypted report, and synchronizing the encrypted report to an electronic medical record system. The thrombus detection comprehensiveness and accuracy are improved through multi-data fusion, dynamic monitoring and risk pre-judgment are achieved, the system adapts to clinical scenes, efficient support is provided for thrombus management of traumatic craniocerebral injury patients, and intervention opportunities and schemes are optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep vein thrombosis detection, and particularly relates to a method for detecting deep vein thrombosis of a traumatic brain injury patient. BACKGROUND

[0002] Traumatic brain injury patients have limited limb movement due to central nervous system damage, long-term bed rest, and abnormal coagulation function caused by trauma stress. The incidence of deep vein thrombosis is significantly higher than that of ordinary patients. If not detected and intervened in time, thrombus shedding may cause pulmonary embolism, which seriously threatens the safety of patients. The current clinical deep vein thrombosis detection methods have obvious limitations: CT venography (CTV) can clearly show the structure of blood vessels, but it has ionizing radiation and is not suitable for traumatic brain injury patients who need multiple reviews; ultrasound examination depends on the experience of the operator, and the recognition accuracy of small thrombi or deep pelvic thrombi is low, which is easy to cause missed judgment and misjudgment due to differences in doctors' subjective judgment, especially at night or in primary hospitals, when there is a shortage of experienced ultrasound doctors, the detection reliability further decreases. In addition, the existing methods focus on single detection, lack of long-term dynamic monitoring design for patients, and cannot capture the subtle changes in the early stage of thrombus formation in time, resulting in missed opportunity for some patients to intervene.

[0003] The existing detection technology lacks data fusion capability, and it is difficult to integrate multi-dimensional information for comprehensive research and judgment. Most methods only rely on single type data, such as judging the stenosis of blood vessel lumen only through image data, or only screening according to physiological parameters such as D-dimer concentration, ignoring the importance of hemodynamic parameters (such as blood flow velocity, direction) and clinical symptoms (such as lower limb swelling, pain). For example, when early thrombus appears in the deep vein, the blood flow velocity will change before the lumen stenosis, if only relying on image detection, it needs to wait until the thrombus volume grows to a certain extent to be identified, which delays early intervention; and simply relying on D-dimer concentration, false positives may occur due to non-specific increase caused by trauma stress, increasing the risk of unnecessary anticoagulant therapy. This single data dimension detection mode cannot form a complete thrombus feature image, resulting in low accuracy of early thrombus detection.

[0004] Meanwhile, the prior art lacks effective means for quantitative assessment and risk prediction of thrombus, and the ability of clinical decision support is insufficient. Traditional methods are difficult to accurately calculate the volume of thrombus, and can only assess the severity of thrombus through qualitative description of "size" and "location", which cannot provide quantitative basis for anticoagulant drug dosage adjustment; the judgment of thrombus stability is also based on subjective observation, which cannot scientifically evaluate the risk of thrombus detachment and early warning of high-risk events. In addition, the risk prediction model relies on limited static parameters, does not combine real-time physiological changes and long-term follow-up data of patients, and the prediction result has low precision, and the model parameters cannot be continuously optimized according to clinical feedback, resulting in insufficient adaptability and iteration ability of the detection method, which is difficult to meet the individualized and dynamic detection needs of patients with traumatic brain injury. SUMMARY

[0005] The present application provides a method for detecting deep vein thrombosis in patients with traumatic brain injury.

[0006] To achieve the above purpose, the present application adopts the following technical scheme: a method for detecting deep vein thrombosis in patients with traumatic brain injury, comprising: Data acquisition and preprocessing: Collecting deep vein image data and physiological parameter data of patients. Image data is obtained by CT device or ultrasonic device, and the image shooting position and scanning parameters are labeled during acquisition; physiological parameters are collected by a physical sign monitor, including patient body temperature, D-dimer concentration and heart rate; the image data is subjected to noise reduction processing and gray scale normalization, and the physiological parameter data is subjected to outlier rejection; Image feature extraction and thrombus region preliminary screening: Extracting vein structure and suspected thrombus features from the preprocessed image data; extracting deep vein vessel contour by edge detection algorithm, determining vessel diameter, structure characteristics and marking vessel lumen stenosis or abnormal protrusion region; based on gray scale feature analysis, calculating the mean and variance of pixel gray scale in the vessel region, when the local region gray scale value is 20%-30% higher than that of the normal vessel region and the variance is less than or equal to 15, it is marked as a thrombus region; at the same time, morphological features are extracted, including area, circularity and texture feature formation feature vector library; Physiological parameter time series analysis and risk stratification: Time trend analysis is performed on the collected physiological parameters to divide the risk grade of thrombus formation; constructing a physiological parameter time series curve, and calculating the parameter change rate by sliding window method; Image-physiological parameter fusion modeling and thrombus recognition: a fusion model is constructed to determine the thrombus attribute of the suspected area; a weighted fusion algorithm is used to assign weights to the image feature vector and physiological parameter risk value, and a fusion decision value is calculated; a decision threshold is set, when the fusion decision value > 0.7, it is determined as a thrombus area, and the thrombus type is labeled at the same time; when the fusion decision value is 0.4-0.7, it is determined as a suspected thrombus area, which needs to be verified in combination with clinical symptoms; when the fusion decision value < 0.4, it is determined as a normal blood vessel area; the area determined as thrombus is marked with position, size and boundary contour; Detection result verification and dynamic monitoring: verify the accuracy of the initial detection result and establish a dynamic monitoring mechanism; select 20%-30% of the thrombus determination area, and manually review by image experts, if the review finds misjudgment area, update the feature extraction algorithm parameters; set a dynamic monitoring period for patients diagnosed with thrombus, continuously collect images and physiological parameters, analyze the size change trend of thrombus; at the same time, record the clinical intervention measures of patients, correlate the thrombus change data after intervention, and form a closed-loop monitoring system.

[0007] Further, it further comprises: Detection report generation and data archiving: generate structured detection reports and complete data storage and archiving; the detection report includes patient basic information, image analysis results, physiological parameter trend chart, risk level assessment, and clinical suggestions; the report data is encrypted using encryption algorithm, stored in hospital PACS system, and synchronized to patient electronic medical record system; Thrombus volume calculation step: calculate the thrombus volume by reconstructing the three-dimensional model of deep vein from image sequence, and provide quantitative basis for clinical intervention; the thrombus volume calculation formula is: , in the formula, V is the thrombus volume, z1 is the z-axis coordinate of the starting layer of the thrombus in the image sequence, z2 is the z-axis coordinate of the terminal layer of the thrombus in the image sequence, and S(z) is the cross-sectional area of the thrombus in one layer of z-axis.

[0008] Thrombosis risk prediction step: based on patient historical data and real-time parameters, a risk prediction model is established to predict the probability of thrombosis in advance; the risk prediction calculation formula is: , in the formula, P is the probability of deep vein thrombosis of the patient within 24 hours in the future, n is the number of risk influencing factors, ai is the weight coefficient of the i th influencing factor, xi is the standardized value of the i th influencing factor, and bi is the bias term of the i th influencing factor.

[0009] Further, the data collection also includes venous hemodynamic parameter collection, blood flow velocity, blood flow direction, and blood flow spectrum morphology in deep veins are collected by an ultrasound Doppler device; during data preprocessing, Fourier transform is used to convert blood flow velocity data to frequency domain, peak frequency and average frequency of blood flow spectrum are extracted, and abnormal spectrum regions are marked; the hemodynamic parameters are fused with image data and conventional physiological parameters to form a three-dimensional data matrix, and the comprehensiveness of subsequent thrombus feature extraction is improved.

[0010] Further, the image feature extraction also includes using a deep learning algorithm to optimize feature extraction accuracy, constructing a lightweight convolutional neural network model, inputting preprocessed image data, extracting low-level texture features through three convolutional layers, reducing feature dimension through two pooling layers, and outputting high-level semantic features through one fully connected layer; cross-validation method is used for model training, the training set contains 1000 cases of deep vein image data of traumatic brain injury patients, and the verification set contains 200 cases of image data; Adam optimizer is used during training, cross-entropy loss function is used as the loss function, and training is stopped when the loss value of the model on the verification set does not decrease for five consecutive rounds; the high-level semantic features extracted by the model are fused with the shape and gray features extracted by traditional algorithms to form a comprehensive feature vector.

[0011] Further, the image-physiological parameter fusion modeling also includes introducing clinical symptom features, collecting patient complaints and physical examination results, and quantifying symptom features; adjusting the weights of the fusion algorithm, the image feature weight is 0.5, the physiological parameter weight is 0.3, and the clinical symptom feature weight is 0.2; and the fusion decision value is recalculated.

[0012] Further, the dynamic monitoring also includes thrombus stability evaluation, the fit degree of the thrombus and the blood vessel wall and the internal density uniformity of the thrombus are calculated through image data; combined with the D-dimer concentration change rate in the physiological parameters, when the fit degree is <0.8 and the density is uneven, and the D-dimer concentration 24h change rate is >20%, it is determined as an unstable thrombus, which prompts the risk of thrombus shedding; the warning information is pushed to the clinician, and it is suggested to take anticoagulation intensive treatment or physical intervention measures.

[0013] Further, it also includes: The detection result visualization step: a three-dimensional model of the deep vein and the thrombus is constructed by using three-dimensional reconstruction technology, and different regions are distinguished by color coding; the thrombus size, position, and stability evaluation results are labeled on the model, the doctor can view the thrombus details by mouse dragging and zooming operation; at the same time, a thrombus development trend curve is generated, which displays the thrombus change law by taking time as the horizontal axis and thrombus volume and D-dimer concentration as the vertical axis; the visualization result can be exported as DICOM format and opened in the hospital's conventional image viewing software.

[0014] Further, the data archiving also includes establishing patient follow-up data association, storing the detection report in association with the subsequent follow-up data of the patient, using a data mining algorithm to analyze the correlation between the detection result and the follow-up result, optimizing the detection model parameters, and updating the optimized model parameters to the detection system to realize continuous iterative optimization of the detection method.

[0015] Further, the data collection is also adapted to a portable ultrasound device to meet the bedside detection requirement, the image data collected by the portable ultrasound device is transmitted to the data processing terminal in real time through 5G communication, a light-weight noise reduction algorithm is used for data preprocessing to adapt to the computing power of the terminal device, and physiological parameter collection is realized by wearing equipment to realize 24h continuous monitoring, and data is synchronized to the terminal through Bluetooth.

[0016] Compared with the prior art, the present application has the following advantages: At the data collection level, the present application integrates medical images, physiological parameters, hemodynamic data and clinical symptom information to avoid the limitations of single data dimension: after image data is denoised and normalized, thrombus features are extracted by combining edge detection and deep learning algorithm to reduce subjective factor interference; the introduction of hemodynamic parameters can capture blood flow changes caused by early thrombus, and the quantitative fusion of clinical symptoms further supplements thrombus-related sign information, making thrombus feature extraction more comprehensive, effectively reducing missed and false judgments caused by data loss, and improving the recognition ability of early microthrombus.

[0017] In terms of detection accuracy and clinical decision support, the present application greatly improves the detection reliability through multi-data fusion modeling and quantitative evaluation: the weighted fusion algorithm of image-physiology-clinical symptoms can comprehensively judge the thrombus attribute by integrating information in each dimension, and reduce false positives and false negatives caused by single data; thrombus volume calculation and stability evaluation provide quantitative basis for clinical practice - volume data can guide anticoagulant drug dosage adjustment, stability analysis can predict thrombus shedding risk, and timely push warning information to avoid serious complications such as pulmonary embolism; the risk prediction model can predict thrombus formation probability in advance by combining historical data and real-time parameters, helping clinicians develop intensive monitoring programs for high-risk patients and improve the timeliness of intervention.

[0018] In terms of clinical applicability and iterative optimization, the present application fully adapts to the special needs of patients with traumatic brain injury: the dynamic monitoring mechanism can adjust the review frequency according to the patient's risk level, avoiding excessive detection or insufficient monitoring; the three-dimensional visualization technology intuitively displays the position, shape and change trend of the thrombus, facilitating multi-department collaborative consultation; bedside detection adapts to portable equipment and 5G transmission, meeting the needs of patients in ICU and neurosurgery who are inconvenient to get out of bed, and shortening the detection time; the follow-up data association and model optimization mechanism can continuously update algorithm parameters using clinical feedback, allowing the detection method to continuously improve with clinical practice and maintain high accuracy for a long time, providing efficient and safe technical support for the whole process management of deep vein thrombosis in patients with traumatic brain injury. BRIEF DESCRIPTION OF DRAWINGS

[0019] Fig. 1 A schematic block diagram of a detection method for deep vein thrombosis in patients with traumatic brain injury according to the present application is provided. Fig. 2 A data fusion weight and score distribution diagram for thrombus determination of a detection method for deep vein thrombosis in patients with traumatic brain injury according to the present application is provided. Fig. 3 A schematic diagram of the relationship between thrombus volume and anticoagulant drug dosage of a detection method for deep vein thrombosis in patients with traumatic brain injury according to the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0021] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0022] Moreover, the terms "first", "second", etc. are used herein for descriptive purposes only and should not be construed as indicating or implying relative importance or a quantity of indicated technical features. Thus, features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise explicitly and specifically limited. In addition, the terms "mounting", "connecting", "connection" should be broadly understood, for example, it can be fixed connection, or detachable connection, or integrally connected; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through intermediate medium, or internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances, and the present application will be further described in detail below with reference to the drawings.

[0023] Reference Figs. 1 to 3 A method for detecting deep vein thrombosis of traumatic brain injury patients; the method for detecting deep vein thrombosis of traumatic brain injury patients is based on medical image and physiological parameter fusion analysis, realizes early and accurate identification of thrombosis, adapts to CT venography (CTV), ultrasound image and other multi-type image data, and the operation process includes the following steps, the data transmission delay of each step is ≤200ms, and the accuracy of the detection result is ≥92%; the steps are as follows: Data acquisition and preprocessing: collect deep vein image data and physiological parameter data of patients. Image data is obtained through a CT device (resolution ≥512x512 pixels, layer thickness 0.5-1mm) or an ultrasound device (frequency 5-10MHz), covering deep vein areas prone to thrombosis in lower limbs, pelvic cavity, etc. Image shooting position (such as supine position, lateral position) and scanning parameters (such as tube voltage 120-140kV) are labeled during acquisition; physiological parameters are collected by a physical sign monitor, including patient body temperature (measurement range 35-42℃, accuracy ±0.1℃), D-dimer concentration (detection range 0-5μg / mL, accuracy ±0.05μg / mL), heart rate (measurement range 40-180 times / min, accuracy ±1 times / min), acquisition frequency 1 / 5min, continuous acquisition for 24h; image data is subjected to noise reduction processing (Gaussian filtering algorithm, standard deviation σ=1.5), gray scale normalization (pixel value is mapped to the range of 0-255), and physiological parameter data is subjected to outlier rejection (3σ principle, data exceeding the mean value ±3 times the standard deviation is rejected), so that the data integrity is ≥98%.

[0024] Image feature extraction and thrombus region preliminary screening: Extract the venous structure and suspected thrombus features from the pre-processed image data. The edge detection algorithm (Canny operator, threshold 100-200) is used to extract the deep vein vessel contour, determine the structural features such as vessel diameter, direction, etc., and mark the vessel lumen stenosis or abnormal protrusion area; based on gray scale feature analysis, calculate the mean and variance of the pixel gray scale in the vessel area, when the local area gray scale value is 20%-30% higher than the normal vessel area and the variance is ≤15, it is marked as a suspected thrombus area; at the same time, the morphological features of the suspected area are extracted, including area (pixel number), circularity (4π×area / perimeter², value 0-1), texture features (contrast and correlation calculated by gray level co-occurrence matrix, contrast value 0-1000, correlation value -1-1), forming a feature vector library.

[0025] Physiological parameter time series analysis and risk stratification: Time series trend analysis is performed on the collected physiological parameters to divide the thrombosis risk level. Construct the physiological parameter time series curve, calculate the parameter change rate by sliding window method (window size 30 min), such as D-dimer concentration change rate within 24 h, heart rate fluctuation amplitude; set the risk stratification standard, when D-dimer concentration > 2 μg / mL and 24 h change rate > 15%, or heart rate fluctuation amplitude > 20 times / min with body temperature > 38℃, it is determined as high risk; when D-dimer concentration 1-2 μg / mL and change rate 5%-15%, or heart rate fluctuation amplitude 10-20 times / min, it is determined as medium risk; otherwise, it is determined as low risk; associate the risk stratification results with the image suspected area features, focus on analyzing the suspected areas with diameter > 5 mm for high-risk patients, and reduce the image analysis frequency for low-risk patients.

[0026] Image-physiological parameter fusion modeling and thrombus identification: Construct a fusion model to determine the thrombus attributes of the suspected area. Use a weighted fusion algorithm to assign weights to the image feature vector and physiological parameter risk value (image feature weight 0.6, physiological parameter weight 0.4), calculate the fusion decision value; set the decision threshold, when the fusion decision value > 0.7, it is determined as a thrombus area, and the thrombus type is also marked (such as red thrombus, white thrombus, based on gray value and texture features: red thrombus with high gray value and uniform texture, white thrombus with medium gray value and rough texture); when the fusion decision value is 0.4-0.7, it is determined as a suspected thrombus area, which needs to be further verified combined with clinical symptoms; when the fusion decision value < 0.4, it is determined as a normal vessel area; for the area determined as a thrombus, mark its position (such as femoral vein, popliteal vein), size (long diameter, short diameter, accurate to 0.1 mm) and boundary contour.

[0027] Detection result verification and dynamic monitoring: verify the accuracy of the initial detection result and establish a dynamic monitoring mechanism. Select 20%-30% of the thrombus judgment area, and manually review it by image experts (Kappa coefficient ≥0.85, to ensure consistency with machine judgment), if the review finds misjudgment area (such as misjudging blood vessel wall calcification as thrombus), update the feature extraction algorithm parameters; for patients diagnosed with thrombus, set a dynamic monitoring period (high-risk patients 1 time / 6h, medium-risk patients 1 time / 12h), continuously collect images and physiological parameters, analyze the trend of thrombus size change (such as 24h within 2mm, prompt thrombus progression risk); at the same time, record the patient's clinical intervention measures (such as the type and dose of anticoagulant drugs), correlate the thrombus change data after intervention, and form a closed-loop monitoring system.

[0028] Detection report generation and data archiving: generate structured detection reports and complete data storage and archiving. The detection report includes patient basic information (name, medical record number, detection time), image analysis results (thrombus location, size, type picture annotation), physiological parameter trend chart, risk level assessment, clinical suggestion (such as whether anticoagulant therapy is needed, review period); use encryption algorithm (SM3 hash algorithm) to encrypt report data, store it in hospital PACS system (support data retention ≥10 years), at the same time, synchronize to patient electronic medical record system, convenient for clinical doctors to call and view at any time, according to hospital information security level, data access permission is divided (such as the attending physician can modify the clinical suggestion, the nurse can only view the report).

[0029] In the present application, it also includes: Thrombus volume calculation step: reconstruct the three-dimensional model of deep vein through image sequence, accurately calculate the volume of thrombus, and provide quantitative basis for clinical intervention. The thrombus volume calculation formula is: , wherein V is the volume of thrombus (unit: mm³), z1 is the z-axis coordinate of the starting layer of thrombus in the image sequence (unit: mm, determined based on the scanning coordinate system of CT or ultrasound equipment), z2 is the z-axis coordinate of the terminal layer of thrombus in the image sequence (unit: mm), S(z) is the cross-sectional area of thrombus at a certain layer of z-axis (unit: mm², calculated by multiplying the number of pixels of thrombus area in the image of the layer by the area of a single pixel, the area of a single pixel = (image resolution * scanning layer thickness) / (pixel matrix row number * pixel matrix column number)). For example, the femoral vein thrombus of a patient has a starting layer z1=10mm and a terminal layer z2=30mm, and the cross-sectional area S(z) of each layer changes linearly with z-axis (S=8mm² when z=10mm, S=12mm² when z=30mm), then V= [(8+(12-8)(z-10) / (30-10))]dz= (8+0.2z-2)dz= =(180+90)-(60+10)=200mm³. Clinically, the dosage of anticoagulant drugs can be adjusted according to the volume. When the volume is >200mm³, anticoagulant therapy needs to be strengthened.

[0030] Thrombosis risk prediction steps: Based on patients' historical data and real-time parameters, establish a risk prediction model to predict the probability of thrombosis in advance. The risk prediction calculation formula is: In the formula, P is the probability of a patient developing deep vein thrombosis within the next 24 hours (dimensionless, ranging from 0 to 1, P ≥ 0.6 indicates high risk, 0.3 ≤ P < 0.6 indicates medium risk, and P < 0.3 indicates low risk), n is the number of risk influencing factors (in this example, n = 5, including D-dimer concentration, bed rest time, heart rate, body temperature, and trauma severity score), and ai is the weight coefficient of the i-th influencing factor (obtained by training historical data using a logistic regression algorithm, such as the weight of D-dimer concentration). =0.3, weight of bed rest time =0.25), xi is the standardized value of the i-th influencing factor (mapping the original data to the range of 0-1, such as D-dimer concentration of 5 μg / mL corresponding to ...). =1, 0 μg / mL corresponds to =0), bi is the bias term of the i-th influencing factor (such as the D-dimer concentration bias term). =0.05, Bed rest time bias term =0.03). For example, a patient =0.8 (D-dimer 4 μg / mL) =0.9 (bed rest time 72h) =0.6 (heart rate 95 beats / min) =0.7 (body temperature 38.2℃), =0.8 (trauma score of 15). - The values ​​are 0.3, 0.25, 0.15, 0.15, and 0.15, respectively. - The values ​​are 0.05, 0.03, 0.02, 0.02, and 0.03 respectively. Therefore, ∑(aixi+bi)=0.3×0.8+0.05+0.25×0.9+0.03+0.15×0.6+0.02+0.15×0.7+0.02+0.15×0.8+0.03=0.24+0.05+0.225+0.03+0.09+0.02+0.105+0.02+0.12+0.03=0.91. P=1-e^(-0.91)≈1-0.402=0.598, which is judged as medium risk. The monitoring cycle needs to be shortened to 8h / time.

[0031] In this invention, the multi-source data acquisition also includes the acquisition of venous hemodynamic parameters. Deep vein blood flow velocity (measurement range 0-100 cm / s, accuracy ±1 cm / s), blood flow direction (towards or away from the probe), and blood flow spectrum morphology (e.g., continuous spectrum, discontinuous spectrum) are acquired using an ultrasonic Doppler device (sampling volume 1-2 mm, angle correction ≤ 60°). During data preprocessing, the blood flow velocity data is transformed to the frequency domain using Fourier transform (sampling frequency 1000 Hz), and the peak frequency and average frequency of the blood flow spectrum are extracted. Abnormal spectral regions are marked (e.g., reverse blood flow, spectral dispersion > 30%). The hemodynamic parameters are fused with imaging data and conventional physiological parameters to form a three-dimensional data matrix, improving the comprehensiveness of subsequent thrombosis feature extraction. For example, when the blood flow velocity is < 10 cm / s and accompanied by reverse blood flow, the corresponding vascular area is highlighted as a suspected high-incidence area for thrombosis.

[0032] In this invention, the image feature extraction further includes optimizing feature extraction accuracy using deep learning algorithms, constructing a lightweight convolutional neural network model (based on the MobileNetV3 architecture, with ≤5M parameters), inputting preprocessed image data, extracting low-level texture features through 3 convolutional layers (3×3 kernel size, stride 1), reducing feature dimensionality through 2 pooling layers (max pooling, 2×2 kernel size), and outputting high-level semantic features (128-dimensional feature dimension) through 1 fully connected layer; model training employs cross-validation (5-fold cross-validation, accuracy ≥93%). The training set contains deep vein imaging data from 1000 patients with traumatic brain injury (including 300 thrombosis images), and the validation set contains 200 images. During training, the Adam optimizer (learning rate 0.001, decay rate 0.0001) is used, and the cross-entropy loss function is used. Training is stopped when the model's loss value on the validation set does not decrease for 5 consecutive rounds. The high-level semantic features extracted by the model are fused with the morphological and grayscale features extracted by traditional algorithms to form a comprehensive feature vector, which improves the accuracy of suspected thrombosis area identification by 5%-8%.

[0033] In this invention, the image-physiological parameter fusion modeling also includes incorporating clinical symptom features, collecting patient complaints (such as lower limb swelling and pain) and physical examination results (such as Homans' sign and Neuhof's sign), quantifying the symptom features (assigning lower limb swelling degree as "no swelling = 0, mild swelling = 1, moderate swelling = 2, severe swelling = 3", and pain degree using a VAS score of 0-10); adjusting the fusion algorithm weights, with image features weighted at 0.5 and physiological parameters weighted at 0.3, and clinical... With a symptom feature weight of 0.2, the fusion judgment value is recalculated. For example, if a patient's imaging feature corresponds to a judgment value of 0.6, physiological parameter risk value of 0.8, and clinical symptom quantification value of 0.7 (moderate swelling + VAS score of 4), then the fusion judgment value = 0.5 × 0.6 + 0.3 × 0.8 + 0.2 × 0.7 = 0.3 + 0.24 + 0.14 = 0.68, which is close to the thrombosis judgment threshold. Further confirmation is needed in conjunction with 24-hour dynamic monitoring results to avoid misjudgment caused by relying solely on imaging or physiological parameters.

[0034] In this invention, the dynamic monitoring also includes thrombus stability assessment. This involves calculating the adhesion between the thrombus and the vessel wall (adhesion = contact area between thrombus and vessel wall / thrombus surface area, value 0-1; adhesion ≥ 0.8 indicates a stable thrombus, < 0.8 indicates an unstable thrombus) and the internal density homogeneity of the thrombus (measured using standard deviation; standard deviation ≤ 10 indicates homogeneity, > 10 indicates non-homogeneity) based on physiological parameters such as the D-dimer concentration change rate. When the adhesion is < 0.8 and the density is non-homogeneous, and the 24-hour D-dimer concentration change rate is > 20%, it is determined to be an unstable thrombus, indicating a risk of thrombus detachment. An early warning message is immediately sent to the clinician, recommending enhanced anticoagulation therapy or physical intervention measures (such as using an intermittent pneumatic pressurization device), and shortening the monitoring cycle to 4 hours / time, continuously monitoring changes in thrombus stability until the adhesion is ≥ 0.8 and the density is homogeneous.

[0035] This invention also includes: Visualization steps for detection results: A 3D reconstruction model of deep veins and thrombi is constructed using 3D reconstruction technology (based on the MarchingCubes algorithm, with a reconstruction accuracy ≤0.5mm). Different regions are distinguished by color coding (normal vessels are blue, stable thrombi are yellow, and unstable thrombi are red). The size, location, and stability assessment results of the thrombus are marked on the model, allowing doctors to view thrombus details (such as the spatial relationship between the thrombus and adjacent vessels) by dragging and zooming with the mouse. At the same time, a thrombus development trend curve is generated, with time as the horizontal axis and thrombus volume and D-dimer concentration as the vertical axis, intuitively showing the thrombus change pattern. The visualization results can be exported to DICOM format, which can be opened in the hospital's conventional image viewing software, facilitating multidisciplinary consultations and improving clinical decision-making efficiency.

[0036] In this invention, the data archiving also includes establishing a correlation between patient follow-up data, linking test reports with subsequent patient follow-up data (such as thrombosis re-examination results 1 month and 3 months after discharge, and whether complications such as pulmonary embolism occurred); using data mining algorithms (such as decision tree algorithms) to analyze the correlation between test results and follow-up results, and optimizing test model parameters (such as adjusting the weights of the fusion algorithm and the coefficients in the risk prediction formula); for example, through follow-up, it was found that the weight of the trauma severity score in the original risk prediction formula was too low, resulting in a lower risk prediction value for some patients with high trauma scores. Based on the follow-up data, the logistic regression model was retrained, and the weight of the trauma severity score was adjusted from 0.15 to 0.2 to improve the accuracy of risk prediction; at the same time, the optimized model parameters were updated to the test system to achieve continuous iterative optimization of the test method.

[0037] In this invention, the data acquisition is also adapted to portable ultrasound equipment to meet bedside testing needs. The image data acquired by the portable ultrasound equipment (weight ≤3kg, battery life ≥4h) is transmitted to the data processing terminal in real time via 5G communication (transmission rate ≥100Mbps). Lightweight noise reduction algorithms (such as bilateral filtering, reducing computational complexity by 40%) are used during data preprocessing to adapt to the computing power of the terminal device. Physiological parameter acquisition uses wearable devices (such as wrist pulse oximeters and patch-type body temperature sensors) to achieve 24-hour continuous monitoring, and the data is synchronized to the terminal via Bluetooth. The bedside testing process takes ≤15 minutes from data acquisition to the output of preliminary test results, meeting the clinical needs of patients with traumatic brain injury who are bedridden and require rapid assessment, and is especially suitable for bedside scenarios such as ICU and neurosurgical wards.

[0038] A specific implementation method for detecting deep vein thrombosis in patients with traumatic brain injury: Example 1: Detection of deep vein thrombosis in ICU patients with severe traumatic brain injury.

[0039] This embodiment describes the use of the detection method of this invention to screen for deep vein thrombosis in the lower extremities of a 50-year-old male patient with traumatic brain injury (GCS score of 8, closed head injury, 5 days post-surgery) admitted to the ICU of a tertiary hospital. The detection process took 12 minutes and the patient was finally diagnosed with acute red thrombosis in the right femoral vein, providing a precise basis for clinical anticoagulation therapy.

[0040] I. Testing Process and Technical Details Multi-source data acquisition and preprocessing: A 128-slice spiral CT scanner (512×512 pixels resolution, 0.6mm slice thickness, 130kV tube voltage, 250mA tube current) was used to acquire images of the patient's bilateral lower extremities and pelvic deep veins. The patient was in a supine position, and scanning parameters were simultaneously labeled. Physiological parameters were acquired using a multi-parameter monitor (Philips IntelliVue MX800): body temperature 37.8℃, D-dimer concentration 4.2μg / mL, heart rate 98 bpm, acquisition frequency 1 time / 5min, continuous acquisition for 24 hours. Additionally, color Doppler ultrasound (GE Logiq E9, 7.5MHz frequency, 1.5mm sampling volume, 55° angle correction) was used to acquire hemodynamic parameters of the right femoral vein: blood flow velocity 8.5cm / s (normal reference value 15-25cm / s), reverse blood flow signal appeared when the blood flow direction was away from the probe, the blood flow spectrum showed discontinuous spectral dispersion of 35%. Gaussian filtering (σ=1.5) was used to reduce noise in CT images, and grayscale normalization mapped pixel values ​​to 0-255. One abnormal heart rate value (132 beats / min, exceeding the mean ± 3 times the standard deviation) was removed using the 3σ principle for physiological parameters, achieving 99% data integrity. Blood flow velocity data was converted to the frequency domain using Fourier transform (sampling frequency 1000Hz), and the peak frequency of 2.3kHz and the average frequency of 1.1kHz were extracted to mark the spectral abnormal regions corresponding to reverse blood flow.

[0041] Image feature extraction and initial screening of thrombus areas: The Canny operator (threshold 120-180) was used to extract the contour of the right femoral vein. The vessel diameter was measured to be 8.2 mm. A stenotic area (stenosis rate of approximately 40%) was marked 1.5 cm from the femoral vein valve. The mean gray value of the vessel area was calculated to be 68, and the mean gray value of the local stenotic area was 89 (31% higher than the normal area), with a variance of 12 (≤15). This area was marked as a suspected thrombus area. The morphological features of this area were extracted: area 2800 pixels (equivalent to an actual area of ​​approximately 0.18 cm²), roundness 0.62, and texture features (contrast 320). With a correlation of 0.75, a 128-dimensional feature vector is formed. At the same time, a lightweight convolutional neural network model (MobileNetV3 architecture, 4.8M parameters, batch size=32, epoch=50) is used as input. The preprocessed CT image is input, and texture features are extracted by 3 convolutional layers (3×3 convolutional kernel, stride 1, activation function ReLU), dimensionality is reduced by 2 max pooling layers (2×2 pooling kernel), and high-level semantic features are output by 1 fully connected layer. The model output is fused with traditional features, and the accuracy of suspected thrombosis area identification reaches 95%, which is 7% higher than that of a single traditional algorithm.

[0042] Physiological parameter time-series analysis and risk stratification: A 24-hour physiological parameter time-series curve was constructed, and a 30-minute sliding window was used for calculation: D-dimer concentration increased from 2.1 μg / mL to 4.2 μg / mL, with a 24-hour change rate of 100%; heart rate fluctuation range was 25 beats / min (mean 92 beats / min, highest 108 beats / min, lowest 83 beats / min); body temperature reached a maximum of 38.1℃, meeting the criteria of "D-dimer > 2 μg / mL and change rate > 15% + heart rate fluctuation range > 20 beats / min + body temperature > 38℃", and was judged as high risk, with a focus on analyzing suspected areas with a diameter > 5 mm.

[0043] Multi-data fusion modeling and thrombus identification: Clinical symptom features were introduced—the patient complained of mild swelling of the right lower extremity (quantitative value 2, moderate swelling), VAS pain score of 3 (mild pain), and positive Homans sign (quantitative value 1); the fusion weights were adjusted: image features 0.5, physiological parameters 0.3, and clinical symptoms 0.2; the fusion judgment value was calculated: the judgment value corresponding to the image features was 0.82 (based on grayscale, morphology, and semantic features), the risk value of the physiological parameters was 0.91 (corresponding to high risk), and the quantitative value of the clinical symptoms was 0.65. The fusion judgment value = 0.5×0.82+0.3×0.91+0.2×0.65=0.41+0.273+0.13=0.813 (>0.7), which was determined to be a thrombus area; combined with high grayscale value (89) and uniform texture (variance 12), it was diagnosed as a red thrombus. The marked location was: the right femoral vein 1.5-3.2cm away from the valve, with a long axis of 17.3mm and a short axis of 6.8mm.

[0044] Thrombus volume calculation and stability assessment: According to the thrombus volume calculation formula, the thrombus originates at a slice z1 of 5 mm and terminates at a slice z2 of 25 mm in the CT imaging sequence. The cross-sectional area S(z) of each slice increases linearly with the z-axis (S=4.2 mm² when z=5 mm, S=9.8 mm² when z=25 mm). Therefore, V= [4.2+(9.8-4.2)(z-5) / (25-5)]dz= (4.2 + 0.28z - 1.4)dz = (2.8 + 0.28z)dz = =(70+87.5)-(14+3.5)=140mm³; Stability assessment: The contact area between the thrombus and the blood vessel wall is 12.3mm², the surface area of ​​the thrombus is 17.1mm², the fit is 12.3 / 17.1≈0.72 (<0.8), the standard deviation of the internal density of the thrombus is 14 (>10), and the 24-hour change rate of D-dimer is 100% (>20%). It is judged as an unstable thrombus, and an alert is sent to the clinician.

[0045] Risk prediction and dynamic monitoring: Calculate the probability of thrombosis in the next 24 hours according to the risk prediction formula, n=5 (D-dimer). =0.84, bed rest time =0.92, heart rate =0.65, body temperature =0.76, trauma score =0.8), weight =0.3、 =0.25、 =0.15、 =0.15、 =0.15, bias term =0.05、 =0.03、 =0.02、 =0.02、 =0.03, ∑(aixi+bi)=0.3×0.84+0.05+0.25×0.92+0.03+0.15×0.65+0.02+0.15×0 .76+0.02+0.15×0.8+0.03=0.252+0.05+0.23+0.03+0.0975+0.02+0.114+0.02+ 0.12 + 0.03 = 0.9335, P = 1 - e^(-0.9335) ≈ 0.61, indicating a high risk. Dynamic monitoring was set to once every 6 hours. Subsequent monitoring showed no increase in the long diameter of the thrombus within 24 hours, and the D-dimer decreased to 2.8 μg / mL. The anticoagulant dosage was adjusted (low molecular weight heparin was changed from 4000 IU / q12h to 3000 IU / q12h).

[0046] Test Report and Data Archiving: Generate a structured report containing basic patient information, CT image annotations (color-coded image of the location and size of the right femoral vein thrombosis), hemodynamic parameter tables (blood flow velocity, spectral characteristics), risk assessment results (high risk, unstable thrombosis), and clinical recommendations (low molecular weight heparin anticoagulation, monitoring every 6 hours). The report is encrypted using the SM3 algorithm and stored in the hospital's PACS system (data retention for 15 years), synchronized with the electronic medical record system, allowing attending physicians to modify treatment recommendations, while nurses only view the monitoring plan. Follow-up data for one month (thrombosis disappearance, no complications) is stored in conjunction with the data, and the trauma score weight in the risk prediction formula is updated based on this (from 0.15 to 0.18), improving the prediction accuracy for similar patients.

[0047] II. Performance Verification Form

[0048] This table is based on statistical analysis of data from 50 ICU patients. Traditional ultrasound examination, relying on physician experience, has an 80% recognition rate for microthrombi <5mm in diameter. Thrombus volume estimation by visual inspection has a 30% error, and stability cannot be quantified, resulting in an 18% false positive rate (due to misjudgment caused by elevated D-dimer levels). This invention, by integrating multiple data sources, increases the microthrombus recognition rate to 95%, reduces volume calculation error to 8%, achieves 92% unstable thrombus warning through adhesion and density analysis, and has a false positive rate of only 5% (excluding the influence of a single D-dimer level). The process time is shortened to 12 minutes (without requiring multiple adjustments to patient position). For example, in this embodiment, traditional ultrasound did not detect abnormal adhesion between the thrombus and the vessel wall. This invention accurately identifies unstable thrombi and issues warnings, avoiding the risk of thrombus detachment due to insufficient anticoagulation, and fully meets the precise, rapid, and safe testing needs of ICU patients.

[0049] Example 2: Detection of deep vein thrombosis in a neurosurgical patient with traumatic brain injury who is on bed rest.

[0050] This embodiment addresses a 42-year-old female patient with traumatic brain injury (GCS score of 12, cerebral contusion, 3 days post-surgery) admitted to the neurosurgery department of a hospital. The bedside detection mode of this invention was used to screen for deep vein thrombosis. The detection process took 15 minutes. Thrombosis was ruled out, but the patient was determined to be at medium risk of thrombosis formation. A targeted prevention plan was then developed.

[0051] I. Testing Process and Technical Details Multi-source data acquisition and preprocessing: Bilateral calf deep vein images (anterior and posterior tibial veins) were acquired bedside using a portable ultrasound device (SonositeEdgeII, 2.8kg, 5h battery life, 6.5MHz frequency). The patient was in a lateral decubitus position (the patient could not lie supine), and the scanning parameters were labeled "frequency 6.5MHz, gain 50dB". Physiological parameters were acquired using a patch-type vital signs monitor (iHealthAM3): body temperature 37.2℃, D-dimer concentration 1.8μg / mL, heart rate 85 bpm, continuously acquired for 24h, acquisition frequency 1 time / 5min. Hemodynamic parameters were acquired using a portable Doppler ultrasound: left posterior tibial vein blood flow velocity 12cm / s, right posterior tibial vein 11cm / s, no reverse blood flow, and the spectrum was continuous (dispersion 18%). Image preprocessing employs lightweight bilateral filtering (reducing computational complexity by 40% and adapting to device computing power), with pixel values ​​ranging from 0 to 255 after grayscale normalization; physiological parameters are removed without outliers using the 3σ principle, resulting in 99.5% data integrity; blood flow velocity data undergoes Fourier transform (sampling frequency 800Hz), with a peak frequency of 1.8kHz and an average frequency of 0.9kHz, and no spectral anomalies are detected.

[0052] Image feature extraction and initial screening of thrombus areas: The Canny operator (threshold 100-160) was used to extract the contours of the bilateral posterior tibial veins. The vessel diameter was 4.5-5.0 mm, with no luminal stenosis or bulge. The mean gray value of the vessel area was calculated to be 58, and the highest gray value in the local area was 65 (12% higher than the normal area, but not reaching the 20% threshold). The variance was 18 (>15), and there were no suspected thrombus areas that met the criteria. A lightweight convolutional neural network model was enabled (same as in Example 1, adapted to the computing power of portable devices, inference speed 25fps). After inputting ultrasound images, high-level semantic features were output. After fusion with traditional morphological features, no suspected areas were marked, and the initial screening result was "no clear signs of thrombosis".

[0053] Physiological parameter time-series analysis and risk stratification: A 24-hour time-series curve was constructed, and a 30-minute sliding window was used for calculation: D-dimer increased from 1.2 μg / mL to 1.8 μg / mL, with a 24-hour change rate of 50%; heart rate fluctuation range was 18 beats / min (mean 82 beats / min, highest 95 beats / min, lowest 77 beats / min); body temperature peaked at 37.5℃ (<38℃), meeting the criteria of "D-dimer 1-2 μg / mL with a change rate of 5%-15%" (actually 50% exceeding the threshold, heart rate fluctuation 10-20 beats / min), and was judged as medium risk, requiring enhanced monitoring (12 hours / time).

[0054] Multi-data fusion modeling and thrombus identification: Clinical symptom features were introduced—patients complained of no swelling in both lower legs (quantitative value 0), no pain (VAS score 0), and negative Homans sign (quantitative value 0); fusion weights were calculated as 0.5 for imaging, 0.3 for physiological, and 0.2 for clinical. The fusion judgment value = 0.5 × 0.3 (no thrombus imaging features) + 0.3 × 0.6 (medium-risk physiological parameters) + 0.2 × 0.1 (no clinical symptoms) = 0.15 + 0.18 + 0.02 = 0.35 (< 0.4), which was judged as a normal vascular area, excluding thrombus.

[0055] Risk prediction and dynamic monitoring: Calculated according to the risk prediction formula, n=5 (D-dimer) =0.36, bed rest time =0.38, heart rate =0.45, body temperature =0.4, trauma score =0.5), weights and biases are the same as in Example 1, ∑(aixi+bi)=0.3×0.36+0.05+0.25×0.38+0.03+0.15×0.45+0.02+0.15×0.4+0.02+0.15×0.5+0.03=0.108+0.05+0.095+0.03+0.0675+0.02+0.06+0.02+0.075+0.03=0.5655, P=1-e^(-0.5655)≈0.43, which is judged as medium risk; the dynamic monitoring cycle is set to 1 time / 12h, and data is continuously collected. On the 3rd day, the D-dimer drops to 1.5μg / mL, with a change rate of 16.7%, and is adjusted to low risk (monitoring every 24h).

[0056] Bedside visualization and data archiving: Lightweight 3D reconstruction (MarchingCubes algorithm, reconstruction accuracy 0.8mm, compatible with portable devices) is used to construct a 3D model of the deep veins in both lower legs (normal vessels are marked in blue), with no thrombosis areas; a simple bedside report is generated (including ultrasound image screenshots, physiological parameter trend charts, and risk levels), which is transmitted in real time to the neurosurgeon's mobile device via 5G (transmission rate 120Mbps); when archiving the data, it is linked to the subsequent 2-week follow-up data (no thrombosis), and based on this, the weight of hemodynamic parameters in risk prediction is optimized (from 0.15 to 0.17), improving the prediction accuracy for intermediate-risk patients.

[0057] Performance Verification Form

[0058] This table is based on statistical analysis of bedside examination data from 80 neurosurgical patients who were typically bedridden. Traditional CT scans require transporting patients to the CT room (high risk during transport of bedridden patients, such as fluctuations in intracranial pressure), the equipment cannot be moved, data transmission and analysis take 60 seconds, and bedside reporting takes 40 minutes. This invention uses a portable device for bedside examination, eliminating the need to move the patient (avoiding transport risks). The device weighs 2.8 kg, 5G transmission takes 5 seconds, and a report is generated on-site in 5 minutes. The false negative rate for low-risk patients is only 3% (compared to 12% with traditional methods). For example, in this embodiment, the patient could not be transported due to unstable intracranial pressure, making traditional CT examination impossible. This invention completes bedside examination and rule out thrombosis within 15 minutes, while accurately determining medium-risk patients, avoiding excessive anticoagulation or insufficient prevention. After optimization of follow-up data association, the accuracy of risk prediction for similar patients in subsequent cases improved by 4%, fully meeting the needs of convenient, low-risk, and accurate preventive bedside examination for neurosurgical patients who are typically bedridden.

[0059] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting deep vein thrombosis in a patient with traumatic brain injury, characterized by, Comprise: Data acquisition and preprocessing: Collect patient deep vein image data and physiological parameter data. Image data is obtained through CT equipment or ultrasonic equipment, and the image shooting body position and scanning parameters are labeled during acquisition; physiological parameters are collected by a physical sign monitor, including patient body temperature, D-dimer concentration, and heart rate; the image data is subjected to noise reduction processing and gray scale normalization, and the physiological parameter data is subjected to outlier rejection; Image feature extraction and thrombus region preliminary screening: extract vein structure and suspected thrombus features from the preprocessed image data; Adopt edge detection algorithm to extract deep vein vessel contour, determine vessel diameter, direction structure characteristics, and mark vessel lumen stenosis or abnormal protrusion region; based on gray scale feature analysis, calculate the mean and variance of the pixel gray scale in the vessel region, and when the local region gray scale value is 20%-30% higher than that of the normal vessel region and the variance is less than or equal to 15, mark it as a thrombus region; at the same time, extract morphological features, including area, circularity, and texture feature to form a feature vector library; Physiological parameter time series analysis and risk stratification: perform time series trend analysis on the collected physiological parameters to stratify thrombus formation risk levels; Construct a physiological parameter time series curve and calculate the parameter change rate by the sliding window method; Image-physiological parameter fusion modeling and thrombus recognition: build a fusion model to determine the thrombus attribute of the suspected region; use a weighted fusion algorithm to assign weights to the image feature vector and physiological parameter risk value, and calculate the fusion decision value; set a decision threshold, when the fusion decision value is greater than 0.7, determine it as a thrombus region, and at the same time label the thrombus type; When the fusion decision value is between 0.4 and 0.7, it is determined as a suspected thrombus region, which needs to be verified in combination with clinical symptoms; when the fusion decision value is less than 0.4, it is determined as a normal vessel region; mark the position, size and boundary contour of the region determined as a thrombus; Detection result verification and dynamic monitoring: verify the accuracy of the initial detection result and establish a dynamic monitoring mechanism; select 20%-30% of the thrombus determination regions, and manually review them by image experts, if the review finds misjudgment regions, update the feature extraction algorithm parameters; set a dynamic monitoring period for patients diagnosed with thrombus, continuously collect image and physiological parameters, and analyze the thrombus size change trend; at the same time, record the patient's clinical intervention measures, correlate the thrombus change data after intervention, and form a closed-loop monitoring system.

2. The method for detecting deep vein thrombosis in traumatic brain injury patients according to claim 1, characterized in that, Also include: Detection report generation and data archiving: generate a structured detection report and complete data storage and archiving; The detection report contains patient basic information, image analysis results, physiological parameter trend chart, risk level evaluation, and clinical suggestions; use encryption algorithm to encrypt the report data, store it in the hospital PACS system, and at the same time synchronize it to the patient electronic medical record system; Thrombus volume calculation step: calculate the thrombus volume by reconstructing the three-dimensional model of deep vein through image sequence, and provide quantitative basis for clinical intervention; the formula for calculating the thrombus volume is: , wherein V is the thrombus volume, z1 is the z-axis coordinate of the starting layer of the thrombus in the image sequence, z2 is the z-axis coordinate of the terminal layer of the thrombus in the image sequence, and S(z) is the cross-sectional area of the thrombus at one layer of the z-axis.

3. The method of claim 1, wherein the method is for detecting deep vein thrombosis in a patient with traumatic brain injury. Also include: Thrombus formation risk prediction step: based on patient historical data and real-time parameters, establish a risk prediction model to predict the probability of thrombus formation in advance; The risk prediction calculation formula is: In the formula, P is the probability of the patient forming a deep vein thrombosis within the next 24 hours, n is the number of risk influencing factors, ai is the weight coefficient of the i th influencing factor, xi is the standardized value of the i th influencing factor, and bi is the bias term of the i th influencing factor.

4. The method of claim 1, wherein the method is for detecting deep vein thrombosis in a patient with traumatic brain injury. The data collection also includes venous hemodynamic parameter collection, blood flow velocity, blood flow direction and blood flow spectrum morphology in deep veins are collected by an ultrasonic Doppler device; during data preprocessing, the blood flow velocity data is converted to a frequency domain by Fourier transform, the peak frequency and average frequency of the blood flow spectrum are extracted, and the abnormal spectrum region is marked; the hemodynamic parameters are fused with the image data and the conventional physiological parameters to form a three-dimensional data matrix, and the comprehensiveness of subsequent thrombus feature extraction is improved.

5. The method for detecting deep vein thrombosis in traumatic brain injury patients according to claim 1, wherein, The image feature extraction also includes optimizing the feature extraction accuracy by using a deep learning algorithm, constructing a lightweight convolutional neural network model, inputting the preprocessed image data, extracting low-level texture features through 3 convolutional layers, reducing the feature dimension through 2 pooling layers, and outputting high-level semantic features through 1 fully connected layer; the model training adopts a cross-validation method, the training set contains 1000 cases of deep vein image data of traumatic brain injury patients, and the verification set contains 200 cases of image data; during the training process, an Adam optimizer is used, a cross-entropy loss function is used as the loss function, and the training is stopped when the loss value of the model on the verification set does not decrease for 5 consecutive rounds; the high-level semantic features extracted by the model are fused with the shape and gray scale features extracted by the traditional algorithm to form a comprehensive feature vector.

6. The method for detecting deep vein thrombosis in traumatic brain injury patients according to claim 1, wherein, The image-physiological parameter fusion modeling also includes introducing clinical symptom features, collecting patient complaint symptoms and physical examination results, and quantifying the symptom features; The weight of the fusion algorithm is adjusted, the image feature weight is 0.5, the physiological parameter weight is 0.3, and the clinical symptom feature weight is 0.2, and the fusion decision value is recalculated.

7. The method of claim 1, wherein the method is for detecting deep vein thrombosis in a patient with traumatic brain injury. The dynamic monitoring also includes thrombus stability evaluation, the fit degree of the thrombus and the blood vessel wall and the internal density uniformity of the thrombus are calculated through the image data; combined with the D-dimer concentration change rate in the physiological parameters, when the fit degree is <0.8 and the density is uneven, and the D-dimer concentration 24h change rate is >20%, it is determined as an unstable thrombus, which prompts the risk of thrombus shedding; the warning information is pushed to the clinician, and it is suggested to take anticoagulation intensive treatment or physical intervention measures.

8. The method of claim 1, wherein the method is for detecting deep vein thrombosis in a patient with traumatic brain injury. Further comprising: The detection result visualization step: a three-dimensional reconstruction technology is used to construct a three-dimensional model of the deep vein and the thrombus, and different regions are distinguished by color coding; the thrombus size, position and stability evaluation result are labeled on the model, the doctor can view the thrombus details through mouse dragging and zooming operation; at the same time, a thrombus development trend curve is generated, the time is taken as the horizontal axis, the thrombus volume and the D-dimer concentration are taken as the vertical axis, and the thrombus change law is displayed; the visualization result can be exported as DICOM format, and can be opened in the hospital's conventional image viewing software.

9. The method of claim 1, wherein the method is for detecting deep vein thrombosis in a patient with traumatic brain injury. The data archiving also includes establishing the association of patient follow-up data, storing the detection report and the subsequent follow-up data of the patient in association; a data mining algorithm is used to analyze the correlation between the detection result and the follow-up result, and the detection model parameters are optimized; at the same time, the optimized model parameters are updated to the detection system, and the continuous iterative optimization of the detection method is realized.

10. The method of claim 1, wherein the method is for detecting deep vein thrombosis in a patient with traumatic brain injury. The data collection is also adapted to a portable ultrasonic device to meet the bedside detection requirement, image data collected by the portable ultrasonic device is transmitted to a data processing terminal in real time through 5G communication, a light-weight noise reduction algorithm is adopted for data preprocessing to adapt to the computing power of the terminal device; physiological parameter collection is realized by wearing equipment for 24h continuous monitoring, and data is synchronized to the terminal through Bluetooth.

Images