A method and system for predicting soft tissue deformation during a puncture procedure based on machine learning

Abstract

The application discloses a kind of soft tissue deformation prediction methods and systems based on machine learning of puncture process, the pose data of puncture needle, force sense signal and ultrasound image sequence in puncture process are collected, and time synchronization processing is carried out to multimodal data, to construct the feature sequence representing tissue dynamics characteristics, then the feature sequence is input into the prediction model with time series modeling ability trained, so that it outputs the deformation prediction result of soft tissue in the future preset time window in the neighborhood of needle tip, the deformation prediction result includes displacement field, key point displacement or target area offset, for reflecting the dynamic change trend of soft tissue in short time scale.The application can estimate the future motion state of soft tissue in advance in puncture process, improve the accuracy of soft tissue motion trend judgment in puncture process, enhance the safety and stability of puncture operation.

Description

Technical Field

[0001] This invention relates to the field of medical intervention technology, specifically to a method for predicting soft tissue deformation during puncture based on machine learning, and more particularly to a technical solution that uses multimodal time-series data to construct a prediction model to predict the future deformation trend of soft tissue near the needle tip during puncture operations. Background Technology

[0002] In ultrasound-guided percutaneous puncture, tumor ablation, biopsy sampling and other medical interventional procedures, soft tissues are affected by a variety of factors during the acupuncture process, such as breathing, heartbeat, patient movement, ultrasound probe contact pressure and needle advancement, which cause significant deformation. This deformation has nonlinear, non-uniform and time-varying characteristics, which causes the target tissue position and puncture path to constantly deviate in actual operation.

[0003] Existing puncture navigation systems mainly rely on real-time ultrasound images, force information, or pose information to estimate the current deformation state of tissues. However, such methods can usually only reflect the instantaneous morphology of tissues at a certain moment and cannot effectively predict the dynamic changes of tissues over a future period of time.

[0004] Because the deformation of soft tissue during puncture has a time delay and cumulative effect, it is often difficult to accurately determine the future direction and degree of tissue movement by relying solely on the current image or mechanical information. This makes the puncture needle prone to deviation due to tissue displacement, affecting the accuracy and safety of interventional surgery.

[0005] On the other hand, although existing studies have developed soft tissue deformation analysis methods based on biomechanical models, such as using finite element models (FEM) to simulate the acupuncture process, these methods involve large computational loads and are highly dependent on model parameters, making them difficult to apply in real time during surgery and unable to accurately predict the future dynamic deformation trend of tissues during puncture.

[0006] For example, application CN110400283A (Real-time and accurate soft tissue deformation prediction) discloses a method for predicting soft tissue deformation. It first calculates the instantaneous state of soft tissue under a given external force based on biomechanical or other tissue-related physical models, and then uses artificial neural networks in machine learning to predict the position of soft tissue voxels or grid nodes at subsequent times based on the instantaneous state. This allows for obtaining soft tissue deformation results at different time points in a short time, which can improve the accuracy and real-time performance of soft tissue imaging and interactive simulation. However, this method is highly dependent on the pre-built biomechanical model and its parameter calibration. It is mainly aimed at solving deformation in general soft tissue scenarios and does not introduce multimodal temporal information such as puncture needle pose, force signal and ultrasound image for the specific operation of puncture. It is difficult to reflect the local dynamic response of soft tissue in the needle tip neighborhood and the target area offset trend during acupuncture in a timely manner. Summary of the Invention

[0007] This invention aims to overcome the problems in existing puncture procedures, such as the susceptibility of soft tissue to complex dynamic deformation due to factors such as breathing, probe pressure, and needle squeezing, the inability of real-time imaging to reflect future tissue movement trends, and the lack of effective prediction methods that make puncture deviation difficult to avoid. It proposes a machine learning-based method for predicting soft tissue deformation during the puncture process.

[0008] This method constructs multimodal temporal features that characterize the dynamic response of soft tissue by simultaneously acquiring the puncture needle pose, force signals, and ultrasound image sequences. A trained deep learning prediction model is then used to predict the displacement field, key point displacement, or target area offset of the soft tissue in the needle tip neighborhood within a preset future time window, thereby obtaining the future deformation trend of the soft tissue. With this predictive capability, this invention can anticipate the dynamic changes of tissue during puncture, significantly improving the accuracy of intraoperative judgment of soft tissue movement trends, risk avoidance, and navigation guidance. This provides a more reliable basis for predicting tissue behavior in intelligent puncture robots, interventional navigation systems, and real-time surgical assistance.

[0009] A machine learning-based method for predicting soft tissue deformation during puncture, comprising the following steps: S1. Data Acquisition and Synchronization: Acquire the pose data, force signals and ultrasound image sequences of the puncture needle during the puncture process, and align the pose data, force signals and ultrasound image sequences at different sampling frequencies with timestamps to generate a synchronized dataset. S2. Feature sequence construction: Based on the synchronous dataset, a feature sequence containing information from multiple consecutive moments is constructed in chronological order. The feature sequence is used to characterize the dynamic changes of the soft tissue near the needle tip during the advancement of the puncture needle, and includes at least one of pose change features, force sensation change features, and image features. S3. Predictive Model Inference: Input the feature sequence into the trained machine learning prediction model, so that the prediction model outputs the deformation prediction result of the soft tissue in the vicinity of the needle tip within a future preset time window based on the feature sequence. S4. Output Deformation Prediction: Based on the deformation prediction results, generate at least one of the displacement field of the soft tissue in the vicinity of the needle tip, the displacement of key points, or the offset of the target area, to characterize the future deformation trend of the soft tissue in a short time scale.

[0010] Preferably, the pose change features in step S2 include at least one of the following: spatial position, advancement speed, advancement direction, and angle changes of the puncture needle at continuous time intervals.

[0011] Preferably, the force sensation change characteristics in step S2 are obtained by performing time-domain differential processing on the force sensation signal, which is used to characterize the local deformation trend caused by the compression of soft tissue by the puncture needle.

[0012] Preferably, the image features in step S2 are obtained by performing convolutional neural network (CNN) analysis on the ultrasound image sequence, and the image features include at least one of tissue boundary location, local texture changes, or optical flow displacement.

[0013] Preferably, the prediction model in step S3 includes a recurrent neural network, a gated recurrent unit network, a long short-term memory network, a temporal Transformer, or an equivalent time series modeling network structure.

[0014] Preferably, in step S3, the prediction model introduces physical constraints to limit the future deformation prediction results to meet the mechanical properties of soft tissue. The physical constraints include at least one of tissue elasticity constraints, continuity constraints, or deformation smoothness constraints.

[0015] Preferably, the prediction model in step S3 is obtained by training on training data that reflects the dynamic response behavior of soft tissue during puncture. The training data includes experimental data, simulation data generated based on tissue mechanics models, or a combination of both.

[0016] Preferably, the deformation prediction results in step S4 include a prediction sequence of multiple future time points, which are used to construct the temporal deformation trajectory of soft tissue.

[0017] Preferably, in step S4, the displacement field is used to characterize the future displacement vectors of multiple spatial locations in the vicinity of the needle tip, and the key point displacement is used to characterize the future spatial offset of the target point, envelope, or risk structure.

[0018] The present invention also provides a machine learning-based soft tissue deformation prediction system for puncture procedures, comprising: Position acquisition unit: used to acquire puncture needle position data and transmit the position data to the signal synchronization module; Force sensing acquisition unit: used to acquire force sensing signals and transmit them to the signal synchronization module; Image acquisition unit: used to acquire ultrasound image sequences and transmit the ultrasound image sequences to the signal synchronization module; Signal synchronization module: used to receive pose data, force signals and ultrasound image sequences, align the pose data, force signals and ultrasound image sequences with timestamps, and transmit the aligned data to the feature extraction module; Feature extraction module: used to receive aligned data, construct feature sequences, and transmit the feature sequences to the machine learning prediction module; Machine learning prediction module: used to receive and process feature sequences, generate prediction results, and transmit the prediction results to the prediction result generation module; Prediction result generation module: Used to receive prediction results, process the prediction results and generate soft tissue displacement field, key point displacement or target area offset.

[0019] In its application, this invention first acquires multimodal time-series data, including needle pose, force signals, and ultrasound image sequences, during the puncture procedure. A data synchronization mechanism establishes the temporal correspondence between these signals. Subsequently, feature sequences reflecting soft tissue dynamics changes are extracted from the synchronized multimodal data. These features are then input into a trained machine learning prediction model, which outputs a deformation prediction result for the soft tissue near the needle tip within a preset time window, including displacement fields, key point displacements, or target area offsets. Based on this prediction, the system can anticipate future soft tissue deformation trends, improving the accuracy of judging soft tissue movement trends during puncture. This provides a predictive basis for intraoperative navigation or risk avoidance modules, enhancing the reliability and safety of the puncture procedure. The method of this invention does not require precise construction of a tissue mechanics model, can be rapidly deployed in different tissue environments, and exhibits good real-time performance and versatility.

[0020] In a preferred embodiment of the present invention, the preset time window can be set to approximately 100ms–500ms according to the real-time requirements of clinical puncture. This time range can cover the short-term dynamic changes of the target tissue under the influence of factors such as respiration, probe pressure changes, and micro-advancement of the puncture needle, ensuring that the prediction results are both forward-looking and real-time available. In practical applications, the time window can also be adjusted according to different tissue types, imaging frame rates, or robot control cycles.

[0021] Regarding training data preparation, the training data can be obtained in various ways, including but not limited to: (1) Data collection: Dynamic images of soft tissue during actual puncture were recorded using ultrasound imaging equipment, and combined with the position of the robotic arm and the force sensor data, the tissue boundaries, key point positions and displacement changes over time were marked manually or by semi-automatic algorithms. (2) Simulation data generation: By constructing a soft tissue finite element model or a rapid simulation model based on tissue mechanical parameters, high-precision displacement fields or key point motion trajectories are automatically generated under virtual puncture conditions. (3) Combination of experimental data and simulation data: used to improve the generalization ability of the model under different tissue conditions and noise levels.

[0022] This invention, by fusing multimodal temporal information such as pose, force sensation, and ultrasound images acquired during the puncture process, and utilizing a deep learning prediction model with time series modeling capabilities, achieves for the first time the ability to predict the future deformation trend of soft tissue in real time. Compared with existing methods that can only reflect the current deformation state of tissue, this invention can predict the dynamic displacement of tissue under the influence of factors such as respiration, probe pressure, and needle compression, significantly reducing the risk of deviation in the puncture path caused by tissue drift. Furthermore, this invention does not rely on complex finite element models or precise tissue parameter calibration, exhibiting higher predictive stability and applicability. It can improve the accuracy of judging the movement trend of soft tissue during puncture, the reliability of intraoperative navigation, and clinical safety, providing crucial soft tissue behavior prediction capabilities for intelligent puncture robots and interventional treatment devices. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall process of a machine learning-based method for predicting soft tissue deformation during puncture. Figure 2 This is a schematic diagram illustrating multimodal data synchronization and feature construction for a machine learning-based method for predicting soft tissue deformation during puncture. Figure 3 This is a schematic diagram of the prediction model structure for a machine learning-based method for predicting soft tissue deformation during puncture. Figure 4 This is a schematic diagram of the training process for a machine learning-based method for predicting soft tissue deformation during puncture. Figure 5 This is a schematic diagram of the soft tissue deformation prediction results of a machine learning-based soft tissue deformation prediction method during puncture. Figure 6 This is a schematic diagram of the structure and data flow of a machine learning-based soft tissue deformation prediction system for puncture procedures. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example 1

[0025] (1) Overall methodology and process like Figure 1 As shown, this embodiment provides a machine learning-based method for predicting soft tissue deformation during puncture, which includes four stages: signal acquisition and synchronization, feature construction, deformation prediction, and prediction result output.

[0026] During the actual puncture procedure, the system first acquires the pose data of the puncture needle in real time from the robotic arm or positioning device, collects the force signal when the puncture needle contacts the tissue from the force sensor, and obtains the ultrasound image sequence corresponding to the puncture area from the ultrasound imaging device. After acquisition, the aforementioned multimodal data is transmitted to the signal synchronization module for processing. Using a unified timestamp alignment strategy, data from different sampling frequencies are resampled to a consistent timeline, thus forming a synchronized dataset for subsequent analysis.

[0027] The synchronized data is sequentially input into the feature extraction module. This module constructs time-series features characterizing tissue dynamics based on the changing trends of different modalities in the synchronized data. Subsequently, the constructed feature sequences are input into a trained machine learning prediction model, which outputs predicted soft tissue deformation within a preset future time window. The prediction result generation module structures and organizes the model output data to obtain deformation information such as displacement field, key point displacement, and target area offset, and outputs the relevant results to the intraoperative navigation interface or the higher-level control system.

[0028] (2) Multimodal data synchronization and feature construction like Figure 2 As shown, this embodiment acquires multimodal input data from three channels: Position data: Spatial position, propulsion direction, propulsion speed and attitude angle information of the puncture needle tip in the robot base coordinate system; Force signal: The sequence of contact force changes over time when a needle tip contacts soft tissue; Ultrasound image sequence: A series of two-dimensional ultrasound image frames acquired by an imaging device, used to describe the movement and deformation of tissue in the imaging plane.

[0029] The sampling frequencies of the three types of data differ, so the signal synchronization module performs unified timestamp alignment processing on them. Specifically, the system establishes a unified timeline based on the image frame timestamp and performs linear or higher-order interpolation on the pose data and force data to ensure that the three quantities at each time point have corresponding data.

[0030] After obtaining the time-aligned synchronized dataset, the feature extraction module constructs the following three types of features in sequence: Posture change characteristics: composed of displacement increments, velocity vectors and angle changes at adjacent time points, used to reflect the advance trend of the puncture needle; Force perception change characteristics: By processing the force perception signal through differential and window smoothing, the force change rate and force change trend are obtained, which are used to depict the dynamic changes in the stress state of the tissue. Image features: Texture features, local boundary changes, and optical flow displacement information are extracted from ultrasound images using convolutional neural networks or classical image processing algorithms to reflect the deformation trend of tissue structures in the image plane.

[0031] By performing convolutional neural network analysis on ultrasound image sequences, high-dimensional image features reflecting tissue boundary displacement, local structural compression or stretching changes can be extracted from continuous images, thereby enhancing the ability to characterize local deformation trends of soft tissues and reducing the impact of single-frame image noise or imaging instability factors on prediction results.

[0032] The three types of features are combined in chronological order to form a fixed-length feature sequence, which provides the input basis for subsequent prediction models.

[0033] When constructing the feature sequence, the synchronized multimodal data is organized in chronological order and the input sequence is constructed using a sliding time window method.

[0034] Each time window contains pose change features, force change features, and / or image features from multiple consecutive moments, which are used to characterize the dynamic response behavior of soft tissue on a short time scale during needle advancement, thereby providing time-correlated input features for subsequent prediction models.

[0035] In practical applications, the selection of pose change features, force change features, and image features in the feature sequence can be flexibly combined according to the data availability, signal quality, and system configuration of the puncture scene, and is not limited to a fixed form.

[0036] When the puncture needle position is stable and the advancement process exerts significant geometric constraints on the soft tissue, the position change feature can be used first to reflect the trend of the influence of the puncture needle movement on the deformation of the surrounding tissue. When the force sensory signal changes significantly during the contact between the puncture needle and soft tissue, and the image quality is greatly affected by noise or occlusion, the force sensory change feature can be emphasized to characterize the local deformation behavior caused by the force on the soft tissue. When the ultrasound imaging quality is high and the tissue boundaries are clear, image features can be introduced to directly reflect the displacement and deformation of soft tissue structures in the imaging plane. When the system simultaneously possesses stable pose acquisition, reliable force feedback, and clear imaging capabilities, the aforementioned features can be combined to enhance the comprehensive characterization of soft tissue dynamics and improve the stability and robustness of deformation prediction results.

[0037] (3) Prediction model structure like Figure 3 As shown, the machine learning prediction model in this embodiment is a deep learning network capable of modeling time series data. This model can be implemented based on recurrent neural networks, gated unit networks, time-series Transformers, or equivalent structures, but in this invention, it is uniformly abstracted as a "deep learning prediction model with time series modeling capabilities".

[0038] The model's input is the aforementioned multimodal feature sequence. By learning the dynamic relationships between different modes in the feature sequence, the model extracts implicit features reflecting the dynamic effects of soft tissue. At the output, the model provides soft tissue deformation predictions for multiple time points within a preset future time window, including but not limited to: the local displacement field in the needle tip neighborhood, the future displacements of several key points, and the overall offset of the target region.

[0039] During training, the model updates its parameters by comparing the error between the predicted output and the actual deformation label. At the same time, physical constraints such as continuity constraints, smoothness constraints, and tissue elasticity constraints can be introduced to make the prediction results more consistent with the actual mechanical properties of soft tissue in time and space.

[0040] In the implementation of this invention, the physical constraints are applied to the deformation prediction results by combining them with the training or inference process of the prediction model.

[0041] Specifically, the tissue elasticity constraint, continuity constraint, and deformation smoothness constraint can be used as regularization constraints during the training phase of the prediction model to limit the deformation variation between adjacent time nodes or adjacent spatial locations, so that the prediction results conform to the mechanical properties of soft tissue continuous media; or they can be used as post-processing rules in the model output phase to smooth the predicted displacement field or key point displacement to avoid abrupt deformations that do not conform to physical laws.

[0042] By introducing the aforementioned physical constraints, it is possible to improve the stability and rationality of the prediction results of future soft tissue deformation in time and space while maintaining data-driven prediction capabilities.

[0043] (4) Training process like Figure 4 As shown, the training process of the model mainly includes four steps: data preparation, feature construction, model training and validation.

[0044] First, the collected experimental data and simulation-generated data are preprocessed, including time rearrangement, outlier removal, filtering, image cropping, and normalization. Then, the synchronized multimodal data is divided into input sequences and corresponding future deformation sequences according to time segments, forming training sample pairs.

[0045] During the training phase, the input feature sequence is fed into the prediction model, and the future real deformation trajectory is used as the supervision label. The model parameters are updated by optimizing the loss function. The loss function may include a position error term, a displacement field smoothness term, and a physical regularization term that conforms to the laws of tissue mechanics. An iterative training method is used, and once the model reaches the preset accuracy on the validation set, it can be solidified and deployed.

[0046] (5) Schematic diagram of deformation prediction results like Figure 5 As shown, the prediction result generation module in this embodiment structures and organizes the future deformation information output by the model to form a result format that can be used for intraoperative navigation display or auxiliary judgment.

[0047] The prediction result generation module performs unified data organization and formatting on the deformation prediction results output by the machine learning prediction module, so that it can be expressed in the form of displacement vector, key point offset or overall displacement trend of target area, and output to the intraoperative navigation display interface or auxiliary judgment module for the operator to refer to during the puncture process.

[0048] For local displacement fields, the displacement vectors of several sampling points in the vicinity of the needle tip at different future time points can be plotted as displacement arrows; for key point displacements, their offset information in the three-dimensional coordinate system can be output; for target areas of interest in certain tasks, their future overall movement direction and trend can be output. The above prediction results can help operators judge the drift of tissue over time and provide a predictive basis for motion assessment during the puncture process.

[0049] (6) System structure and data flow like Figure 6 As shown, the system constituting this embodiment includes multiple functional modules. The pose acquisition unit, force acquisition unit, and image acquisition unit are responsible for data acquisition, respectively; the signal synchronization module realizes time alignment; the feature extraction module constructs a multimodal feature sequence for prediction; the machine learning prediction module is responsible for providing future deformation prediction results; and the prediction result generation module parses and formats the deformation information and transmits the results to the navigation interface or the host system.

[0050] The data transmission process between modules is clearly defined, enabling real-time provision of future soft tissue deformation trends during puncture, thus providing a reliable predictive basis for the clinical auxiliary navigation system. Example 2

[0051] This embodiment is basically the same as Embodiment 1 in terms of overall process, data acquisition method, multimodal feature construction and prediction result generation method. The difference is that this embodiment adopts a hybrid prediction model structure that combines physical prior constraints to improve the stability and interpretability of future soft tissue deformation prediction.

[0052] In the following description, only the parts that differ from Example 1 will be described in detail.

[0053] like Figure 1 As shown, this embodiment still employs a complete prediction process consisting of signal acquisition and synchronization, feature construction, deformation prediction, and result generation. The acquired pose data, force signals, and ultrasound image sequences are synchronized in time before entering the feature construction stage. This embodiment focuses on introducing another implementation method for the prediction model structure, used to replace the time-series deep learning network in Embodiment 1.

[0054] like Figure 2 As shown, this embodiment still uses the feature construction method in Embodiment 1, including the extraction of pose change features, force perception change features, and image features. However, in this embodiment, the feature sequence is not only used as input to the machine learning model, but also some features are directly used as parameters of the physical equations for solving the subsequent physical constraint module.

[0055] Among them, the rate of change of force perception can be used as the external force term in the soft tissue mechanical response equation; the pose change characteristics can be used as boundary conditions; and the image features can roughly reflect the position of the tissue boundary, providing a structural reference for the physical constraint model.

[0056] like Figure 3 As shown, this embodiment adopts a dual-channel parallel fusion structure of "deep learning model + physical constraint module" to replace the single deep learning sequence model in embodiment 1.

[0057] The structure consists of two sub-modules: 1) Data-driven prediction submodule: This module is similar to the time-series deep learning prediction model in Example 1. It takes a pre-constructed feature sequence as input and outputs the deformation prediction results within a preset future time window. This module is mainly used to capture the coupling relationship between nonlinear and high-dimensional features.

[0058] 2) Physical Constraint Prediction Submodule: Based on the elastic mechanical properties of soft tissue, this embodiment introduces a physical prediction submodule. This module establishes prediction relationships in the following way: Use the rate of change of force perception as the external force input; Use the change in needle tip pose as the boundary driving condition; The local tissue structure extracted from the image is used as a simplified geometric reference; Use continuum mechanics models (such as linear elastic or hyperelastic) to estimate the deformation trend of soft tissue over a short period of time.

[0059] This submodule outputs a physically reasonable deformation estimate, such as the local displacement field or the drift direction of key points.

[0060] The two submodules output the prediction results of the deep model and the prediction results generated by the physical constraints, respectively.

[0061] The fusion module performs weighted or dynamic fusion of the two: when the data quality (such as high image noise) is poor, the weight of the physical module is increased; when the boundary conditions of the physical model are insufficient, the weight of the deep learning module is increased; an adaptive weight strategy is used to make the fusion result more stable in different scenarios, and finally obtains the deformation prediction output result as in this embodiment.

[0062] like Figure 4 As shown, the training process in this embodiment is generally the same as that in Embodiment 1, but the key point is: (1) The training method of the deep learning module remains unchanged, and the supervised learning method is still adopted, using the future deformation sequence as the label.

[0063] (2) The physical constraint module does not require parameter training, but parameter calibration can be performed: tissue elasticity parameters can be calibrated experimentally, and boundary conditions can be estimated based on organ anatomy.

[0064] (3) The fusion strategy can be trained by training the fusion module in the following ways: Optimize the fusion weights using an error minimization strategy on the validation set or learn how to dynamically adjust the fusion ratio by introducing a lightweight neural network to analyze the input features. like Figure 5 As shown, the prediction results in this embodiment include: Local displacement fields at multiple future time points; Key point displacement prediction; The direction and magnitude of the target region's offset in the short term.

[0065] Because this embodiment incorporates physical constraint results, the prediction results have higher smoothness and stability under conditions of rapid change and large changes in tissue boundaries, and are closer to the actual tissue behavior.

[0066] like Figure 6 As shown, the system structure in this embodiment is the same as that in Embodiment 1, except for the internal structure of the machine learning prediction module. The system still includes: a pose acquisition unit, a force sensing acquisition unit, an image acquisition unit, a signal synchronization module, a feature extraction module, a machine learning prediction module (which adopts a hybrid structure internally), and a prediction result generation module.

[0067] The prediction module contains a data-driven submodule, a physical constraint submodule, and a fusion module. The differences lie in the implementation method, but do not change the system interface and data flow structure.

[0068] The hybrid prediction model provided in this embodiment has the following advantages: 1) It still maintains high stability in complex organizational dynamic scenarios (such as severe organizational drift); 2) Predictive outputs that incorporate physical constraints are more interpretable and less prone to abrupt changes in predictions that do not conform to organizational mechanics; 3) It is more robust to experimental noise and image noise; 4) It can automatically adjust the prediction ratio in different scenarios through fusion strategies to improve generalization ability.

[0069] This embodiment complements Embodiment 1 and together constitutes various possible implementations of the present invention.

[0070] The above description is only a preferred embodiment 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 predicting soft tissue deformation during puncture based on machine learning, characterized in that, Includes the following steps: S1. Data Acquisition and Synchronization: Acquire the pose data, force signals and ultrasound image sequences of the puncture needle during the puncture process, and align the pose data, force signals and ultrasound image sequences at different sampling frequencies with timestamps to generate a synchronized dataset. S2. Feature sequence construction: Based on the synchronous dataset, a feature sequence containing information from multiple consecutive moments is constructed in chronological order. The feature sequence is used to characterize the dynamic changes of the soft tissue near the needle tip during the advancement of the puncture needle, and includes at least one of pose change features, force sensation change features, and image features. S3. Predictive Model Inference: Input the feature sequence into the trained machine learning prediction model, so that the prediction model outputs the deformation prediction result of the soft tissue in the vicinity of the needle tip within a future preset time window based on the feature sequence. S4. Output Deformation Prediction: Based on the deformation prediction results, generate at least one of the displacement field of the soft tissue in the vicinity of the needle tip, the displacement of key points, or the offset of the target area, to characterize the future deformation trend of the soft tissue in a short time scale.

2. The method for predicting soft tissue deformation during puncture based on machine learning according to claim 1, characterized in that, The pose change features mentioned in step S2 include at least one of the following: spatial position, advancement speed, advancement direction, and angle changes of the puncture needle at continuous time intervals.

3. The method for predicting soft tissue deformation during puncture based on machine learning according to claim 1, characterized in that, The force sensation change characteristics mentioned in step S2 are obtained by performing time-domain differential processing on the force sensation signal, and are used to characterize the local deformation trend caused by the compression of soft tissue by the puncture needle.

4. The method for predicting soft tissue deformation during puncture based on machine learning according to claim 1, characterized in that, The image features mentioned in step S2 are obtained by performing convolutional neural network analysis on the ultrasound image sequence. The image features include at least one of tissue boundary location, local texture changes, or optical flow displacement.

5. The method for predicting soft tissue deformation during puncture based on machine learning according to claim 1, characterized in that, The prediction model described in step S3 includes recurrent neural networks, gated recurrent unit networks, long short-term memory networks, temporal Transformers, or equivalent time series modeling network structures.

6. The method for predicting soft tissue deformation during puncture based on machine learning according to claim 1, characterized in that, The prediction model in step S3 introduces physical constraints to limit the future deformation prediction results to meet the mechanical properties of soft tissue. The physical constraints include at least one of tissue elasticity constraints, continuity constraints, or deformation smoothness constraints.

7. The method for predicting soft tissue deformation during puncture based on machine learning according to claim 1, characterized in that, The prediction model described in step S3 is obtained by training on training data that reflects the dynamic response behavior of soft tissue during puncture. The training data includes experimental data, simulation data generated based on tissue mechanics models, or a combination of both.

8. The method for predicting soft tissue deformation during puncture based on machine learning according to claim 1, characterized in that, The deformation prediction results in step S4 include a prediction sequence for multiple future time points, which are used to construct the temporal deformation trajectory of soft tissue.

9. The method for predicting soft tissue deformation during puncture based on machine learning according to claim 1, characterized in that, The displacement field described in step S4 is used to characterize the future displacement vectors of multiple spatial locations in the vicinity of the needle tip, and the key point displacement is used to characterize the future spatial offset of the target point, envelope, or risk structure.

10. A machine learning-based system for predicting soft tissue deformation during puncture, characterized in that, include: Position acquisition unit: used to acquire puncture needle position data and transmit the position data to the signal synchronization module; Force sensing acquisition unit: used to acquire force sensing signals and transmit the force sensing signals to the signal synchronization module; Image acquisition unit: used to acquire ultrasound image sequences and transmit the ultrasound image sequences to the signal synchronization module; Signal synchronization module: used to receive the pose data, the force signal and the ultrasound image sequence, align the pose data, the force signal and the ultrasound image sequence with timestamps, and transmit the aligned data to the feature extraction module; Feature extraction module: used to receive the aligned data, construct a feature sequence, and transmit the feature sequence to the machine learning prediction module; Machine learning prediction module: used to receive and process the feature sequence, generate prediction results, and transmit the prediction results to the prediction result generation module; Prediction result generation module: used to receive the prediction results, process the prediction results and generate soft tissue displacement field, key point displacement or target area offset.

Images