Lung tissue puncture early warning method based on multi-dimensional force signals

Abstract

The invention belongs to the field of intelligent operation assistance, and relates to a lung tissue puncture early warning method based on a multi-dimensional force signal, which comprises the following steps: embedding a puncture needle into a sensor, acquiring a total force signal by the sensor, acquiring a space coordinate of a needle tip by using an optical tracking system, and decoupling the total force signal; extracting time domain, frequency domain and force-displacement curve characteristics according to the decoupled multi-dimensional force signals and the space coordinates of the needle tip; acquiring a boundary point set of the high-risk structure of the lung tissue, and calculating spatial distance features according to the spatial coordinates of the puncture needle tip and the boundary point set; fusing the time domain, the frequency domain, the force-displacement curve and the spatial distance features; inputting the fusion feature sequence into a trained risk prediction model to obtain a risk probability; performing early warning decision according to the risk probability; according to the method, time domain, frequency domain and force-displacement characteristics are constructed, deep fusion with the minimum distance of a needle tip-high-risk structure is carried out, mechanical change and space closeness are combined for modeling, and the early warning accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of medical robot and intelligent surgical assistance technology, and relates to a lung tissue puncture early warning method based on multidimensional force signals. Background Technology

[0002] In clinical percutaneous lung biopsy and pulmonary interventional procedures, avoiding accidental injury to adjacent blood vessels, trachea, or bullae remains a challenge. Traditional puncture navigation relies heavily on preoperative imaging (such as CT and MRI) for path planning, with intraoperative guidance primarily based on the surgeon's experience and image analysis. However, the lung position dynamically changes during respiration, and preoperative images cannot reflect the real-time situation. Registration errors may exist between different image modalities, and combined with differences in individual surgeon experience, it is often difficult to perceive the location of risky tissues around the needle tip in a timely and accurate manner. This means that while existing puncture-assisted systems have improved path planning, their ability to identify and avoid risky tissues during the procedure remains limited.

[0003] In recent years, force feedback technology has been introduced into puncture surgery, hoping to perceive tissue characteristics and risks through changes in the interaction force between the needle tip and tissue. However, the nonlinear mechanical behavior of pulmonary soft tissue is complex and affected by respiratory movements. Traditional force feedback models are insufficient to fully describe this dynamic coupling characteristic, and their real-time performance and environmental adaptability need improvement. For example, when the needle tip approaches the pulmonary blood vessel wall or the boundary of a bulla, subtle changes in tissue resistance may occur, which inexperienced operators may find difficult to detect by touch alone. In such cases, the lack of an effective real-time warning mechanism can easily lead to accidental puncture injuries.

[0004] In summary, existing technologies lack a method that can integrate multi-source information and identify lung puncture risks in real time. Summary of the Invention

[0005] To address the aforementioned problems in the prior art, this invention employs a lung tissue puncture early warning method based on multidimensional force signals, comprising:

[0006] S1. A sensor is embedded inside the puncture needle. The sensor collects the total force signal of the puncture needle in real time. The spatial coordinates of the puncture needle tip are collected in real time using an optical tracking system. The collected total force signal is preprocessed.

[0007] S2. Decouple the preprocessed total force signal to obtain the decoupled multidimensional force signal; extract time-domain features, frequency-domain features, and force-displacement curve features based on the decoupled multidimensional force signal and the spatial coordinates of the puncture needle tip.

[0008] S3. Obtain preoperative three-dimensional images of lung tissue, segment the preoperative three-dimensional images of lung tissue, and obtain the boundary point set of high-risk structures of lung tissue.

[0009] S4. Calculate the minimum Euclidean distance from the puncture needle to the high-risk structure in the lung tissue based on the spatial coordinates of the puncture needle tip and the boundary point set of the high-risk structure to obtain the spatial distance characteristics.

[0010] S5. The time domain features, frequency domain features, force-displacement curve features, and spatial distance features are fused to obtain a fused feature sequence;

[0011] S6. Obtain the trained risk prediction model, input the fused feature sequence into the trained risk prediction model to obtain the predicted risk probability; make early warning decisions based on the risk probability.

[0012] Beneficial effects:

[0013] 1. This invention achieves high-precision decoupling of axial force, lateral force, and torsional load by attaching fiber Bragg gratings at multiple points along the puncture needle axis and combining them with coupling matrix calibration. Based on this, it constructs time-domain, frequency-domain, and force-displacement multimodal features, which can meticulously characterize the mechanical response behavior of different tissues and tissue interfaces. Furthermore, it deeply integrates this with the needle tip-high-risk structure minimum distance obtained based on optical positioning and 3D image registration, jointly modeling "mechanical mode change" and "spatial proximity," effectively distinguishing ordinary tissue interfaces from high-risk situations such as adjacent blood vessels and pulmonary bullae. 2. This invention constructs a CNN+LSTM deep learning risk prediction model. The convolutional network automatically extracts local mutation patterns in the force signal sequence, and the LSTM... By capturing the temporal evolution features across windows, typical patterns close to blood vessels, such as "gradual force increase superimposed with small-amplitude oscillations," can be identified, and tissue type and risk probability can be output, thereby significantly improving recognition accuracy and robustness in complex mechanical scenarios; 3. This invention proposes a dynamic threshold function that adaptively tightens with puncture depth, and performs multi-source fusion judgment on the risk probability output by deep learning, the feature deviation index based on statistics / distance, and the spatial distance threshold, which effectively suppresses false alarms and false negatives while improving sensitivity and timeliness. Attached Figure Description

[0014] Figure 1 A flowchart of a lung tissue puncture early warning method based on multidimensional force signals provided in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of multidimensional force signal decoupling and feature extraction provided in an embodiment of the present invention;

[0016] Figure 3 A schematic diagram of the CNN-LSTM risk prediction network structure provided in an embodiment of the present invention;

[0017] Figure 4 The flowchart of dynamic threshold decision and early warning logic provided in the embodiments of the present invention is shown. Detailed Implementation

[0018] 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.

[0019] like Figure 1 As shown, this embodiment of the invention employs a lung tissue puncture early warning method based on multidimensional force signals, comprising:

[0020] S1. An FBG sensor (fiber optic Bragg grating sensor) is embedded inside the puncture needle. The FBG sensor collects the total force signal of the puncture needle in real time. The optical tracking system collects the spatial coordinates of the puncture needle tip in real time and preprocesses the collected total force signal.

[0021] Total force signal Where t is the current time, For time indexing, The length of the signal segment is denoted by ; the total force signal is the sum of the axial force, lateral force, and torsional torque signals, which reflect the interaction mechanical state between the needle tip and different tissues. Since the raw signal is often mixed with environmental noise, sensor baseline drift, and electromagnetic interference, preprocessing is necessary.

[0022] Relying solely on force signals can easily lead to ambiguity. To improve recognition accuracy, this invention introduces an optical tracking system to obtain the spatial coordinates of the puncture needle tip. .

[0023] Preprocessing of the acquired total force signal includes:

[0024] A bandpass filter is used to remove low-frequency drift below 0.2Hz and high-frequency noise above 30Hz from the acquired total force signal, while retaining the main human tissue response frequency band to obtain the first total force signal;

[0025] Wavelet denoising or adaptive filtering is used to eliminate non-stationary noise in the first total force signal, improve the signal-to-noise ratio, and obtain the second total force signal;

[0026] The signals of different channels in the second total force signal are z-score normalized to ensure that the dimensions of each channel are uniform, thus obtaining the third total force signal.

[0027] A zero-point baseline was established using a static loading experiment. The third total force signal was then corrected based on the zero-point baseline to remove the influence of temperature changes and long-term drift on the signal, resulting in a preprocessed total force signal.

[0028] S2. Decouple the preprocessed total force signal to obtain the decoupled multidimensional force signal; extract multimodal feature vectors based on the decoupled multidimensional force signal and the spatial coordinates of the puncture needle tip; the multimodal feature vectors include: time-domain features. Frequency domain characteristics and the characteristics of the force-displacement curve ;

[0029] like Figure 2 As shown, since the FBG sensor is attached to the inner wall of the puncture needle at multiple points, the spectral response of different grating channels is simultaneously affected by multi-directional forces and temperature coupling. To obtain the true physical quantity, the preprocessed total force signal needs to be decoupled. The specific process includes:

[0030] The estimated coupling matrix is ​​obtained using sensor calibration experiments;

[0031] Specifically, during the calibration phase, known standard axial force, standard transverse force, and standard torsional torque are applied to the puncture needle, and the total force signal collected by the sensor is... That is, the wavelength change of each grating channel;

[0032] Constructing linear equations:

[0033]

[0034] Where A is the coupling matrix, For standard axial force, M represents the standard lateral force, and M represents the standard torsional moment.

[0035] Based on the total force signal, standard axial force, standard lateral force, and standard torsional moment collected by the sensor, the linear equation is solved by the least squares method to obtain the estimated coupling matrix.

[0036] During the actual puncture, the force signal from the sensor was measured. The decoupled multidimensional force signal can be obtained by solving the inverse matrix of the estimated coupling matrix: .

[0037] The decoupled multidimensional force signal can accurately reflect axial insertion force, lateral offset force and torsional moment, providing a basis for tissue differentiation.

[0038] Extracting the time-domain features of the decoupled multidimensional force signal includes: calculating the mean, variance, skewness, and kurtosis of the force signal in each dimension after decoupling, reflecting the overall signal distribution and fluctuations; and calculating the maximum rate of change of the force signal F in each dimension after decoupling. It is used to identify the breakthrough point of the organizational interface; the mean, variance, skewness, kurtosis and maximum rate of change of force signals in all dimensions are combined to obtain the time-domain characteristics, that is, the time-domain characteristics of the current time t. .

[0039] Extracting the frequency domain features of the decoupled multidimensional force signal includes: performing a Fast Fourier Transform on the force signal of each decoupled dimension to obtain the spectrum of the force signal of each dimension. Due to differences in viscoelasticity, different tissues exhibit characteristic differences in their spectral distribution. The power spectral density or amplitude of the main frequency components of the force signal spectrum of each dimension is extracted. The power spectral density or amplitude of the main frequency components of the force signal of all dimensions is combined to obtain the frequency domain features, i.e., the frequency domain features at the current time t. .

[0040] The force-displacement curve features extracted from the decoupled multidimensional force signal include:

[0041] Based on the spatial coordinates of the puncture needle tip Calculate the displacement of the puncture needle tip Based on the axial force signal after decoupling of the puncture needle and the displacement of the needle tip Constructing force-displacement curves ;

[0042] Force-displacement curves can be manipulated using first-order differences or smoothed derivatives. Differentiation (i.e.) ), thus obtaining the stiffness-displacement curve ; Apply first-order differences or smoothed derivatives to the stiffness-displacement curves Differentiation (i.e.) ), thus obtaining the stiffness change rate-displacement curve;

[0043] Force-displacement curves obtained through first-order difference Differentiation includes: calculating the force-displacement curve Local stiffness of each point i ;in, Let be the change in local force at point i at the current time and the previous time. Let be the change in local displacement of point i between the current time and the previous time.

[0044] Using a sliding window to analyze the force-displacement curves Stiffness-displacement curve The process is performed to obtain the force-displacement curve. Stiffness-displacement curve The extreme values ​​within each window will be used to determine the force-displacement curve. Stiffness-displacement curve The extreme values ​​within each window are combined, and the combined result is used as the force-displacement curve feature, that is, the force-displacement curve feature at the current time t. ;

[0045] Force-displacement curve The extreme values ​​are: Stiffness-displacement curve The extreme value is .

[0046] In one embodiment, based on lateral force Calculate the lateral resultant force and the rate of change of the lateral resultant force with displacement. Add the lateral resultant force and its rate of change to the force-displacement curve features to help determine whether the needle body is bent or laterally squeezed.

[0047] These features constitute a multidimensional vector describing the mechanical environment of the needle tip, providing rich input for subsequent deep learning models.

[0048] S3. Obtain preoperative three-dimensional images of lung tissue, segment the preoperative three-dimensional images of lung tissue, and obtain the boundary point set of high-risk structures in the lung tissue. ;

[0049] High-risk structures in the lungs include blood vessels, trachea, and bullae.

[0050] S4. Calculate the minimum Euclidean distance from the puncture needle to the high-risk structures in the lung tissue based on the spatial coordinates of the needle tip and the boundary point set of the high-risk structures, thus obtaining the spatial distance characteristics. ;

[0051] Minimum Euclidean distance from needle tip to high-risk structures in lung tissue Defined as:

[0052]

[0053] in, For the set of boundary points The i-th point in the array.

[0054] As an additional input, working in conjunction with mechanical characteristics, it can effectively distinguish between "force signal abrupt changes originating from tissue interfaces" and "proximity to high-risk structures." For example, if force signal fluctuations are accompanied by a sharp decrease in ΔD, it is highly likely that the needle tip is close to the blood vessel wall.

[0055] S5. The time domain features, frequency domain features, force-displacement curve features, and spatial distance features are fused to obtain a fused feature sequence;

[0056] The specific steps for fusing time-domain features, frequency-domain features, force-displacement curve features, and spatial distance features include:

[0057] Time-domain features Frequency domain characteristics Force-displacement characteristics Spatial distance characteristics By combining the features, a comprehensive feature sequence is obtained. ;

[0058] Using a sliding window (e.g., length L, step size S) to synthesize feature sequences Processing is performed to obtain the feature sequence. , where C is the feature dimension.

[0059] For the comprehensive feature sequence The processing includes: sequentially acquiring the comprehensive feature sequence using a sliding window on the time axis. The feature sequences within the sliding window are concatenated to obtain a concatenated feature sequence. These concatenated features from all time windows are then combined to obtain a fused feature sequence. .

[0060] S6. Obtain the trained risk prediction model, input the fused feature sequence into the trained risk prediction model to obtain the predicted risk probability; make early warning decisions based on the risk probability.

[0061] like Figure 3 As shown, the risk prediction model includes a convolutional neural network, a long short-term memory network, and a multi-task output network. The trained risk prediction model processes the fused feature sequence by: inputting the fused feature sequence into the convolutional neural network and the long short-term memory network respectively; fusing the output features of the convolutional neural network and the long short-term memory network; and inputting the fused features into the multi-task output network to obtain the predicted organization type and risk probability.

[0062] Convolutional neural networks (CNNs), by extracting local patterns through one-dimensional convolutional kernels (1D Conv), can identify typical abrupt changes and subtle fluctuations in force signal sequences. For example, the instantaneous force peak that occurs when the pleura is breached can be sensitively captured by the convolutional layer. Specifically, the CNN in this embodiment of the invention includes a first 1D Conv layer, a first Pooling layer, a second 1D Conv layer, and a second Pooling layer connected in series.

[0063] Long Short-Term Memory (LSTM) networks utilize memory units to analyze long-term dependencies and capture the slow upward trend of force signals as the needle tip gradually approaches the blood vessel. Specifically, the LSTM network in this embodiment of the invention includes a first LSTM layer and a second LSTM layer connected in series.

[0064] The multi-task output network includes a tissue type classification layer and a risk probability prediction layer, both of which are fully connected layers. The tissue type classification layer uses Softmax to classify the current tissue type (such as normal lung tissue, adjacent blood vessels, adjacent bullae, etc.), and the risk probability prediction layer outputs the risk probability using a Sigmoid function. .

[0065] The predicted tissue type does not directly participate in the threshold logic for automatically triggering warnings, but it is used for real-time display on the visual navigation interface. Informing doctors that the needle tip is currently located in the "lung parenchyma," "near the blood vessel wall," or "pleura" can assist doctors in making proactive judgments, verifying the credibility of the risk probability, and avoiding misjudgments that may be caused by a single probability indicator.

[0066] The training process for the risk prediction model includes:

[0067] A lung tissue biopsy training set is obtained, which includes the total force signal of the puncture needle and the spatial coordinates of the needle tip. Temporal features, frequency domain features, force-displacement curve features, and spatial distance features are constructed based on the lung tissue biopsy dataset. These features are then fused to obtain a fused feature sequence. This fused feature sequence is input into a risk prediction model to obtain the predicted tissue type and the predicted risk probability. A loss function value is calculated based on the predicted tissue type, the actual tissue type, the predicted risk probability, and the actual risk label (safe or unsafe). The parameters of the risk prediction model are updated based on the loss function value. When the loss function value is minimized, the trained risk prediction model is obtained.

[0068] In one embodiment, during training, hierarchical cross-validation is employed to ensure a balance among different sample types, and data augmentation (such as time jitter, slight noise perturbation, and random time shift) is introduced to improve the model's generalization ability. Simultaneously, methods such as temperature scaling are used to calibrate the probability output, making the risk probabilities more realistic.

[0069] like Figure 4 As shown, early warning decisions based on risk probability include:

[0070] Setting a dynamic threshold function ;in, The initial threshold is typically in the range of 0.6–0.8, and the optimal value is obtained through parameter tuning using a validation set. For puncture process factors, Let be the displacement of the needle tip at time t. This is the preset target puncture depth or the maximum effective stroke of the puncture needle. This value is determined before the surgery based on the puncture path length planned in the preoperative CT image (i.e., the Euclidean distance from the needle entry point on the skin to the target lesion); α is the tightening coefficient, generally taken as 0.1~0.3; as the puncture depth increases, Gradually reduce, i.e. tighten the threshold, to improve sensitivity in deep high-risk areas.

[0071] Set distance threshold Minimum distance between the needle tip and a high-risk structure Related to anatomical safety clearance, based on experimental data, this invention will use a distance threshold. The value range is set to 2-4 mm, and the specific value is selected through ROC curve analysis and safety preference, which is used as a criterion for spatial proximity.

[0072] All samples labeled "safe" (i.e., located within normal lung parenchyma, far from blood vessels and trachea) were selected from the lung tissue biopsy training set. Statistical analysis was performed on the fusion feature sequences of all safe samples to obtain the mean of the feature distribution of the fusion feature sequences. With covariance matrix This represents the statistical baseline for "normal puncture status";

[0073] Based on the characteristic distribution mean With covariance matrix The fused feature sequence output by step S5 using Mahalanobis distance calculation Characteristic deviation :

[0074]

[0075] in, Represents the covariance matrix The inverse matrix is ​​used to eliminate correlations and dimensional differences between different feature dimensions.

[0076] when A significant increase indicates that the current mechanical mode has deviated significantly from the normal safe mode.

[0077] Set three early warning conditions:

[0078] The predicted risk probability is greater than the current dynamic threshold: when At that time, the needle tip was considered to be in a high-risk area;

[0079] Feature deviation is greater than the current dynamic threshold: when At that time, it was considered that the mechanical mode had shifted abnormally;

[0080] Spatial distance features are less than the distance threshold: when At that time, it was believed that the needle tip was too close to the high-risk structure.

[0081] If any one of the warning conditions is met, a risk warning is triggered; otherwise, no risk warning is triggered, data collection continues, and a judgment is made.

[0082] When an alert is triggered, the system enters the decision output phase. Its functions include: audible and visual alarm: immediately emitting sound and light signals to alert the operator; interface prompts: highlighting the needle tip position on the surgical navigation interface and indicating the risk category; path correction suggestions: based on model analysis results, prompting the operator to adjust the needle tip direction or withdraw it; robot control linkage: in automatic puncture mode, the system can directly pause the robotic arm movement to ensure safety.

[0083] The entire process achieves closed-loop control of signal acquisition, feature construction, model prediction, threshold judgment, and early warning output, which can ensure the safety of puncture surgery in real time with high frequency and low latency.

[0084] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early warning of lung tissue puncture based on multidimensional force signals, characterized in that, include: S1. A sensor is embedded inside the puncture needle. The sensor collects the total force signal of the puncture needle in real time. The spatial coordinates of the puncture needle tip are collected in real time using an optical tracking system. The collected total force signal is preprocessed. S2. Decouple the preprocessed total force signal to obtain the decoupled multidimensional force signal; Time-domain features, frequency-domain features, and force-displacement curve features are extracted based on the decoupled multidimensional force signal and the spatial coordinates of the puncture needle tip. S3. Obtain preoperative three-dimensional images of lung tissue, segment the preoperative three-dimensional images of lung tissue, and obtain the boundary point set of high-risk structures of lung tissue. S4. Calculate the minimum Euclidean distance from the puncture needle to the high-risk structure in the lung tissue based on the spatial coordinates of the puncture needle tip and the boundary point set of the high-risk structure to obtain the spatial distance characteristics. S5. The time domain features, frequency domain features, force-displacement curve features, and spatial distance features are fused to obtain a fused feature sequence; S6. Obtain the trained risk prediction model, input the fused feature sequence into the trained risk prediction model to obtain the predicted risk probability; make early warning decisions based on the risk probability.

2. The method for early warning of lung tissue puncture based on multidimensional force signals according to claim 1, characterized in that, The preprocessing of the acquired total force signal includes: using a bandpass filter to remove low-frequency drift and high-frequency noise from the acquired total force signal to obtain the first total force signal; using wavelet denoising or adaptive filtering to eliminate non-stationary noise in the first total force signal to obtain the second total force signal; normalizing the signals of different dimensions in the second total force signal to obtain the third total force signal; establishing a zero-point baseline using a static loading experiment, and correcting the third total force signal based on the zero-point baseline to obtain the preprocessed total force signal.

3. The method for early warning of lung tissue puncture based on multidimensional force signals according to claim 1, characterized in that, Decoupling the preprocessed total force signal includes: A known standard axial force, standard transverse force, and standard torsional torque are applied to the puncture needle, and the total force signal acquired by the sensor is recorded. ; Constructing linear equations: Where A is the coupling matrix, For standard axial force, M is the standard lateral force, and M is the standard torsional moment. Based on the total force signal, standard axial force, standard lateral force, and standard torsional moment collected by the sensor, the linear equation is solved by the least squares method to obtain the estimated coupling matrix; The preprocessed total force signal is decoupled based on the estimated coupling matrix to obtain the decoupled multidimensional force signal.

4. The method for early warning of lung tissue puncture based on multidimensional force signals according to claim 1, characterized in that, Extracting the time-domain features of the decoupled multidimensional force signal includes: calculating the mean, variance, skewness, kurtosis, and maximum rate of change of the force signal in each dimension after decoupling; and combining the mean, variance, skewness, kurtosis, and maximum rate of change of the force signals in all dimensions to obtain the time-domain features.

5. A method for early warning of lung tissue puncture based on multidimensional force signals according to claim 1, characterized in that, Extracting the frequency domain features of the decoupled multidimensional force signal includes: performing a fast Fourier transform on each dimension of the decoupled force signal to obtain the spectrum of the force signal in each dimension; extracting the power spectral density of the spectrum of the force signal in each dimension; and combining the power spectral densities of the force signals in all dimensions to obtain the frequency domain features.

6. The method for early warning of lung tissue puncture based on multidimensional force signals according to claim 1, characterized in that, The decoupled multidimensional force signals include: axial force, lateral force, and torsional moment; the force-displacement curve features extracted from the decoupled multidimensional force signals include: Calculate the displacement of the puncture needle tip based on its spatial coordinates. Based on the axial force signal after decoupling of the puncture needle and the displacement of the needle tip Constructing force-displacement curves ; Force-displacement curves obtained through first-order difference Differentiating the values ​​yields the stiffness-displacement curve. Stiffness-displacement curves were obtained through first-order differences. Differentiating the coefficients yields the stiffness-displacement curve. Using a sliding window to analyze the force-displacement curves Stiffness-displacement curve The process is performed to obtain the force-displacement curve. Stiffness-displacement curve The extreme values ​​within each window will be used to determine the force-displacement curve. Stiffness-displacement curve The extreme values ​​within each window are combined, and the combined result is used as the feature of the force-displacement curve.

7. The method for early warning of lung tissue puncture based on multidimensional force signals according to claim 1, characterized in that, The fusion of time-domain features, frequency-domain features, force-displacement features, and spatial distance features includes: combining time-domain features, frequency-domain features, force-displacement features, and spatial distance features to obtain a comprehensive feature sequence; and processing the comprehensive feature sequence using a sliding window to obtain a fused feature sequence.

8. A method for early warning of lung tissue puncture based on multidimensional force signals according to claim 1, characterized in that, The risk prediction model includes a convolutional neural network, a long short-term memory network, and a multi-task output network. The trained risk prediction model processes the fused feature sequence by inputting the fused feature sequence into the convolutional neural network and the long short-term memory network respectively, fusing the output features of the convolutional neural network and the long short-term memory network, and inputting the fused features into the multi-task output network to obtain the predicted organization type and risk probability.

9. A method for early warning of lung tissue puncture based on multidimensional force signals according to claim 1, characterized in that, Early warning decisions based on risk probability include: Set a dynamic threshold function, and obtain the current dynamic threshold based on the dynamic threshold function; set a distance threshold. Obtain the training set used by the risk prediction model during the training phase, and select all samples marked as "safe" from the training set; perform statistical analysis on the fusion feature sequences of all samples marked as "safe" to obtain the feature distribution mean and covariance matrix; The feature bias of the fused feature sequence output in step S5 is calculated based on the feature distribution mean and covariance matrix. Three early warning conditions are set: the predicted risk probability is greater than the dynamic threshold, the feature deviation is greater than the dynamic threshold, and the spatial distance feature is less than the distance threshold. If any one of the warning conditions is met, a risk warning will be triggered; otherwise, a risk warning will not be triggered.

10. A method for early warning of lung tissue puncture based on multidimensional force signals according to claim 9, characterized in that, Dynamic threshold function ;in, As the initial threshold, Let be the puncture progress factor at time t. The preset puncture target depth, Let α be the displacement of the puncture needle at time t, and α be the tightening coefficient.

Images