Methods and equipment for liver ultrasound vascular tracking for estimating the motion of thoracic and abdominal tumors

Abstract

This invention discloses a method and device for liver ultrasound vascular tracking for motion estimation of thoracic and abdominal tumors, comprising: real-time acquisition of ultrasound vascular images of the patient's liver; real-time vascular detection of each frame of the image using a trained detection model; modeling the vascular motion using a Kalman filter; matching the detection box obtained by the detection model with the tracking box predicted by the motion model using a Hungarian algorithm model, and updating the motion model according to the matching result; and extracting the vascular motion trajectory based on the tracking box after Kalman filtering. This invention utilizes ultrasound vascular images to extract vascular motion curves, achieving non-invasive, radiation-free, and marker-free precise treatment.

Description

Technical Field

[0001] This invention belongs to the field of radiotherapy technology, specifically relating to a liver ultrasound vascular tracking method and device for estimating the motion of thoracic and abdominal tumors. Background Technology

[0002] Radiotherapy plays an increasingly prominent role in cancer treatment and has become one of the main methods for treating malignant tumors. It uses high-dose radiation beams to irradiate the target area of ​​the patient's lesion to kill cancer cells. However, most organs and tissues in the human body undergo displacement and deformation due to respiratory movements, especially lung and liver tumors. To increase the dose to the target area and minimize additional irradiation to organs and normal tissues, accurate and real-time modeling of tumor movement is one of the key technologies for achieving precision radiotherapy.

[0003] The movement of a tumor can be inferred by measuring other organs, tissues, or other physiological signals that are well correlated with and easy to measure. These signals can be simply classified into two types: in vivo signals and in vitro signals.

[0004] Compared to in vivo respiratory signals, external respiratory signals are a form of indirect modeling. Respiration is accompanied by the rise and fall of the chest and abdominal walls, and the inflow and outflow of air through the respiratory tract. These signals are strongly correlated with respiratory movements, and changes in these external signals can be measured to understand lung function. Commonly used external signals in respiratory movement measurement include: movement of sampling points on the body surface, movement of the body surface curvature, airflow velocity during respiration, abdominal wall pressure during respiration, and the temperature of the respiratory airflow, among others.

[0005] In vivo signals mainly include motion information from tumor tissue, landmark organs, and implanted reference markers. In vivo motion signals can be categorized into four types: motion of the tumor target area, motion of artificially implanted markers, motion of landmark organs that move synchronously with the tumor, and motion of anatomical structures rigidly connected to the tumor. Methods involving measuring anatomical structures rigidly connected to the tumor are only suitable for intracranial and spinal tumors with relatively stable tumor locations and rigid anatomical structures, and are not applicable to measurements in the thoracic and abdominal motion regions.

[0006] In summary, existing techniques for measuring tumor motion involve measuring the motion signals of other surrogates that are well correlated with tumor motion. However, in vitro signals are an indirect measurement method compared to in vivo signals, requiring the establishment of a separate correlation model between in vitro motion signals and in vivo tumor motion. This process is cumbersome and inaccurate.

[0007] Currently, the most common clinical method is to track tumor movement by implanting metal markers or other markers into or near the tumor, such as the CyberKnife radiotherapy system. This system can measure the movement of markers using X-rays and electromagnetic positioning devices, and has a high correlation with tumor movement. However, this method requires puncture to implant the marker, and the patient will receive additional radiation doses during imaging, which is quite harmful to the patient.

[0008] Respiratory movement is an irregular and unpredictable physiological movement, and radiotherapy is usually performed in 20-30 sessions. The patient's respiratory movement pattern may change during each treatment. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention proposes a liver ultrasound vascular tracking method and device for estimating the motion of thoracic and abdominal tumors.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows:

[0011] On one hand, this invention discloses a liver ultrasound vascular tracking method for estimating the motion of thoracic and abdominal tumors, comprising:

[0012] Step S1: Real-time acquisition of ultrasound vascular images of the patient's liver;

[0013] Step S2: Use the trained detection model to perform real-time blood vessel detection on each frame of the image;

[0014] The motion model of blood vessels is modeled using a Kalman filter;

[0015] The detection bounding boxes obtained by the detection model and the tracking bounding boxes predicted by the motion model are matched using the Hungarian algorithm model, and the motion model is updated based on the matching results.

[0016] Step S3: Extract the blood vessel motion trajectory based on the tracking box after Kalman filtering.

[0017] Based on the above technical solution, the following improvements can be made:

[0018] As a preferred approach, the training methods for the detection model include:

[0019] Step A: Pre-train the feature extraction network using a large natural image dataset to enable the feature extraction network to have basic feature extraction capabilities;

[0020] Step B: Construct a patient-specific liver ultrasound vascular dataset;

[0021] Step C: Based on the liver ultrasound vascular dataset, the pre-trained feature extraction network is fine-tuned through transfer learning to obtain the trained detection model.

[0022] As a preferred option, step S2 specifically includes:

[0023] Step S2.1: Use the trained detection model to perform real-time frame-by-frame blood vessel detection on the image, generate a corresponding Kalman filter for each newly appearing detection box, and model the motion model of the blood vessel using the Kalman filter.

[0024] Step S2.2: Predict the position of the tracking box in the next frame image based on the motion model, and input the prediction result into the Hungarian algorithm model.

[0025] At the same time, for the image of the current frame, the detection results of the detection model are also input into the Hungarian algorithm model;

[0026] Step S2.3: The Hungarian algorithm model matches the tracking boxes predicted by the motion model with the detection boxes obtained by the detection model using the IoU metric;

[0027] For successfully matched tracking boxes, they are marked. If N consecutive frames are successfully matched, the motion model is updated using a Kalman filter and the tracking boxes are predicted, where N≥3.

[0028] For newly appearing detection boxes, initialize new tracking boxes;

[0029] For unmatched tracking boxes, continue tracking for several frames. If no match is found, delete the tracking box.

[0030] As a preferred embodiment, step S3 further comprises: performing low-pass filtering on the tracking box after Kalman filtering to extract the blood vessel motion trajectory.

[0031] On the other hand, the present invention also discloses a liver ultrasound vascular tracking device for estimating the motion of thoracic and abdominal tumors, comprising:

[0032] The image acquisition module is used to acquire ultrasound vascular images of the patient's liver in real time; the vascular tracking module is used to perform real-time vascular detection on each frame of the image using a trained detection model.

[0033] The motion model of blood vessels is modeled using a Kalman filter;

[0034] The detection bounding boxes obtained by the detection model and the tracking bounding boxes predicted by the motion model are matched using the Hungarian algorithm model, and the motion model is updated based on the matching results.

[0035] The blood vessel extraction module is used to extract the blood vessel motion trajectory based on the tracking box after Kalman filtering.

[0036] As a preferred approach, the detection model is trained using a training module, which includes:

[0037] The pre-training unit is used to pre-train the feature extraction network using a large natural image dataset, enabling the feature extraction network to have basic feature extraction capabilities.

[0038] Dataset building unit, used to build patient-specific liver ultrasound vascular datasets;

[0039] The fine-tuning unit is used to fine-tune the pre-trained feature extraction network based on the liver ultrasound vascular dataset through transfer learning to obtain the trained detection model.

[0040] As a preferred embodiment, the blood vessel tracking module specifically includes:

[0041] The processing unit is used to perform real-time frame-by-frame blood vessel detection on the image using the trained detection model, and to generate a corresponding Kalman filter for each newly appearing detection box. The Kalman filter models the motion model of the blood vessel.

[0042] The input unit is used to predict the position of the tracking box in the next frame image based on the motion model, and input the prediction result into the Hungarian algorithm model.

[0043] At the same time, for the image of the current frame, the detection results of the detection model are also input into the Hungarian algorithm model;

[0044] The matching unit, the Hungarian algorithm model, uses the IoU metric to match the tracking boxes predicted by the motion model with the detection boxes obtained by the detection model;

[0045] For successfully matched tracking boxes, they are marked. If N consecutive frames are successfully matched, the motion model is updated using a Kalman filter and the tracking boxes are predicted, where N≥3.

[0046] For newly appearing detection boxes, initialize new tracking boxes;

[0047] For unmatched tracking boxes, continue tracking for several frames. If no match is found, delete the tracking box.

[0048] As a preferred embodiment, the blood vessel extraction module includes a low-pass filter, which is used to perform low-pass filtering on the tracking box after Kalman filtering.

[0049] This invention discloses a liver ultrasound vascular tracking method and device for estimating the motion of thoracic and abdominal tumors. It tracks the movement of liver vessels using ultrasound, detects vessel positions in real time using a detection model, and establishes a more accurate vascular motion model through a Kalman filter to obtain vascular motion curves. This achieves precise treatment without invasiveness, radiation, or additional markers. It has the following features:

[0050] Beneficial effects:

[0051] First, the present invention uses ultrasound vascular images to extract vascular motion curves, and then obtains reference respiratory motion curves, without requiring any external devices to track respiratory motion or landmarks.

[0052] Secondly, compared with similar methods, this invention is non-invasive and radiation-free, and does not require additional marking points on the body, making the treatment process more comfortable.

[0053] Third, this invention can be used in conjunction with most ultrasound equipment. It only requires providing a stable image sequence with a stable acquisition frequency to achieve real-time modeling of target area motion information, making it widely applicable. Attached Figure Description

[0054] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart of a liver ultrasound vascular tracking method provided in an embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of the liver ultrasound vascular tracking method provided in an embodiment of the present invention.

[0057] Figure 3 This is a rendering of the detection box and tracking box on the ultrasound vascular image provided in an embodiment of the present invention.

[0058] Figure 4 This is a schematic diagram illustrating the training of the detection model provided in an embodiment of the present invention.

[0059] Figure 5 This is a schematic diagram illustrating the detection and tracking of ultrasound vascular images provided in an embodiment of the present invention.

[0060] Figure 6 (a) and (b) are motion curves of the detection box and the tracking box provided in the embodiments of the present invention, respectively.

[0061] Figure 7The motion curve of the tracking frame after low-pass filtering is provided for an embodiment of the present invention.

[0062] Wherein: 1-ultrasonic probe, 2-clamp. Detailed Implementation

[0063] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0064] 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.

[0065] The expression “includes” is an “open-ended” expression, which means that there is a corresponding component and should not be interpreted as excluding additional components.

[0066] To achieve the objectives of this invention, some embodiments of a liver ultrasound vascular tracking method and device for estimating the motion of thoracic and abdominal tumors are provided, such as... Figure 1 As shown, the liver ultrasound vascular tracking method includes:

[0067] Step S1: Real-time acquisition of ultrasound vascular images of the patient's liver;

[0068] Step S2: Use the trained detection model to perform real-time blood vessel detection on each frame of the image;

[0069] The motion model of blood vessels is modeled using a Kalman filter;

[0070] The detection bounding boxes obtained by the detection model and the tracking bounding boxes predicted by the motion model are matched using the Hungarian algorithm model, and the motion model is updated based on the matching results.

[0071] Step S3: Extract the blood vessel motion trajectory based on the tracking box after Kalman filtering, and output the blood vessel motion curve.

[0072] Based on the inventor's creative thinking, it was discovered that the hepatic vascular movement caused by breathing during radiotherapy is highly correlated with the movement of tumors in the chest and abdomen. Therefore, this invention uses hepatic vascular movement as a substitute for target area movement.

[0073] This invention uses ultrasound guidance to track liver blood vessels during radiotherapy and estimates the motion of tumors in the patient's chest and abdomen in real time. Compared with existing technologies, this invention is non-invasive, radiation-free, and requires no additional markers.

[0074] Each step will be explained in detail below.

[0075] like Figure 2 As shown, before radiotherapy, ultrasound probe 1 is fixed to the patient's chest and abdomen to acquire ultrasound vascular images of the patient's liver.

[0076] Furthermore, clamp 2 can be used to fix the ultrasound probe to the patient's chest and abdomen. One end of the clamp is fixed to the treatment bed and rigidly connected to it, while the other end clamps the ultrasound probe.

[0077] In another embodiment, the gripper may also be replaced by a programmable robotic arm, which supports and guides the ultrasonic probe.

[0078] In another embodiment, the clamp can also be directly adhered to the patient's skin surface using patch ultrasound.

[0079] The ultrasonic probes using convex array probe technology are shown in Table 1:

[0080] Table 1 Parameters of the ultrasound probe

[0081]

[0082] In this embodiment, during radiotherapy, the treatment room is filled with high-energy rays. Ultrasound vascular images are transmitted wirelessly to an ultrasound workstation. The transmission method can be WiFi (IEEE 802.11), and the operating frequency band can be selected from 2.4G and 5G. Alternatively, Bluetooth (IEEE 802.15.1) can also be used.

[0083] The transmitted ultrasound vascular images are processed in real time by the doctor's workstation using an NVIDIA graphics processing unit (GPU) with CUDA to accelerate image data processing. Figure 3 As shown. For workstations without a GPU, the processing unit can be replaced by a CPU.

[0084] The above detection model is based on a deep convolutional neural network, and YOLOv5 or YOLOv8 can be used as the detection model, such as... Figure 4 As shown, its training methods include:

[0085] Step A: Pre-train the feature extraction network using a large natural image dataset to enable the feature extraction network to have basic feature extraction capabilities;

[0086] Step B: Construct a patient-specific liver ultrasound vascular dataset;

[0087] Step C: Based on the liver ultrasound vascular dataset, the pre-trained feature extraction network is fine-tuned through transfer learning to obtain the trained detection model.

[0088] This invention first uses a feature extraction network pre-trained on a large natural image dataset to solve the problem of the small quantity and low quality of medical ultrasound images.

[0089] The feature extraction network used has few parameters, and the model parameters are updated rapidly. Only a few hundred patient images need to be collected and labeled for fine-tuning the detection model. The data collection, labeling, and training of this invention can be completed within half an hour, and the number of model parameters is small, with a file size of only 2MB, making it suitable for widespread application.

[0090] The ultrasound vascular image stream acquired by the ultrasound probe is transmitted to the doctor's workstation via a wireless network for display and processing. The detection model detects the position of each blood vessel in the image online and in real time, then calculates the motion characteristics of each blood vessel, and then uses a similarity metric to calculate the probability that the tracking box and the detection box belong to the same target. Finally, an ID identifier can be assigned to each object through data association.

[0091] Specifically, such as Figure 5 As shown, step S2 specifically includes:

[0092] Step S2.1: Use the trained detection model to perform real-time frame-by-frame blood vessel detection on the image, generate a corresponding Kalman filter for each newly appearing detection box, and model the motion model of the blood vessel using the Kalman filter.

[0093] Step S2.2: Predict the position of the tracking box in the next frame image based on the motion model, and input the prediction result into the Hungarian algorithm model.

[0094] At the same time, for the image of the current frame, the detection results of the detection model are also input into the Hungarian algorithm model;

[0095] Step S2.3: The Hungarian algorithm model matches the tracking boxes predicted by the motion model with the detection boxes obtained by the detection model using the IoU metric;

[0096] For successfully matched tracking boxes, they are marked. If three consecutive frames are successfully matched, the motion model is updated using a Kalman filter and the tracking boxes are predicted.

[0097] For newly appearing detection boxes, initialize new tracking boxes;

[0098] For unmatched tracking boxes, continue tracking for several frames. If no match is found, delete the tracking box.

[0099] The modeling of vascular motion state is described in detail below:

[0100] The Kalman filter, a state-optimal estimation method that considers noise and interference, initially ignores noise and interference. It fuses the prior estimate of the previous state (i.e., the prediction of the current state) with the observed data (measurements of the detection boxes) to correct the previous estimate. This process is iterated until the posterior estimate is closer to the true state value. The Kalman filter is described by a state equation providing state information and an observation equation providing position information. The algorithm consists of two parts: prediction and update, with iterative estimation yielding the optimal result.

[0101]

[0102] x w =Ax w-1 +Bu w-1 +ω w-1 (2)

[0103] Z w =(x,y,r,h) T (3)

[0104] Z w =Hx w +V w (4)

[0105] Equation (1) represents the system state x in the Wth frame. w The vector definition of (the estimated value of frame W) provides state estimation for the system, where (x, y) represents the center position of the tracking box, (r, h) represent the aspect ratio and height of the tracking box, respectively, and the other four parameters... This represents the corresponding velocity component.

[0106] In formula (2), A, B, u w-1 These represent the state transition matrix, control input matrix, and system control quantity, respectively, ω. w-1 It is Gaussian process noise with covariance Q.

[0107] Equation (3) represents the system observation value Z in the Wth frame. w The vector definition of (detection value of frame W, target detection box information) provides position information for the system.

[0108] In equation (4), H represents the observation matrix, which is the detection result provided by the detection model, and V w It is Gaussian observation noise with covariance R.

[0109] The prediction of vasomotor status is explained in detail below:

[0110] In frame W, the state transition matrix A and the optimal estimate from the previous frame are used... Predict the state of the current W frame, and simultaneously observe that state to obtain the predicted value Z.w Using Kalman gain K to Z w and By performing fusion, the optimal estimate for the current moment can be obtained. It also serves as the predicted value of the prior state at the next moment. Input.

[0111]

[0112]

[0113] Equations (5) and (6) are the state prediction equation and the covariance prediction equation, respectively, where This is the state prediction value for the Wth frame. This is the optimal estimate for the W-1th time.

[0114] The specific details regarding the vasomotor status update are as follows:

[0115]

[0116]

[0117]

[0118] Equation (7) is the Kalman gain update equation, where K w For Kalman gain, The prior estimate of the covariance of the Wth frame is, i.e. The covariance.

[0119] Equation (8) is the posterior state estimation update equation, where This is the optimal estimate for the Wth frame.

[0120] Equation (9) is the posterior estimation covariance equation, where I is the identity matrix.

[0121] In this invention, the similarity measure between the tracking box and the detection box is the intersection-union ratio (IUGR), which is used to calculate the matching degree of the motion model. The tracking box (bbox) predicted by the motion model is then used to... t and the detection bounding box obtained by the detection model d The cost matrix is ​​obtained by performing IoU measurement on each step, and the calculation formula is shown below:

[0122]

[0123] in:

[0124] σ is the matching threshold, which can be adjusted to control the degree of matching between the tracking box and the detection box. In this embodiment, σ is set to 0.3.

[0125] In the Hungarian algorithm, all values ​​less than σ in the cost matrix are set to 0. Then, the cost matrix is ​​used as input to the Hungarian algorithm model to obtain the matching results of the detection box and the tracking box.

[0126] Figure 6 (a) and (b) are the motion curves of the detection box and the tracking box of the present invention, respectively. The motion model generated by the tracking box after Kalman filtering is more stable and can continue to give prediction results even when the detection fails.

[0127] Specifically, step S3 further involves: performing low-pass filtering on the tracking box after Kalman filtering to extract the blood vessel motion trajectory.

[0128] This invention applies a low-pass filter to the tracking frame after Kalman filtering to remove noise caused by the tracking frame's own jitter and vascular movement caused by heartbeat. The respiratory and heart rate in a resting state are approximately 0.2 Hz and 1.2 Hz, respectively. Using a low-pass filter can remove most of the high-frequency noise in the signal, resulting in a smoother tracking frame motion curve and effectively reducing glitches. Figure 7 As shown.

[0129] In summary, using a Kalman filter to model the motion of blood vessels and updating the model in real time during treatment results in a more accurate motion model. Furthermore, it can predict the reappearance of blood vessels even when their disappearance in the image is caused by patient movement, thus solving the problem of sudden blood vessel disappearance.

[0130] Furthermore, embodiments of the present invention also disclose a liver ultrasound vascular tracking device for estimating the motion of thoracic and abdominal tumors, comprising:

[0131] The image acquisition module is used to acquire ultrasound vascular images of the patient's liver in real time; the vascular tracking module is used to perform real-time vascular detection on each frame of the image using a trained detection model.

[0132] The motion model of blood vessels is modeled using a Kalman filter;

[0133] The detection bounding boxes obtained by the detection model and the tracking bounding boxes predicted by the motion model are matched using the Hungarian algorithm model, and the motion model is updated based on the matching results.

[0134] The blood vessel extraction module is used to extract the blood vessel motion trajectory based on the tracking box after Kalman filtering.

[0135] Furthermore, the detection model is trained using a training module, which includes:

[0136] The pre-training unit is used to pre-train the feature extraction network using a large natural image dataset, enabling the feature extraction network to have basic feature extraction capabilities.

[0137] Dataset building unit, used to build patient-specific liver ultrasound vascular datasets;

[0138] The fine-tuning unit is used to fine-tune the pre-trained feature extraction network based on the liver ultrasound vascular dataset through transfer learning to obtain the trained detection model.

[0139] Furthermore, the blood vessel tracking module specifically includes:

[0140] The processing unit is used to perform real-time frame-by-frame blood vessel detection on the image using the trained detection model, and to generate a corresponding Kalman filter for each newly appearing detection box. The Kalman filter models the motion model of the blood vessel.

[0141] The input unit is used to predict the position of the tracking box in the next frame image based on the motion model, and input the prediction result into the Hungarian algorithm model.

[0142] At the same time, for the image of the current frame, the detection results of the detection model are also input into the Hungarian algorithm model;

[0143] The matching unit, the Hungarian algorithm model, uses the IoU metric to match the tracking boxes predicted by the motion model with the detection boxes obtained by the detection model;

[0144] For successfully matched tracking boxes, they are marked. If three consecutive frames are successfully matched, the motion model is updated using a Kalman filter and the tracking boxes are predicted.

[0145] For newly appearing detection boxes, initialize new tracking boxes;

[0146] For unmatched tracking boxes, continue tracking for several frames. If no match is found, delete the tracking box.

[0147] Furthermore, the blood vessel extraction module includes a low-pass filter, which is used to perform low-pass filtering on the tracking box after Kalman filtering.

[0148] It is worth noting that the technical solution of the liver ultrasound vascular tracking device is similar to that of the liver ultrasound vascular tracking method, and will not be described in detail here.

[0149] This invention discloses a liver ultrasound vascular tracking method and device for estimating the motion of thoracic and abdominal tumors. It tracks the movement of liver vessels using ultrasound, detects vessel positions in real time using a detection model, and establishes a more accurate vascular motion model through a Kalman filter to obtain vascular motion curves. This achieves precise treatment without invasiveness, radiation, or additional markers. It has the following features:

[0150] Beneficial effects:

[0151] Fourth, this invention utilizes ultrasound vascular images to extract vascular motion curves, thereby obtaining reference respiratory motion curves, without requiring any external devices to track respiratory motion or landmarks.

[0152] Fifth, compared with similar methods, this invention is non-invasive and radiation-free, and does not require additional marking points on the body, making the treatment process more comfortable.

[0153] Sixth, this invention can be used in conjunction with most ultrasound equipment. It only requires providing an image sequence with a stable acquisition frequency to achieve real-time modeling of target area motion information, making it widely applicable.

[0154] It should be understood that the various techniques described herein can be implemented in combination with hardware or software, or a combination thereof. Thus, the methods and apparatus of the present invention, or certain aspects or portions thereof, can take the form of program code (i.e., instructions) embedded in a tangible medium, such as a floppy disk, CD-ROM, hard disk, or any other machine-readable storage medium, wherein when the program is loaded into and executed by a machine such as a computer, that machine becomes an apparatus for practicing the present invention.

[0155] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it. They should not be used to limit the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A liver ultrasound blood vessel tracking method for thoraco-abdominal tumor motion estimation, characterized by, Correlation between liver blood vessel movement caused by patient respiration during radiotherapy and chest and abdominal tumor movement, using liver blood vessel movement as a target region movement surrogate for chest and abdominal tumor movement estimation, comprising: Step S1: Real-time acquisition of ultrasound blood vessel images of the patient's liver; Step S2: Real-time blood vessel detection of each frame of image using the trained detection model; Modeling the motion model of the blood vessels as the target region movement surrogate using a Kalman filter; Matching the detection box obtained by the detection model and the tracking box predicted by the motion model of the blood vessels as the target region movement surrogate using a Hungarian algorithm model, and updating the motion model according to the matching result; Step S3: Extracting the blood vessel motion trajectory as the target region movement surrogate according to the tracking box after Kalman filtering to realize chest and abdominal tumor motion estimation; The step S2 specifically comprises: Step S2.1: Real-time frame-by-frame blood vessel detection of the image using the trained detection model, generating a corresponding Kalman filter for each newly appearing detection box, and modeling the motion model of the blood vessels by the Kalman filter; Step S2.2: Predicting the position of the tracking box in the next frame of image according to the motion model, and inputting the prediction result into a Hungarian algorithm model, At the same time, the detection result of the detection model is also input into the Hungarian algorithm model for the current frame of image; Step S2.3: The Hungarian algorithm model matches the tracking box predicted by the motion model and the detection box obtained by the detection model through IoU measurement; For the tracking box that matches successfully, mark it, if it matches successfully for consecutive N frames of image, update the motion model using the Kalman filter and predict the tracking box, where N≥3; For the newly appearing detection box, initialize a new tracking box; For the unmatched tracking box, continue tracking for several frames of image, if it is always not matched with the detection box, delete the tracking box.

2. The liver ultrasound blood vessel tracking method of claim 1, wherein, The training method of the detection model comprises: Step A: Pre-training the feature extraction network using a large natural image dataset to enable the feature extraction network to have basic feature extraction capability; Step B: Constructing a patient-specific liver ultrasound blood vessel dataset; Step C: Fine-tuning the pre-trained feature extraction network through transfer learning based on the liver ultrasound blood vessel dataset to obtain the trained detection model.

3. The liver ultrasound blood vessel tracking method of claim 1, wherein, The step S3 further comprises: low-pass filtering the tracking box after Kalman filtering to extract the blood vessel motion trajectory.

4. Liver ultrasound blood vessel tracking device for thoraco-abdominal tumor motion estimation, characterized by, Correlation between liver blood vessel movement caused by patient respiration during radiotherapy and chest and abdominal tumor movement, using liver blood vessel movement as a target region movement surrogate for chest and abdominal tumor movement estimation, comprising: An image acquisition module for real-time acquisition of ultrasound blood vessel images of the patient's liver; a blood vessel tracking module for real-time blood vessel detection of each frame of image using a trained detection model; Modeling the motion model of the blood vessels as the target region movement surrogate using a Kalman filter; Matching the detection box obtained by the detection model and the tracking box predicted by the motion model of the blood vessels as the target region movement surrogate using a Hungarian algorithm model, and updating the motion model according to the matching result; A blood vessel extraction module is configured to extract a blood vessel motion trajectory of the blood vessel as a motion surrogate of a target region according to the tracking box after Kalman filtering, so as to realize thoraco-abdominal tumor motion estimation. The blood vessel tracking module specifically comprises: A processing unit is configured to perform real-time frame-by-frame blood vessel detection on images by using a trained detection model, and generate a corresponding Kalman filter for each newly appearing detection box, wherein the Kalman filter models a motion model of the blood vessel; An input unit is configured to predict a position of a tracking box in a next frame of image according to the motion model, and input a prediction result into a Hungarian algorithm model, Meanwhile, for the image of the current frame, a detection result of the detection model is also input into the Hungarian algorithm model; A matching unit is configured to match the tracking box predicted by the motion model and the detection box obtained by the detection model by using an IoU metric. For a tracking box that is successfully matched, a mark is performed, and if the tracking box is successfully matched for consecutive N frames of images, the motion model is updated by using a Kalman filter and the tracking box is predicted, wherein N is greater than or equal to 3. For a newly appearing detection box, a new tracking box is initialized. For a tracking box that is not matched, the tracking box is continuously tracked for several frames of images, and if the tracking box is not matched with a detection box all the time, the tracking box is deleted.

5. The liver ultrasound blood vessel tracking device of claim 4, wherein, The detection model is trained by using a training module, and the training module comprises: A pre-training unit is configured to pre-train a feature extraction network by using a large natural image dataset, so that the feature extraction network has a basic feature extraction capability; A dataset construction unit is configured to construct a patient-specific liver ultrasound blood vessel dataset; A fine-tuning unit is configured to fine-tune the pre-trained feature extraction network by using a liver ultrasound blood vessel dataset through transfer learning, so as to obtain a trained detection model.

6. The liver ultrasound blood vessel tracking device of claim 4, wherein, The blood vessel extraction module comprises a low-pass filter configured to perform low-pass filtering on the tracking box after Kalman filtering.

Images