A deep learning-based puncture path risk quantification system and method

Abstract

The present application relates to the technical field of deep learning and puncture design cross, and particularly relates to a puncture path risk quantification system and method based on deep learning, which integrates three major core puncture risk indicators of needle path deviation, collision risk and tissue interaction, comprehensively captures potential risk points of each link in the puncture process, relies on a puncture risk quantification deep model to accurately quantify the risk of each initial alternative puncture path in a single puncture process, and presents the risk characteristics of each initial alternative puncture path in a stereoscopic manner through multi-dimensional indicators complementing each other and progressing layer by layer, so that the risk assessment result is more comprehensive and has more reference value, and the puncture execution path with the lowest comprehensive risk is accurately selected from multiple initial alternative puncture paths.

Description

Technical Field

[0001] This invention relates to the field of deep learning and puncture design technology, and more specifically, to a puncture path risk quantification system and method based on deep learning. Background Technology

[0002] Current clinical puncture procedure planning relies heavily on the experience and judgment of medical personnel, lacking systematic quantitative assessment tools, which significantly limits risk identification during the puncture process. Traditional methods often focus only on single types of risk factors, failing to comprehensively integrate multi-dimensional puncture risk indicators such as needle deviation, collision risk, and tissue interaction. This makes it difficult to fully capture the dynamic changes of needle tip deviation from the preset path, and also fails to adequately assess the risk of contact between the puncture needle and prohibited areas, as well as the potential impact of tissue resistance during needle insertion. Consequently, the comprehensiveness and accuracy of risk assessment are insufficient, potentially overlooking critical surgical risk points.

[0003] Existing methods for quantifying puncture risk mostly employ static, linear assessment logic, failing to adequately adapt to the dynamic fluctuations in patients' physiological states. Assessment results based solely on a single, fixed physiological scenario are insufficient to reflect the impact of physiological changes on risk during actual surgery. Furthermore, traditional quantification methods lack intelligent model support, failing to effectively map disparate risk indicators into a unified risk quantification result. This makes it difficult to intuitively compare risks from different pathways, limiting the reliability and practicality of risk quantification and hindering its ability to meet the refined puncture risk assessment requirements of precision medicine.

[0004] Especially in difficult or high-risk puncture scenarios, the selection of puncture pathways becomes even more challenging. These challenges include situations such as small or deep lesions, lesions adjacent to critical risk structures like blood vessels, nerves, or bones, anatomical variations in the patient, limitations imposed by scarring, ulceration, or dense subcutaneous vascularization on the available needle insertion area, puncture pathways traversing multiple layers of tissue with varying resistance, and dynamic changes in the target and prohibited areas due to respiratory or heart rate fluctuations. Medical personnel often struggle to reliably assess the risk differences between multiple candidate pathways based solely on experience. Therefore, it is necessary to establish a pathway risk quantification method that comprehensively considers needle deviation, collision risks in prohibited areas, and tissue interaction risks for these difficult puncture scenarios. This will help in selecting lower-risk puncture pathways more suitable for actual puncture procedures.

[0005] Based on the above, this application proposes a deep learning-based system and method for quantifying puncture path risk. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a puncture path risk quantification system and method based on deep learning.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based puncture path risk quantification system includes: The 3D anatomical model module is used to build a 3D anatomical model of the puncture area; The initial alternative puncture path module generates multiple initial alternative puncture paths based on the three-dimensional anatomical model of the puncture area. The comprehensive risk quantification module for puncture paths is used to build a simulation model of puncture paths and further obtain the comprehensive puncture risk index of each initial candidate puncture path. The path puncture risk index of an initial candidate puncture path is obtained as follows: Select an initial candidate puncture path, obtain the path puncture risk index of this initial candidate puncture path in the i-th puncture process, and then obtain the average path puncture risk index (Avs) and the risk fluctuation range (RDRI). Based on the comparison result between the path puncture risk index and the path puncture risk threshold, determine whether the path puncture risk index is a high-risk index. Finally, sum the number of high-risk indices and mark them as Nusq. Mark the input feature with the highest risk contribution weight corresponding to the high-risk index as the preferred risk feature. Perform pairwise comparisons on all preferred risk features. When two preferred risk features are the same, increase the risk focusing count by one; when two preferred risk features are not the same, increase the risk dispersion count by one. Finally, mark the total number of pairwise comparisons as TT, sum the risk focusing counts and mark them as WED, sum the risk dispersion counts and mark them as LKP, and then use the formula... The comprehensive risk index of the initial alternative puncture path was calculated. ;in, , All are weighting coefficients;

[0008] The puncture execution path determination module marks the initial alternative puncture path with the lowest comprehensive puncture risk index as the puncture execution path, and performs puncture surgery on the patient according to the puncture execution path.

[0009] Furthermore, the average path puncture risk index Avs is obtained by summing and averaging the path puncture risk indices of the i puncture processes.

[0010] Furthermore, the Risk Volatility Range (RDRI) is obtained by calculating the difference between the maximum and minimum path puncture risk indices during the i-th puncture process.

[0011] Furthermore, the path puncture risk index of the initial candidate puncture path during a single puncture process is obtained as follows: Select a set of physiological data, import the set of physiological data into the puncture path simulation model to construct the corresponding physiological scenario, simultaneously import the parameter package of the initial candidate puncture path, start the puncture simulation, after the simulation puncture is completed, obtain the basic data of various puncture risk indicators, import the basic data of various puncture risk indicators into the puncture risk quantification depth model, and the puncture risk quantification depth model outputs the path puncture risk index.

[0012] Furthermore, the puncture risk indicators include the following three items: needle deviation indicator, collision risk indicator, and tissue interaction indicator.

[0013] Furthermore, the needle deviation index includes the following basic data: maximum deviation, standard deviation of deviation fluctuation, and duration of deviation exceeding the standard.

[0014] Furthermore, the collision risk indicators include the following basic data: prohibited area contact time, prohibited area contact depth, and contact type risk index.

[0015] Furthermore, the organizational interaction indicators include the following basic data: resistance fluctuation amplitude and needle insertion resistance peak index.

[0016] Furthermore, a deep learning-based method for quantifying puncture path risk is presented, with the following specific steps: Step 1: Construct a three-dimensional anatomical model of the puncture area; Step 2: Generate multiple initial alternative puncture paths; Step 3: Quantify the overall risk of path penetration for each initial candidate path; Step 4: Determine and execute the puncture path.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The system and method of this invention integrate three major categories of core puncture risk indicators: needle path deviation, collision risk, and tissue interaction. It comprehensively captures potential risk points at each stage of the puncture process. Relying on a puncture risk quantification depth model, it accurately quantifies the risk of each initial candidate puncture path in a single puncture process. It simulates the risk data under multiple physiological fluctuation scenarios obtained from path puncture and constructs a multi-dimensional comprehensive risk analysis system. The average path puncture risk index reflects the overall risk level of the path, the risk fluctuation range reflects the impact of physiological fluctuations on puncture risk, the number of high-risk indices clarifies the exposure frequency of high-risk scenarios, and the risk focusing and dispersion times accurately characterize the distribution characteristics of risk sources. These multi-dimensional indicators complement each other and progress layer by layer, presenting a comprehensive view of the risk characteristics of each initial candidate puncture path, making the risk assessment results more comprehensive and valuable. The puncture path with the lowest overall risk is accurately selected from multiple initial candidate puncture paths, avoiding the subjectivity and uncertainty of traditional path selection based on experience, and providing a clear direction for surgical intervention through precise risk source localization. The combination of three types of puncture risk indicators and a deep model ensures the accuracy of risk quantification, while the construction of a multi-dimensional comprehensive risk index achieves comprehensive risk assessment. The synergistic effect of both makes puncture path planning more scientific and feasible. Attached Figure Description

[0018] Figure 1 A flowchart of a deep learning-based method for quantifying the risk of puncture paths; Figure 2 Flowchart for building a simulation model of the puncture path; Figure 3 This is a flowchart for generating initial alternative puncture paths. Detailed Implementation

[0019] Example 1: Refer to Figures 1 to 3 A deep learning-based puncture path risk quantification system includes a three-dimensional anatomical model module, an initial alternative puncture path module, a comprehensive puncture path risk quantification module, and a puncture execution path determination module.

[0020] The 3D anatomical model module is used to build a 3D anatomical model of the puncture area.

[0021] The specific construction process for the 3D anatomical model of the puncture area is as follows: Thin-slice scanning of the puncture area is performed using CT (1-2mm slice thickness) or MRI (T1-weighted sequence) to ensure image resolution of fine structures such as blood vessels, nerves, and lesions. The data output is in DICOM standard format, containing grayscale information of the patient's individualized anatomical structures, covering skin, subcutaneous tissue, lesions, key risk structures (blood vessels, nerves, bones), and organs. Based on the 3D Slicer or MITK platform, denoising, grayscale calibration, and inter-slice interpolation are performed to eliminate image artifacts and structural distortions caused by slice thickness differences, laying the foundation for accurate segmentation. A combination of automatic segmentation and manual correction is used to separate different anatomical structures. Automatic segmentation: For large structures such as bones and organs, a threshold segmentation algorithm is used; for fine structures such as blood vessels and nerves, a deep learning segmentation model is used, trained on a pre-labeled clinical image dataset. Manual correction: Doctors fine-tune the segmentation boundaries through the software interface to correct deviations in automatic segmentation (such as missed blood vessel branches or blurred lesion edges), ensuring no key risk structures are missed. The model employs a MarchingCubes algorithm to convert segmented 2D image sequences into 3D mesh models with a resolution of 0.5–1 mm, balancing accuracy and efficiency. Different structures are assigned independent labels and attributes (e.g., lesions: red, blood vessels: blue, nerves: yellow, bones: gray), and their physical properties (blood vessel diameter, organ capsule thickness, etc.) and 3D coordinates are recorded. The polygon count of non-critical area meshes (e.g., subcutaneous tissue) is reduced by 30%–50%, while retaining refined meshes for critical structures (blood vessels, lesions) to prevent stuttering caused by excessively large models. The model is exported in STL format (compatible with Unity3D simulation modules) while preserving coordinate information and attribute labels for each structure, ensuring collision detection and distance calculation functions are possible after importing the simulation model. Doctors can manually adjust the model's structural positions based on the patient's actual condition (e.g., anatomical variations) to ensure consistency between the model and the patient's actual anatomical structures.

[0022] The initial alternative puncture path module generates multiple initial alternative puncture paths based on the three-dimensional anatomical model of the puncture area.

[0023] Multiple initial alternative puncture paths are generated based on the 3D anatomical model of the puncture area. The specific process is as follows: The prohibited areas (blood vessels, nerves, bones, with collision object attributes) and safe areas (subcutaneous tissue, intermuscular spaces, without collision object attributes) in the 3D anatomical model are identified. The 3D coordinates of the lesion center, puncture needle constraints (needle length L (30~150mm, set according to the surgical scenario), maximum bending angle ≤15° (to avoid excessive path bending leading to puncture failure), effective working length ≤0.8L, where L is the total length of the puncture needle), and path difference threshold (Euclidean distance from the needle entry point (skin surface starting point)) are determined. The distance between the needle insertion point and the percutaneous needle insertion angle (the angle between the path and the normal to the skin surface) is ≥15°; a rectangular / circular candidate area (area ≥5cm×5cm) is delineated within the safe area on the skin surface, and areas with scars, ulcers, and dense subcutaneous blood vessels are removed, and 50~100 potential needle insertion points are generated in advance (evenly distributed); a spherical area with a radius of 5mm centered on the lesion is used as the target buffer zone (the path endpoint falling into this area is considered "target hit"); the shortest distance between any node on the path and the surface of the prohibited area is ≥3mm (routine puncture) / ≥5mm (high-risk puncture), which is used as the threshold for collision detection. Two random trees are established: T_start (starting point tree) with a random point within the needle insertion point candidate area as the root node, and T_goal (target tree) with a random point in the target point buffer area as the root node. Algorithm parameters are set as follows: step size Δ = 3~5mm (adjusted according to the accuracy requirements of the puncture area, e.g., Δ = 3mm for lung puncture, Δ = 5mm for breast puncture), maximum number of iterations N_max = 5000 (to avoid infinite loops), and connection threshold d_connect = 5mm (connection is attempted if the distance between nodes in the two trees is ≤ this value). 70% of sampling points are randomly generated from the safe area, and 30% are generated from the needle insertion point candidate area / target point buffer area to reduce invalid sampling in prohibited areas. If a generated sampling point falls into a prohibited area or the safe distance is insufficient, it is automatically discarded and resampled until a valid sampling point is obtained. If a path has already been generated, subsequent sampling prioritizes areas ≥ 10mm away from existing needle insertion points to ensure path differentiation. For the root node of T_start, search for the node closest to the sampling point, and expand the node by a step size Δ along the line connecting the two points to generate a new node. If the distance between the new node and the sampling point is less than Δ, then the sampling point is directly used as the new node. Based on the collision properties of the 3D anatomical model, a point-to-surface collision detection algorithm is used to determine whether the new node and the expanded line segment touch the prohibited area: if the distance between any point of the line segment and the prohibited area is less than the safe distance for avoiding the prohibited area, it is determined to be a collision, and the new node is discarded; if there is no collision, the new node is added to T_start, and the 3D coordinates and needle insertion angle of the node are recorded. Target tree expansion is synchronized: T_goal is expanded using the same logic to achieve bidirectional search acceleration.When the distance between the new node of T_start and the node of T_goal is ≤ d_connect, attempt to directly connect the two nodes to generate a complete path; Path validity verification: Target reach verification: Whether the end point of the path falls into the target point buffer zone; Avoidance verification: Whether the entire path meets the safe distance from the prohibited area; Physical constraint verification: Whether the path curvature is ≤ θ_max (determined by calculating the sum of the angles between adjacent line segments of the path), and whether the path length is ≤ needle length L. Initial path generation: Repeat the above steps to generate 8~10 valid paths (more than the number of targets, used for filtering); Needle entry point distance judgment: Calculate the Euclidean distance between the needle entry point of the new path and the needle entry point of the retained path. If it is ≥10mm, it meets the requirements; Needle entry angle judgment: Calculate the difference between the needle entry direction of the new path (the initial line segment direction of the needle entry point → target point) and the needle entry angle of the retained path. If it is ≥15°, it meets the requirements; Deduplication processing: Remove duplicate or highly similar paths (path overlap > 70%), and retain 3~5 paths that meet the difference threshold.

[0024] The comprehensive risk quantification module for puncture paths is used to build a simulation model of puncture paths and further obtain the comprehensive puncture risk index of each initial candidate puncture path.

[0025] The comprehensive risk index of an initial candidate puncture path is obtained as follows: Select an initial candidate puncture path and obtain the path puncture risk index for that initial candidate puncture path in the i-th puncture process (each puncture process corresponds to a path puncture risk index, and the physiological data corresponding to each puncture process are not the same, which conforms to the patient's normal physiological fluctuation range, such as respiratory rate 10~30 breaths / minute, heart rate 60~120 beats / minute). Each puncture process corresponds to a set of physiological data of the patient, which is divided according to the degree of physiological fluctuation and generated in an orderly manner. For example, the set of physiological data corresponding to the first puncture process is as follows: The first set of physiological data for the first puncture was: respiratory rate 10 breaths / min + heart rate 60 beats / min; the second set of physiological data for the first puncture was: respiratory rate 12 breaths / min + heart rate 70 beats / min; the third set of physiological data for the i-th puncture was: respiratory rate 30 breaths / min + heart rate 120 beats / min (all simulated). The average path puncture risk index (Avs) was calculated by summing and averaging the path puncture risk indices for the i-th puncture. The difference between the maximum and minimum path puncture risk indices for the i-th puncture was calculated to obtain the risk fluctuation range (RDRI). A path puncture risk threshold was then set (path...). The puncture risk threshold is obtained through joint calibration based on clinical safety boundaries and model quantitative validation (value can be 0.7). When the puncture risk index is higher than the puncture risk threshold, the corresponding puncture risk index is marked as a high-risk index (no marking is needed if it is not higher). Finally, the number of high-risk indices is summed and marked as Nusq. The input feature with the highest risk contribution weight corresponding to the high-risk index is marked as the preferred risk feature (e.g., if the input feature with the highest risk contribution weight corresponding to a high-risk index is the maximum offset, then the maximum offset is marked as the preferred risk feature). All preferred features are then marked as high-risk features. Select risk features and perform pairwise comparisons (e.g., if there are three preferred risk features, namely preferred risk feature A, preferred risk feature B, and preferred risk feature C, then perform AB comparison, BC comparison, and AC comparison). When two preferred risk features are the same (e.g., both preferred risk features are the maximum offset), increase the risk focusing count by one. When two preferred risk features are not the same, increase the risk dispersion count by one. Finally, mark the total number of pairwise comparisons as TT, sum the risk focusing counts and mark them as WED, sum the risk dispersion counts and mark them as LKP, and use the formula... The comprehensive risk index of the initial alternative puncture path was calculated. ;in, , All are weighting coefficients. Because the more focused the risk sources, the easier it is to optimize them specifically; the more dispersed the risk sources, the more difficult it is to control the overall risk. The value can be 0.4. The value can be 0.6.

[0026] The path puncture risk index of the initial candidate puncture path in a single puncture procedure is obtained as follows: Select a set of physiological data and import it into the puncture path simulation model to construct the corresponding physiological scenario. Simultaneously import the parameter package of the initial candidate puncture path (the puncture path simulation model can simulate the complete puncture process of the initial candidate puncture path under this physiological scenario). Start the puncture simulation. After the simulation is completed, obtain the basic data of various puncture risk indicators and import the basic data of various puncture risk indicators into the puncture risk quantification depth model (the puncture risk quantification depth model is a deep learning model, which will be fine-tuned for each patient; for each patient, it will first be...). CT / MRI scans are used to obtain a personalized 3D anatomical model (STL format). After importing into Unity, a patient-specific simulation sample (e.g., 100 samples covering the range of physiological fluctuations) is generated. These samples are used to fine-tune the puncture risk quantification depth model (only updating the weights of the last 1-2 hidden layers). The fine-tuned model will learn the anatomical / tissue characteristics of the patient and, under the same physiological data, output weights that match the individual patient, ensuring the personalized accuracy of risk assessment. Fine-tuning requires only a small number of patient-specific samples (100 samples), training time is less than 10 minutes, can be completed on a regular medical workstation, and is fully compatible with the clinical preoperative preparation process. The puncture risk quantification depth model outputs a path puncture risk index.

[0027] The puncture risk indicators include the following three items: needle deviation indicator, collision risk indicator, and tissue interaction indicator; the needle deviation indicator includes the following basic data: maximum deviation, standard deviation of deviation fluctuation, and duration of deviation exceeding the standard; the collision risk indicator includes the following basic data: contact time in prohibited areas, contact depth in prohibited areas, and contact type risk index; the tissue interaction indicator includes the following basic data: resistance fluctuation amplitude and peak needle insertion resistance index.

[0028] The maximum offset is obtained as follows: During the simulated puncture process, at each sampling time point (the interval between sampling time points is determined by the sampling frequency; a sampling frequency of 10Hz can be set to meet the data acquisition requirements of puncture surgery), the needle tip coordinates corresponding to the sampling time point are obtained. Simultaneously obtain the node closest to the needle tip coordinates on the initial candidate puncture path. (The nearest node is determined by minimizing the Euclidean distance), and the offset at that sampling time is calculated using the three-dimensional Euclidean distance formula: traversing all sampling times The maximum value is the maximum offset; The standard deviation of the offset fluctuation is obtained as follows: Obtain the offset at all sampling times. Calculate the mean of the offsets at all sampling times. ; k is the total number of sampling times; through the formula The standard deviation of the offset fluctuation is calculated.

[0029] The duration of offset exceeding the limit is obtained as follows: Set an offset exceeding the limit threshold T (e.g., 3mm), and iterate through all sampling times... Statistical satisfaction The number of times n is used to obtain the sampling interval Δt = 1 / sampling frequency = 0.1 seconds (10Hz sampling), and the offset excess time T. over =n×Δt.

[0030] The contact time in the prohibited area is obtained as follows: when the collision sensor of the puncture needle enters the collision body in the prohibited area, it is marked as being in contact; when it leaves the range of the prohibited area, it is marked as not in contact. Since the sampling is at the same time difference (e.g., 10Hz, 0.1 seconds interval), all sampling times are traversed, and the number of times n is in contact is counted; the contact time in the prohibited area = n × sampling interval Δt.

[0031] The contact depth in the prohibited area is obtained as follows: Calculated only at sampling moments during contact, based on the needle tip coordinates in the contact state. The surface distance of the restricted area can be obtained using Unity's Physics.ClosestPoint method. The nearest point S; calculate the depth using the 3D Euclidean distance formula: ;in,( , , Let S be the coordinates of point S; iterate through all contact moments. The maximum value is taken as the contact depth of the prohibited area.

[0032] The contact type risk index is obtained as follows: Based on the clinical risk level of the prohibited area, preset base weights are used (example): Blood vessels: 1.0 (high risk of bleeding); Nerves: 0.8 (risk of nerve damage); Bones: 0.5 (usually non-penetrable, low contact risk). The depth coefficient for the corresponding prohibited area is: Depth coefficient = min(contact depth / 5mm, 1) (5mm is preset as the severe depth threshold); the time coefficient for the corresponding prohibited area is: Time coefficient = min(contact time / 2 seconds, 1) (2 seconds is preset as the severe duration threshold). The contact type risk index = preset base weight of blood vessels × depth coefficient of blood vessels × time coefficient of blood vessels + preset base weight of nerves × depth coefficient of nerves × time coefficient of nerves + preset base weight of bones × depth coefficient of bones × time coefficient of bones.

[0033] The resistance fluctuation amplitude is obtained as follows: The resistance value corresponding to the needle tip at each sampling moment is obtained. The average resistance value is calculated by summing and averaging the resistance values ​​corresponding to the needle tip at all sampling times. The sample standard deviation formula is used: The resistance fluctuation amplitude was calculated.

[0034] The peak needle insertion resistance index is obtained as follows: The tissue layer where the needle tip is located is identified in real-time using Unity's tissue layer collider tags (e.g., subcutaneous, muscle, organ); resistance data is collected within each tissue layer, and the maximum resistance value within that layer is taken as the peak needle insertion resistance index (e.g., subcutaneous peak value). peak muscle mass organ peak Basic resistance weight allocation for tissue layers: The weights of each tissue layer are preset according to the clinical risk level (example): subcutaneous: 0.2 (low resistance, low risk); muscle: 0.3 (medium resistance, moderate risk); organs: 0.5 (high resistance, high risk of damage). .

[0035] The specific construction process of the deep learning model for puncture risk quantification is as follows: Through a puncture path simulation model, generate simulation samples with different initial alternative paths and different physiological data in batches, covering the entire risk range (low / medium / high risk); generate at least 10,000 valid samples (the deep learning model needs sufficient data to support its generalization ability). Each sample includes input features + labels: Input features (8-dimensional structured data): maximum offset, standard deviation of offset fluctuation, duration of offset exceeding the standard, contact time in the prohibited area, contact depth in the prohibited area, contact type risk index, and resistance fluctuation amplitude. 1. Peak resistance index; Label (target value): Risk label in the 0-1 range (needs to be combined with clinical expert scores). Labeling logic: Invite 3-5 clinical puncture physicians to score each sample from 1 to 10 based on the clinical risk significance of 8 basic data (e.g., contact depth > 3mm = high risk, deviation duration > 1 second = high risk); linearly map the 1-10 scores to a 0-1 range label (e.g., 1 point → 0.0, 5 points → 0.5, 10 points → 1.0), and take the average of the expert scores as the final label to ensure the clinical rationality of the label. The 8 basic data units are different (e.g., length mm, force N, time s), and all features need to be compressed to the [0,1] range through Min-Max normalization. Formula: (in , The minimum and maximum values ​​of this feature in the sample set are preset based on the distribution of the simulation data, such as the maximum offset. Contact depth Extreme anomalies caused by simulation errors (such as contact depth > 10mm, peak resistance > 20N) were removed, and valid data were screened using the 3σ principle. The dataset was divided into a training set (7000 records), a validation set (2000 records), and a test set (1000 records) in a 7:2:1 ratio to ensure consistent risk label distribution across sets (avoiding data skew). Considering the input is 8-dimensional structured data (non-image / sequence data), a lightweight fully connected neural network (MLP) was chosen as the basic architecture to avoid overfitting and inference delays caused by complex models. The architecture details are as follows: the input layer has 8 neurons to receive 8 normalized basic data points; hidden layer 1 has 64 neurons, using ReLU as the activation function; regularization layer 1 uses Dropout (0.2) as the activation function; hidden layer 2 has 32 neurons, using ReLU as the activation function; regularization layer 2 uses L2 regularization (λ=0.01) as the activation function; hidden layer 3 has 16 neurons. ReLU is used as the activation function; the output layer has only one neuron and uses Sigmoid as the activation function; the core logic of the architecture design is: lightweight: the total number of parameters is less than 10,000 (8×64+64×32+32×16+16×1=3920), the inference speed is fast (millisecond level), and it can meet the real-time requirements of Unity simulation; nonlinear fitting: 3 hidden layers + ReLU activation function, which can fully fit the complex nonlinear relationship between 8 features and risk index (such as the collaborative risk of "contact depth + contact time"); anti-overfitting: through Dropout (random deactivation) + L2 regularization double insurance, overfitting in small sample scenarios is avoided. Core training parameter settings: Loss function: Weighted combination of mean squared error (MSE) and Huber loss (weight ratio 7:3): MSE ensures overall fitting accuracy, while Huber loss is more robust to outliers (avoiding training fluctuations caused by expert labeling bias); Optimizer: Adam optimizer, initial learning rate = 0.001, learning rate scheduling strategy = ReduceLROnPlateau (the learning rate is halved if the validation set loss does not decrease after 3 epochs); Training epochs: 50-100 epochs, Batch size: 32-64 (adapted to ordinary GPU / CPU training); Early stopping mechanism: Training stops when the validation set loss does not decrease for 5 consecutive epochs, saving the optimal model (avoiding overfitting). Training environment configuration: Framework selection: TensorFlow / PyTorch (lightweight framework, supports ONNX format export, compatible with Unity calls). Because the Sigmoid output may exhibit a mismatch between predicted and actual risk (e.g., a prediction of 0.8 but an actual risk of medium), it needs to be calibrated using Platt scaling: The calibration function is trained using validation set data. ; The model output is shown below, and a and b are calibration parameters. Calibration objective: Ensure that the risk index classification of "risk index below 0.3 = low risk, 0.3-0.7 = medium risk, and above 0.7 = high risk" aligns with clinical expert judgment. Quantitative accuracy indicators: Mean absolute error (MAE): <0.05 (risk index prediction bias controlled within 5%); Coefficient of determination (R²): >0.85 (the model can explain more than 85% of risk variation); Root mean square error (RMSE): <0.08. Risk grading accuracy: Graded into "low / medium / high" levels, accuracy on the test set >90%; High-risk recall: Identification rate of true high-risk samples (label >0.7) on the test set >95% (avoiding missed high-risk paths). After removing the needle path deviation indicator, collision risk indicator, and tissue interaction indicator respectively, the model was retrained, and the performance degradation was compared: Removing any indicator required a decrease in R² of >0.15, proving that all three indicators are core risk factors with no redundancy. Export Format: Export the trained model to ONNX format (cross-platform compatible, Unity supports direct calls); Model Compression: Compress the model weights from 32-bit floating-point to 16-bit using Quantization-Aware Training (QAT), further reducing inference latency (inference time < 10ms). Interface Development: Write the model call interface in C# script (using Unity's ONNX Runtime plugin). Core Process: After the simulated puncture, read 8 basic data points; normalize the data according to the Min-Max parameters during training; input the normalized features into the ONNX model to obtain the calibrated risk index; output the risk index (keeping 2 decimal places, such as 0.23, 0.67). Stability Guarantee: Add exception handling logic in Unity (e.g., automatically adopt the default safe value when the feature value exceeds the preset range) to avoid model inference crashes. After the model path puncture risk index YDRI, calculate the model path puncture risk index for each input feature. Partial derivatives of (normalized values ​​of 8 basic indicators) Calculate the gradient-feature product for each feature: (Absolute value ensures positive contribution, avoiding offsetting positive and negative contributions); Normalization yields the risk contribution weight: (The sum of the 8 weights is 1, such as the contact type risk index of 0.35 and the maximum offset of 0.25).

[0036] The specific construction process for the puncture path simulation model is as follows: Import the optimized 3D anatomical model (STL format) into Unity3D, and assign collider attributes according to structural type: Trigger colliders are set for prohibited areas (blood vessels, nerves) to record contact parameters; rigid colliders are set for bones (simulating impenetrable characteristics); and no colliders are set for safe areas. Structural labels (red - lesion, blue - blood vessel, etc.) and physical properties (blood vessel diameter, organ elasticity) are retained and mapped to Unity components. The model coordinate system is unified with the clinical surgical coordinate system (with the patient's surface markers as the origin) to ensure that the 3D coordinates of the simulation path are consistent with the actual position within the patient's body. A rigid puncture needle model is created in Unity according to the pre-set puncture needle parameters (length, diameter, bending angle threshold), and PhysX physical properties (density, rigidity) are bound. A collision detection sensor is installed at the needle tip to support real-time capture of the needle insertion position and posture. Based on the Unity Animation system, keyframe animations are created for structures such as the lungs and chest wall. An inspiratory-expiratory cycle is defined, and respiratory rate parameters are bound to scripts to synchronously drive the model's expansion / contraction deformation, simulating the positional fluctuations of target and prohibited areas caused by respiration. Vertex animation technology is used to set displacement keyframes for the wall vertices of the vascular model, driving small vibrations of the vertices according to the heart rate cycle, reproducing the dynamic impact of vascular pulsation on the path safety distance. Alternative path parsing and loading: Import parameter packages (needle insertion point coordinates, angles, node sequences) for 3-5 initial alternative puncture paths, visualize and annotate the path trajectory in the Unity scene, and support one-click switching of the simulation path. Based on the PhysX physics engine, the interaction between the puncture needle and tissue is simulated—elastic deformation components are used for subcutaneous tissue (creating indentation upon needle insertion and rebound after needle withdrawal), and resistance gradient components are used for muscles / fascia (simulating differences in needle insertion resistance among different tissues); the needle insertion speed is set in stages (10mm / s for subcutaneous segment and 5mm / s for deep segment) to recreate the realistic puncture feel. For non-critical tissues (such as subcutaneous fat), LOD (Level of Detail) technology is used to reduce model accuracy during long-distance rendering, ensuring stable simulation frame rate.

[0037] The puncture execution path determination module marks the initial alternative puncture path with the lowest comprehensive puncture risk index as the puncture execution path, and performs puncture surgery on the patient according to the puncture execution path.

[0038] Before and during the actual puncture according to the puncture execution path, a corresponding aseptic operation area can be marked on the patient's body surface based on the needle entry point coordinates, needle entry angle, and body surface projection range of the puncture execution path. This aseptic operation area is used as an auxiliary verification area during the operation. After the medical staff completes skin disinfection, aseptic draping, and aseptic coverage of the puncture needle / catheter and guide components, the puncture is performed according to the needle entry point and needle entry angle output by the system. When it is necessary to adjust the needle entry angle or change the puncture instrument during the puncture, the aseptic operation area is used as a reference to determine whether the adjustment operation is still within the aseptic coverage range. If the adjustment range exceeds the preset aseptic operation area or the non-sterile part of the puncture instrument may enter the puncture channel, it will prompt to re-disinfect the local area, replace the aseptic instrument, or re-lay the aseptic barrier, thereby helping to maintain the aseptic operation conditions of the puncture process while performing a low-risk puncture path.

[0039] After the puncture is completed, the actual position data of the catheter tip or puncture needle tip can be further obtained. The actual position data can be obtained from intraoperative ultrasound, CT / fluoroscopy review images, navigation positioning sensors or other positioning methods that can identify the tip position. The actual position data is registered with the coordinate system of the aforementioned three-dimensional anatomical model to obtain the actual three-dimensional coordinates of the catheter tip or puncture needle tip. The system compares the actual three-dimensional coordinates with the distance between the target buffer zone, the end point of the puncture execution path and the prohibited area. When the actual three-dimensional coordinates fall into the target buffer zone or the deviation from the end point of the puncture execution path is less than the preset success threshold, and there is no excessive contact with prohibited areas such as blood vessels, nerves, and bones, the system outputs an auxiliary judgment result of successful puncture. Otherwise, the system outputs prompts such as tip deviation, failure to reach the target or risk of contact with adjacent risk structures to assist medical personnel in judging the position of the catheter tip and whether the puncture is successful.

[0040] Example 2: A deep learning-based method for quantifying the risk of puncture paths, the specific steps of which are as follows: Step 1: Construct a three-dimensional anatomical model of the puncture area; Step 2: Generate multiple initial alternative puncture paths; Step 3: Quantify the overall risk of path penetration for each initial candidate path; Step 4: Determine and execute the puncture path.

[0041] The above formulas are all dimensionless calculations, and the preset parameters in the formulas should be set by those skilled in the art according to the actual situation.

[0042] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0043] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0044] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0045] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0046] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0047] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0048] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based puncture path risk quantification system, characterized in that, include: The 3D anatomical model module is used to build a 3D anatomical model of the puncture area; The initial alternative puncture path module generates multiple initial alternative puncture paths based on the three-dimensional anatomical model of the puncture area. The comprehensive risk quantification module for puncture paths is used to build a simulation model of puncture paths and further obtain the comprehensive puncture risk index of each initial candidate puncture path. The path puncture risk index of an initial candidate puncture path is obtained as follows: Select an initial candidate puncture path, obtain the path puncture risk index of this initial candidate puncture path in the i-th puncture process, and then obtain the average path puncture risk index (Avs) and the risk fluctuation range (RDRI). Based on the comparison result between the path puncture risk index and the path puncture risk threshold, determine whether the path puncture risk index is a high-risk index. Finally, sum the number of high-risk indices and mark them as Nusq. Mark the input feature with the highest risk contribution weight corresponding to the high-risk index as the preferred risk feature. Perform pairwise comparisons on all preferred risk features. When two preferred risk features are the same, increase the risk focusing count by one; when two preferred risk features are not the same, increase the risk dispersion count by one. Finally, mark the total number of pairwise comparisons as TT, sum the risk focusing counts and mark them as WED, sum the risk dispersion counts and mark them as LKP, and then use the formula... The comprehensive risk index of the initial alternative puncture path was calculated. ;in, , All are weighting coefficients; The puncture execution path determination module marks the initial alternative puncture path with the lowest comprehensive puncture risk index as the puncture execution path, and performs puncture surgery on the patient according to the puncture execution path.

2. The puncture path risk quantification system based on deep learning according to claim 1, characterized in that, The average path puncture risk index Avs is obtained by summing and averaging the path puncture risk indices of the i puncture processes.

3. The puncture path risk quantification system based on deep learning according to claim 1, characterized in that, The Risk Volatility Range (RDRI) is obtained by calculating the difference between the maximum and minimum path puncture risk indices during the i-th puncture process.

4. The puncture path risk quantification system based on deep learning according to claim 1, characterized in that, The path puncture risk index of the initial candidate puncture path in a single puncture process is obtained as follows: Select a set of physiological data, import the set of physiological data into the puncture path simulation model to construct the corresponding physiological scenario, simultaneously import the parameter package of the initial candidate puncture path, start the puncture simulation, after the simulation puncture is completed, obtain the basic data of various puncture risk indicators, import the basic data of various puncture risk indicators into the puncture risk quantification depth model, and the puncture risk quantification depth model outputs the path puncture risk index.

5. The puncture path risk quantification system based on deep learning according to claim 4, characterized in that, Puncture risk indicators include the following three items: needle deviation indicator, collision risk indicator, and tissue interaction indicator.

6. The puncture path risk quantification system based on deep learning according to claim 5, characterized in that, The needle deviation index includes the following basic data: maximum deviation, standard deviation of deviation fluctuation, and duration of deviation exceeding the standard.

7. The puncture path risk quantification system based on deep learning according to claim 5, characterized in that, The collision risk index includes the following basic data: prohibited area contact time and prohibited area contact depth.

8. The puncture path risk quantification system based on deep learning according to claim 5, characterized in that, Collision risk indicators also include the contact type risk index.

9. A puncture path risk quantification system based on deep learning according to claim 5, characterized in that, The organizational interaction indicator includes the following basic data: resistance fluctuation amplitude and needle insertion resistance peak index.

10. A deep learning-based method for quantifying puncture path risk, applied to the deep learning-based puncture path risk quantification system described in any one of claims 1-9, characterized in that, The specific steps are as follows: Step 1: Construct a three-dimensional anatomical model of the puncture area; Step 2: Generate multiple initial alternative puncture paths; Step 3: Quantify the overall risk of path penetration for each initial candidate path; Step 4: Determine and execute the puncture path.

Images