Three-dimensional dynamic path planning algorithm for liver intervention puncture operation

Abstract

The invention belongs to the technical field of DAS, discloses a three-dimensional dynamic path planning algorithm for a liver interventional puncture operation, and aims to solve the problem that the complexity of an operation environment is increased due to abdominal movement caused by respiratory movement change of a patient in a liver interventional operation. According to the method, collision punishment is calculated through a respiratory movement model to avoid the influence of abdominal movement on path planning, focus positioning is actively updated through a reward function according to observation data collected in an operation, and dynamic active tracking of the accurate position of a target spot is achieved. The three-dimensional optimal path from puncture to the target spot is sought in a dynamic environment, accurate dynamic surgery simulation is achieved, and the most safe, accurate and effective treatment purpose aiming at the treatment target spot is achieved.

Description

Technical Field

[0001] This invention relates to the field of DSA technology, and in particular to a three-dimensional dynamic path planning algorithm for liver interventional puncture surgery. Background Technology

[0002] Interventional liver therapy, utilizing percutaneous, transvascular, and transcirculatory approaches, employs methods such as embolization, ablation, internal irradiation, and shunt establishment. It has become an important clinical local treatment option for patients with advanced cancer and cirrhosis, effectively reducing tumor burden, lowering portal pressure, and improving clinical prognosis. However, clinical practice of interventional liver therapy still faces several bottlenecks requiring breakthroughs: First, accuracy is lacking. For some poorly vascularized lesions, localization is difficult, and the feeding arteries are hard to identify. Respiratory movements also affect the accuracy, necessitating repeated adjustments to the instrument's angle and direction to reach the lesion and blood vessels. Second, intelligence is insufficient. 2D-DSA is affected by overlapping vascular images, rapid contrast agent clearance, and respiratory movements, making the identification of the lesion's feeding vessels and catheter insertion path primarily dependent on personal experience. During puncture, manual measurement is required to plan the puncture path. Therefore, it is necessary to research a three-dimensional dynamic path planning algorithm for liver interventional puncture surgery. Summary of the Invention

[0003] The purpose of this invention is to provide a three-dimensional dynamic path planning algorithm for liver interventional puncture surgery, which can accurately achieve dynamic simulation of the surgery.

[0004] The above-mentioned technical objective of this invention is achieved through the following technical solution: a three-dimensional dynamic path planning algorithm for liver interventional puncture surgery, assuming the starting point of the path is... The target point is The path with the lowest energy consumption from the starting point to the destination is Assume there are K obstacles in the environment. The total energy function of a path point consists of a distance function and a collision penalty function, i.e. , of which E l Represents the distance traveled along the entire path. E C Here is the collision penalty function: In the formula, gth k (·) The collision detection model represents the Kth obstacle; to avoid the influence of abdominal movement on path planning, a respiratory motion model is used to calculate the collision penalty, assuming that the motion voxel of the puncture robot when the operating mechanism reaches the point is... α i ,use α i Updating obstacle positions using prior knowledge yields the collision penalty function for motion priors: The collision penalty function Ec mentioned above is dynamically adjusted according to the motion voxels when the puncture robot reaches each node in the path during the movement process, thereby realizing the dynamic planning of the 3D path of the puncture robot.

[0005] A further feature of the present invention is that it includes a reward function, which is used to actively update the lesion location based on the collected observation data during surgery, thereby achieving dynamic and active tracking of the precise location of the target.

[0006] A further setting of the present invention is: the reward function is: ; in, and w These represent the target's position and orientation relative to the tracker, respectively. d The distance constant between the tracker and the target is defined. c Represents the regularization factor.

[0007] A further provision of the present invention is that the reward function includes: assuming that the tracking system and the observation system are dynamically linear, and that the system noise and observation noise conform to a Gaussian distribution with a mean of 0, the observed values ​​can be expressed as... , H For observing system parameters, This represents observation noise. According to the recursive process of the Kalman filter, t The state at time is S t This can be represented as, where, m t Indicates system input, F and B For system parameters, This indicates system noise.

[0008] A further provision of the present invention is that the reward function further includes: calculating the relevant error matrix. and Kalman gain ,in V Given the covariance matrix of the observation error, the corrected Kalman gain can be obtained by iteratively updating the Kalman gain. t The best state at any moment , ; Assumption t The action at a given moment is The compensation function is expressed as r t The optimal state of the Kalman filter can be obtained. value function ,in Representing cumulative compensation, the action value function can be expressed as: The relative importance of actions can be represented by an advantage function. To measure.

[0009] A further provision of this invention is: constructing an active tracking deep learning network structure, training an active tracker for abdominal movements, the active tracker including an observation encoder, a sequence encoder, and an actor-critic network, and using a fusion function. f s （·） Obtain encoding with timing information This encoding is simultaneously input into both the critic network and the actor network to calculate the expected reward and action decision, the reward value function. Action decision strategy function .

[0010] A further provision of the present invention is that, based on the actor-critic algorithm, the policy function can be obtained. and value function The update gradient equation is as follows: ; in, , m Represents learning efficiency. H（·） For information entropy regularization, These are the parameters obtained in the previous step.

[0011] A further provision of this invention is that, during the training process of the active tracking deep learning network structure, the network parameters are updated using an update gradient equation. During tracking, the optimal state is first recursively corrected using a Kalman filter, and then the optimal action of the state is calculated using a reinforcement network. To adjust the tracker to achieve active tracking of the target.

[0012] The beneficial effects of this invention are: Preoperative planning of the surgical path and simulation of the operation process helps to improve surgical accuracy and save surgical time. It is an important step in the "one-stop" liver interventional surgery that integrates intelligence and precision. In the existing technology, the use of surgical navigation technology to perform transvascular implantation surgery has made some progress, but the following reasons limit its clinical use: (1) Most general surgical navigation technologies only use the patient's static preoperative images; (2) Intraoperative navigation only relies on the patient's preoperative images and cannot utilize the image information collected by the patient during the operation.

[0013] In liver interventional surgery, changes in the patient's respiratory movements can lead to abdominal movements, increasing the complexity of the surgical environment. This invention utilizes a respiratory motion model to calculate collision penalties to avoid the impact of abdominal movements on path planning. Using a reward function, it actively updates lesion localization based on observation data collected during surgery, achieving dynamic and active tracking of the target's precise location. By seeking the optimal three-dimensional path to the target in a dynamic environment, it achieves precise dynamic surgical simulation, resulting in the safest and most accurate treatment targeting the therapeutic target. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the active tracking deep learning network structure for the target region of this invention. Detailed Implementation

[0015] A three-dimensional dynamic path planning algorithm for liver interventional puncture surgery, characterized in that: the starting point of the path is assumed to be... The target point is The path with the lowest energy consumption from the starting point to the destination is Assume there are K obstacles in the environment. The total energy function of a path point consists of a distance function and a collision penalty function, i.e. , of which E l Represents the distance traveled along the entire path. E C Here is the collision penalty function: In the formula, gth k (·) The collision detection model represents the Kth obstacle; to avoid the influence of abdominal movement on path planning, a respiratory motion model is used to calculate the collision penalty, assuming that the motion voxel of the puncture robot when the operating mechanism reaches the point is... α i ,use α i Updating obstacle positions using prior knowledge yields the collision penalty function for motion priors: The collision penalty function Ec mentioned above is dynamically adjusted according to the motion voxels when the puncture robot reaches each node in the path during the movement process, thereby realizing the dynamic planning of the 3D path of the puncture robot.

[0016] The three-dimensional dynamic path planning algorithm for liver interventional puncture surgery includes a reward function. This reward function is used during the procedure to actively update the lesion localization based on the collected observation data, achieving precise dynamic and active tracking of the target location. Assuming the tracking and observation systems are dynamically linear, and the system noise and observation noise follow a Gaussian distribution with a mean of 0, the observed values ​​can be represented as... , H For observing system parameters, This represents observation noise. According to the recursive process of the Kalman filter, t The state at time is S t This can be represented as, where, m t Indicates system input, F and B For system parameters, This indicates system noise.

[0017] Calculate the correlation error matrix. and Kalman gain ,in V Given the covariance matrix of the observation error, the corrected Kalman gain can be obtained by iteratively updating the Kalman gain. t The best state at any moment , ; Assumption t The action at a given moment is The compensation function is expressed as r t The optimal state of the Kalman filter can be obtained. value function ,in Representing cumulative compensation, the action value function can be expressed as: The relative importance of actions can be represented by an advantage function. To measure.

[0018] For active target tracking tasks, a reward function is designed to guide the agent's learning: a reward is given when the target approaches the desired position, and a penalty is given when it moves away. The reward function is as follows: ; in, and w These represent the target's position and orientation relative to the tracker, respectively. d The distance constant between the tracker and the target is defined. c Represents the regularization factor.

[0019] An active tracking deep learning network structure was constructed to train an active tracker for abdominal movements. The active tracker includes an observation encoder, a sequence encoder, and an actor-critic network, which are then fused using a fusion function. f s （·） Obtain encoding with timing information This encoding is simultaneously input into both the critic network and the actor network to calculate the expected reward and action decision, the reward value function. Action decision strategy function According to the actor-critic algorithm, the policy function can be obtained. and value function The update gradient equation is as follows: ; in, , m Represents learning efficiency. H（·） For information entropy regularization, These are the parameters obtained in the previous step. During the training of the active tracking deep learning network structure, the network parameters are updated using the gradient update equation. During tracking, the optimal state is first recursively corrected using a Kalman filter, and then the optimal action for the state is calculated using a reinforcement network. To adjust the tracker to achieve active tracking of the target.

Claims

1. A three-dimensional dynamic path planning algorithm for liver interventional puncture surgery, characterized in that: Assuming the starting point of the path is The target point is The path with the lowest energy consumption from the starting point to the destination is Assume there are K obstacles in the environment. The total energy function of a path point consists of a distance function and a collision penalty function, i.e. E l Represents the distance traveled along the entire path. E C Here is the collision penalty function: In the formula, gθ k (·) The collision detection model represents the Kth obstacle; to avoid the influence of abdominal movement on path planning, a respiratory motion model is used to calculate the collision penalty, assuming that the motion voxel of the puncture robot when the operating mechanism reaches the point is... α i ,use α i Updating obstacle positions using prior knowledge yields the collision penalty function for motion priors: The collision penalty function Ec mentioned above is dynamically adjusted according to the motion voxels when the puncture robot reaches each node in the path during the movement process, thereby realizing the dynamic planning of the 3D path of the puncture robot.

2. The three-dimensional dynamic path planning algorithm for liver interventional puncture surgery according to claim 1, characterized in that: This includes a reward function, which is used during surgery to actively update the lesion location based on the collected observation data, thereby achieving dynamic and active tracking of the precise target location.

3. The three-dimensional dynamic path planning algorithm for liver interventional puncture surgery according to claim 2, characterized in that: The reward function is: ; in, and w These represent the target's position and orientation relative to the tracker, respectively. d The distance constant between the tracker and the target is defined. c Represents the regularization factor.

4. The three-dimensional dynamic path planning algorithm for liver interventional puncture surgery according to claim 3, characterized in that: The reward function includes: assuming the tracking and observation systems are dynamically linear, and the system noise and observation noise follow a Gaussian distribution with a mean of 0, the observations can be expressed as... , H For observing system parameters, This represents observation noise. According to the recursive process of the Kalman filter, t The state at time is S t This can be represented as, where, μ t Indicates system input, F and B For system parameters, This indicates system noise.

5. The three-dimensional dynamic path planning algorithm for liver interventional puncture surgery according to claim 4, characterized in that: The reward function also includes: calculating the correlation error matrix. and Kalman gain ,in V Given the covariance matrix of the observation error, the corrected Kalman gain can be obtained by iteratively updating the Kalman gain. t The best state at any moment , ; Assumption t The action at a given moment is The compensation function is expressed as r t The optimal state of the Kalman filter can be obtained. value function ,in Representing cumulative compensation, the action value function can be expressed as: The relative importance of actions can be represented by an advantage function. To measure.

6. The three-dimensional dynamic path planning algorithm for liver interventional puncture surgery according to claim 5, characterized in that: An active tracking deep learning network structure was constructed to train an active tracker for abdominal movements. The active tracker includes an observation encoder, a sequence encoder, and an actor-critic network, which are then fused using a fusion function. f s （·） Obtain encoding with timing information This encoding is simultaneously input into both the critic network and the actor network to calculate the expected reward and action decision, the reward value function. Action decision strategy function .

7. The three-dimensional dynamic path planning algorithm for liver interventional puncture surgery according to claim 6, characterized in that: Based on the actor-critic algorithm, the policy function can be obtained. and value function The update gradient equation is as follows: ; in, , μ Represents learning efficiency. H（·） For information entropy regularization, These are the parameters obtained in the previous step.

8. The three-dimensional dynamic path planning algorithm for liver interventional puncture surgery according to claim 7, characterized in that: During the training of the active tracking deep learning network structure, the network parameters are updated using an update gradient equation. During tracking, the optimal state is first recursively corrected using a Kalman filter, and then the optimal action for the state is calculated using a reinforcement network. To adjust the tracker to achieve active target tracking.

Images