Prostate precision puncture robot path planning method based on deep learning

Abstract

The application discloses a prostate precise puncture robot path planning method based on deep learning, relates to the field of path planning, and first carries out global path planning based on MRI data before operation, and abstracts a puncture environment and a planned path into a graph structure model.In operation, the dynamic deformation of the prostate is perceived through real-time ultrasound images, and the affected local dangerous area in the initial path graph is accurately identified based on the deformation information.Subsequently, the graph neural network technology is introduced, the affected local subgraph is subjected to rapid and intelligent rerouting calculation, a new safe local path is generated, and finally the new path is seamlessly integrated back into the original global path graph, and efficient correction of the overall path is completed.In this way, the time-consuming bottleneck of global calculation can be avoided, so that the real-time problem of intraoperative path dynamic correction is solved under the premise of ensuring high precision, and the path planning ability of the robot in a dynamic environment is significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of path planning, and more particularly, to a prostate precise puncture robot path planning method based on deep learning. BACKGROUND

[0002] As one of the common malignant tumors of men in the world, prostate cancer, its early diagnosis and precise treatment are crucial to improve the prognosis of patients. Transrectal ultrasound-guided prostate biopsy is the gold standard for obtaining tissue samples for pathological diagnosis. To overcome the limitations of traditional manual puncture, which relies on the experience of doctors and has unstable precision, a robot-assisted puncture system has emerged, which can significantly improve the target hit rate and surgical safety of puncture through precise mechanical arm control. In the robot-assisted puncture process, accurate path planning is the key to success. This step usually needs to integrate preoperative high-resolution magnetic resonance imaging (MRI) data to identify lesions and plan the best puncture path that avoids the urethra, nerve and blood vessel bundles and other key tissues. In recent years, deep learning technology has provided a new way for realizing automatic and high-precision intelligent path planning due to its excellent ability in medical image segmentation, registration and feature extraction. Building a precise puncture robot path planning scheme based on deep learning has become a research hotspot and development direction in this field.

[0003] In the prior art, some schemes try to use complex non-rigid registration algorithms to track the changes of anatomical structures in real time and re-plan the entire puncture path on this basis. Although this method can theoretically achieve high correction accuracy, it involves intensive computing of full three-dimensional images, such as biomechanical models or complex voxel-level deformation field solving, which requires huge computing resources and is difficult to meet the stringent real-time requirements in surgery, often leading to operation delays and affecting the smoothness of the operation. On the contrary, some real-time methods use simplified models or local registration strategies, which shorten the calculation time but may sacrifice the accuracy of registration and path planning, and cannot effectively compensate for complex nonlinear deformation, resulting in new puncture deviations. Therefore, how to realize real-time and efficient correction of the path under the premise of ensuring high accuracy in the presence of dynamic deformation is a technical bottleneck that needs to be solved in the current prostate puncture robot field.

[0004] Therefore, there is an urgent need for an optimized prostate precise puncture robot path planning method based on deep learning. SUMMARY

[0005] To solve the above technical problems, the present application is proposed.

[0006] According to an aspect of the present application, a deep learning-based prostate precise puncture robot path planning method is provided, which comprises: performing preoperative global path generation and graph structuring on preoperative MRI scan data to obtain an initial path graph and an initial 3D safety map; performing real-time updating on the initial 3D safety map based on a captured current ultrasound frame to obtain a real-time 3D safety map and a three-dimensional deformation field; performing path deformation perception and risk assessment on the initial path graph based on the three-dimensional deformation field and the real-time 3D safety map to obtain a segmented risk score vector and an affected subgraph; performing local path rerouting on the affected subgraph based on a graph neural network to obtain a new local path; and performing path graph integration on the initial path graph based on the new local path to obtain a revised path graph.

[0007] Compared with the prior art, the deep learning-based prostate precise puncture robot path planning method provided by the present application firstly performs global path planning based on MRI data before operation, and abstracts the puncture environment and the planned path into a graph structure model. In the operation, the dynamic deformation of the prostate is perceived through real-time ultrasound images, and the affected local dangerous area in the initial path graph is accurately identified based on the deformation information. Subsequently, the graph neural network technology is introduced, and the affected local subgraph is quickly and intelligently rerouted to generate a new safe local path, and finally the new path is seamlessly integrated into the original global path graph to complete the efficient revision of the overall path. In this way, the time-consuming bottleneck of global calculation can be avoided, thereby solving the real-time problem of intraoperative path dynamic revision under the premise of ensuring high precision, and significantly improving the path planning ability of the robot in a dynamic environment. BRIEF DESCRIPTION OF DRAWINGS

[0008] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application taken in conjunction with the accompanying drawings. The accompanying drawings are provided to aid in the understanding of the present application and constitute a part of the specification, together with the written description, to explain the present application and, together with the written description, serve to explain the present application. The drawings do not limit the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0009] Figure 1 A flowchart of the deep learning-based prostate precise puncture robot path planning method according to the embodiments of the present application.

[0010] Figure 2 A data flow diagram of the deep learning-based prostate precise puncture robot path planning method according to the embodiments of the present application.

[0011] Figure 3 A flowchart of S1 in the deep learning-based prostate precise puncture robot path planning method according to the embodiments of the present application.

[0012] Figure 4 Flowchart of S13 in the deep learning-based prostate precise puncture robot path planning method according to the embodiment of the application.

[0013] Figure 5 Flowchart of S2 in the deep learning-based prostate precise puncture robot path planning method according to the embodiment of the application.

[0014] Figure 6 Flowchart of S3 in the deep learning-based prostate precise puncture robot path planning method according to the embodiment of the application.

[0015] Figure 7 Flowchart of S31 in the deep learning-based prostate precise puncture robot path planning method according to the embodiment of the application. DETAILED DESCRIPTION

[0016] Embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. Although some embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms, and should not be interpreted as being limited to the embodiments set forth herein, but rather, these embodiments are provided to more thoroughly and completely understand the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for exemplary purposes only, and are not intended to limit the scope of protection of the present disclosure.

[0017] In view of the problems in the above background art, the present application proposes a deep learning-based prostate precise puncture robot path planning method. Figure 1 Flowchart of the deep learning-based prostate precise puncture robot path planning method according to the embodiment of the application. Figure 2 Data flow diagram of the deep learning-based prostate precise puncture robot path planning method according to the embodiment of the application. As shown in Figure 1 and Figure 2 The deep learning-based prostate precise puncture robot path planning method 100 includes: S1, performing preoperative global path generation and graph structuring on preoperative MRI scan data to obtain an initial path graph and an initial 3D safety map; S2, performing real-time updating on the initial 3D safety map based on a captured current ultrasound frame to obtain a real-time 3D safety map and a three-dimensional deformation field; S3, performing path deformation perception and risk assessment on the initial path graph based on the three-dimensional deformation field and the real-time 3D safety map to obtain a segmented risk score vector and an affected subgraph; S4, performing local path re-routing on the affected subgraph based on a graph neural network to obtain a new local path; and S5, performing path graph integration on the initial path graph based on the new local path to obtain a corrected path graph.

[0018] In the aforementioned deep learning-based robotic path planning method for precise prostate puncture, step S1 involves generating a global path and structuring a graph from preoperative MRI scan data to obtain an initial path map and an initial 3D safety map. It should be understood that preoperative MRI scan data, due to its high resolution, can clearly present the anatomical structures of the prostate, target lesion, urethra, neurovascular bundles, and other key dangerous tissues, providing a precise anatomical basis for puncture path planning. Intraoperative real-time imaging technologies (such as ultrasound), while reflecting dynamic changes, have lower resolution and cannot support precise global path planning. Therefore, this application further processes the preoperative MRI scan data to generate an initial global path and a structured graph model, thereby providing a basic safety path framework for the entire puncture process. The path and environmental information are abstracted in the form of a graph structure, facilitating efficient identification and correction of affected local paths during the procedure. This ensures the accuracy of the initial path planning, while the abstraction of the graph structure provides a structured basis for local calculations during intraoperative dynamic correction, avoiding the high computational overhead of intraoperative global replanning. Thus, while ensuring the overall path safety, the real-time performance of intraoperative path correction is improved.

[0019] In particular, in one specific embodiment, Figure 3 This is a flowchart of step S1 in the deep learning-based robotic path planning method for precise prostate puncture according to an embodiment of this application. Figure 3 As shown, preoperative global path generation and graph structuring are performed on preoperative MRI scan data to obtain an initial path map and an initial 3D safety map, including: S11, performing semantic segmentation on preoperative MRI scan data based on a pre-trained 3D U-Net model to obtain an initial 3D safety map; S12, inputting the initial 3D safety map into the optimal path planning module to obtain the original optimal path; S13, converting the original optimal path into the initial path map.

[0020] Specifically, step S11, the preoperative MRI scan data is subjected to semantic segmentation based on a pre-trained 3D U-Net model to obtain an initial 3D safety map. It should be understood that, although the preoperative MRI scan data can clearly present the three-dimensional anatomical structure of the prostate, the target lesion and the surrounding key tissues such as the urethra, the nerve and vascular bundle, these structures exist in the form of gray scale information in the original data and are not clearly distinguished, so direct use for path planning cannot accurately identify the safety area and the forbidden area. Therefore, the pre-trained 3D U-Net model is used to perform semantic segmentation on the MRI data, wherein the 3D U-Net model is a deep learning model specially used for processing three-dimensional data, which is based on the classic U-Net architecture and can effectively process the segmentation task of the preoperative MRI scan data by expanding to three-dimensional space, so as to accurately divide the prostate area, the target lesion and the forbidden area to be avoided, and construct an initial 3D safety map containing the spatial distribution of these areas. In this way, clear spatial constraints can be provided for subsequent path planning, ensuring that the planned path avoids dangerous tissues and accurately reaches the target lesion, laying a foundation for puncture safety and accuracy.

[0021] In particular, in one possible embodiment, the implementation process of step S11 is as follows: first, the pre-training process of the 3D U-Net model adopts a semi-supervised strategy, trains on a dataset containing sparse labeled slices, focuses on learning in the labeled area through a weighted cross-entropy loss function, and combines three-dimensional rotation, elastic deformation and other data enhancement methods to improve robustness, while using the Adam optimizer to iterate training to convergence, ensuring that the model has high-precision segmentation capability for the prostate and surrounding tissues. Second, obtain the preoperative MRI scan data of the patient, which is three-dimensional voxel data containing gray scale information of the prostate and surrounding tissues. Then, input the MRI data into the pre-trained 3D U-Net model, which extracts local features through multi-layer convolution operations, realizes feature upsampling through deconvolution operations, and fuses feature information of different scales through skip connection, so as to accurately identify and segment the prostate area, the target lesion and the forbidden area. Finally, according to the segmentation result, specific labels are assigned to different areas, such as 1 for the prostate area, 2 for the target lesion, 3 for the forbidden area, and 0 for the background, to construct an initial 3D safety map of the distribution of each area in three-dimensional space.

[0022] Specifically, in step S12, the initial 3D safety map is input into the optimal path planning module to obtain the original optimal path. It should be understood that, since the initial 3D safety map has clearly marked the safety area, the target lesion and the forbidden area, it provides clear spatial constraints for path planning, but it is necessary to find an optimal path from the puncture entry point to the target lesion under the constraints to ensure the safety and efficiency of the puncture process. Therefore, the initial 3D safety map is input into the optimal path planning module to plan an original optimal path that has optimal length, good smoothness and can accurately reach the target lesion while avoiding the forbidden area. In this way, the path can have high execution efficiency on the basis of safety, reducing the puncture time, and the smoothness of the path can reduce the difficulty of robot operation and improve the puncture accuracy.

[0023] In particular, in one possible embodiment, the implementation process of step S12 is as follows: first, the three-dimensional coordinates of the puncture entry point and the target lesion are determined, the entry point is selected at a safe position of the rectal wall or perineum, and the target lesion coordinates are determined by the lesion mark in the initial 3D safety map. Second, the initial 3D safety map is input into the optimal path planning module, which uses the RRT* algorithm to generate a tree structure by random sampling in three-dimensional space, continuously expanding the branches to explore the reachable space, while evaluating each potential path to select a path that avoids the forbidden area, has shorter length and smaller curvature. Finally, the optimal solution is selected from the generated paths as the original optimal path, which starts from the entry point and is located in the safety area throughout the journey, and finally accurately reaches the target lesion.

[0024] Specifically, in step S13, the original optimal path is converted into an initial path graph. It should be understood that, since the original optimal path is a continuous geometric trajectory, it is difficult to efficiently identify the affected local area and make targeted corrections when the prostate deforms intraoperatively, while the graph structure can clearly represent the topological relationship of the path through nodes and edges, facilitating the extraction and recalculation of local subgraphs. Therefore, the original optimal path is further converted into an initial path graph to discretize the continuous path into an initial path graph composed of nodes and edges, where the nodes represent key points on the path, and the edges represent the connection relationship between the nodes. In this way, the local subgraph affected by deformation can be quickly located intraoperatively, and only this part needs to be recalculated, avoiding the high time consumption of global re-planning and significantly improving the real-time performance of dynamic path correction.

[0025] In particular, in one specific embodiment, Figure 4 The flowchart of step S13 in the deep learning-based prostate accurate puncture robot path planning method according to the embodiment of the present application is shown in FIG. 13. As shown in FIG. 13, the step S13 includes steps S131 and S132. Figure 4As shown, converting the original optimal path into an initial path graph includes: S131, discretizing the original optimal path into a series of path points; S132, taking the path points as nodes of the initial path graph and taking the connections between adjacent path points as edges of the initial path graph.

[0026] More specifically, in step S131, the original optimal path is discretized into a series of path points. It should be understood that, since the original optimal path is a continuous geometric trajectory from the puncture entry point to the target lesion, its continuous form cannot be directly used to construct a graph structure with topological relationships, and when the prostate deforms intraoperatively, it is difficult to accurately locate the specific path segment affected, resulting in a lack of clear operation objects for local correction. Therefore, the original optimal path is discretized into a series of path points to convert the continuous trajectory into quantifiable and locatable discrete points, providing a basic unit for subsequent construction of a graph structure. In this way, the spatial distribution of the path can be broken down into specific coordinate points, facilitating intraoperative analysis of the displacement and risk changes of each path point to accurately lock the area affected by deformation, providing a clear operation object for local path correction, while reducing the complexity of subsequent graph structure processing and improving the efficiency of path dynamic adjustment.

[0027] In particular, in one possible embodiment, the implementation process of step S131 is as follows: first, a sampling interval parameter is determined, which is set according to the puncture accuracy requirement and the path length to ensure that the discretized path points can both fully reflect the direction of the original path and not increase the computational load due to excessive number of points, for example, setting the interval to 1 mm to balance accuracy and efficiency. Secondly, starting from the puncture entry point along the original optimal path, uniform sampling is performed according to the set interval, and each point on the path is obtained in turn, each point recording its three-dimensional spatial coordinates (x, y, z), the coordinate values being determined based on the coordinate system of the preoperative MRI to ensure consistency with the spatial reference of the initial 3D safety map. Finally, the path points obtained by sampling are numbered in the order of their position on the path to form an ordered path point sequence, the first point of the sequence being the puncture entry point and the last point being the target lesion center point, and the intermediate points corresponding to the intermediate positions on the path in turn, thereby fully retaining the spatial direction information of the original path.

[0028] More specifically, in step S132, path points are used as nodes of the initial path graph, and connections between adjacent path points are used as edges of the initial path graph. It should be understood that since the discretized path points are isolated spatial coordinates, lacking a description of the connection relationships between path segments, they cannot reflect the topological structure of the path, making it difficult to identify and replan local path segments. Therefore, this application further uses path points as nodes of the initial path graph and connections between adjacent path points as edges of the initial path graph, thereby constructing a graph structure containing nodes and edges, clarifying the topological relationships between path points and the spatial attributes of path segments. In this way, the overall topological relationship of the path can be clearly presented through the graph structure, with each path segment corresponding to an edge in the graph and each path point corresponding to a node. Thus, during the operation, by analyzing the displacement of nodes and the risk changes of edges, local subgraphs affected by deformation (composed of related nodes and edges) can be quickly identified, and rerouting calculations can be performed only on these subgraphs, avoiding global processing of the entire path and significantly improving the real-time performance and efficiency of dynamic path correction.

[0029] In the aforementioned deep learning-based robotic path planning method for precise prostate puncture, step S2 involves updating the initial 3D safety map in real time based on the captured current ultrasound frame to obtain a real-time 3D safety map and a three-dimensional deformation field. It should be understood that the prostate gland undergoes non-rigid deformation during surgery due to factors such as patient respiration, ultrasound probe pressure, and muscle contraction, leading to deviations between the initial 3D safety map and the actual anatomical location. Therefore, this application updates the initial 3D safety map in real time based on the captured current ultrasound frame to obtain a real-time 3D safety map reflecting the actual anatomical structure during surgery, and simultaneously calculates the three-dimensional deformation field describing organ deformation. This ensures that the safety map remains consistent with the actual intraoperative situation, providing accurate spatial references and deformation data for subsequent path deformation perception, risk assessment, and local rerouting, ensuring that the puncture path always avoids dynamically changing dangerous tissues, and improving the safety and precision of the surgery.

[0030] In particular, in one specific embodiment, Figure 5 This is a flowchart of step S2 in the deep learning-based robotic path planning method for precise prostate puncture according to an embodiment of this application. Figure 5 As shown, the initial 3D safety map is updated in real time based on the captured current ultrasound frame to obtain a real-time 3D safety map and a three-dimensional deformation field, including: S21, inputting the current ultrasound frame and preoperative MRI scan data into a trained unsupervised registration network to obtain a three-dimensional deformation field; S22, applying the three-dimensional deformation field to the initial 3D safety map to obtain the real-time 3D safety map.

[0031] Specifically, in step S21, the current ultrasound frame and the preoperative MRI scan data are input into the trained unsupervised registration network to obtain a three-dimensional deformation field. It should be understood that, since the preoperative MRI scan data and the intraoperative ultrasound frame belong to different imaging modalities, there are significant differences in gray scale features, resolution and coordinate systems between the two, and direct association between the two cannot accurately establish the spatial correspondence. Therefore, the current ultrasound frame and the preoperative MRI scan data are input into the trained unsupervised registration network in the present application. The trained unsupervised registration network is an improved model based on the VoxelMorph architecture, which adopts an encoder-decoder structure, wherein the decoder is symmetrical to the encoder. The deep features of the MRI and ultrasound images are extracted by gradually reducing the spatial dimensions, and the spatial dimensions are gradually restored by transposed convolution. Finally, a three-dimensional deformation field with the same size as the input image is output, so as to learn the internal correlation between the two modal images through the network and obtain a three-dimensional deformation field capable of describing the spatial mapping relationship from the preoperative MRI coordinate system to the intraoperative ultrasound coordinate system. In this way, the non-rigid deformation of the prostate and the surrounding tissues can be accurately quantified, and accurate spatial transformation basis is provided for subsequent safe map updating, so as to ensure that the position information of the anatomical structure in the operation is effectively associated with the preoperative data, and accurate deformation data basis is laid for path risk assessment.

[0032] In particular, in one possible embodiment, the implementation process of step S21 is as follows: first, the input data is preprocessed. The preoperative MRI scan data is defined as a fixed image, and its original gray scale information and coordinate system are retained. The current ultrasound frame is denoised and gray scale normalized to obtain a preprocessed ultrasound frame as a moving image. The prostate mask in the ultrasound frame is extracted by a lightweight segmentation network as an anatomical structure reference. Secondly, the preprocessed fixed image and the moving image are input into the trained unsupervised registration network, such as VoxelMorph. The network includes a feature extraction encoder and a deformation field decoder. The encoder extracts deep features of the two modal images through convolutional layers, and the decoder outputs an initial deformation field based on feature differences. During network training, image similarity loss is used to ensure gray scale feature alignment, and Dice loss of the prostate mask is used to ensure anatomical structure consistency, so that the output deformation field not only conforms to the image feature correlation, but also follows the anatomical structure continuity. Finally, a three-dimensional deformation field is output through network inference. The deformation field is a vector field with the same size as the fixed image. Each voxel position stores a three-dimensional displacement vector, which accurately describes the spatial offset of the position from the preoperative MRI coordinate system to the intraoperative ultrasound coordinate system, and completely reflects the non-rigid deformation such as translation, stretching and twisting of the prostate and the surrounding tissues.

[0033] Specifically, in step S22, the three-dimensional deformation field is applied to the initial 3D safety map to obtain the real-time 3D safety map. It should be understood that since the initial 3D safety map is constructed based on preoperative MRI scan data, the label information (such as forbidden area and target lesion position) thereof corresponds to the anatomical structure in the preoperative coordinate system, and the deformation of the prostate during the operation has caused the position of these structures to change. If the initial 3D safety map is continued to be used, the actual distribution of the safety area and the forbidden area cannot be reflected, which may cause deviation in the path evaluation. Therefore, the three-dimensional deformation field is further applied to the initial 3D safety map to migrate the label information in the initial map to the intraoperative coordinate system through the spatial mapping relationship of the deformation field, and generate the real-time 3D safety map reflecting the current anatomical structure distribution. In this way, the forbidden area, target lesion and safety area position in the safety map can be completely matched with the actual situation during the operation, providing an accurate spatial reference for path deformation perception and risk assessment, ensuring that the subsequent path correction can be based on the real safety constraints, and improving the safety of the puncture.

[0034] In particular, in one possible embodiment, the implementation process of step S22 is as follows: first, the initial 3D safety map is three-dimensional voxel label data, different integer labels correspond to different anatomical structures, such as 0 for background, 1 for prostate, 2 for urethra, 3 for nerve and blood vessel bundle, and 4 for target lesion, and the label position is determined based on the preoperative MRI coordinate system. Secondly, for each voxel position in the real-time 3D safety map, the corresponding source position thereof in the initial 3D safety map is queried according to the three-dimensional deformation field, that is, the original position of the current position in the preoperative coordinate system is calculated through the displacement vector in the deformation field. Subsequently, the initial map labels around the source position are sampled by using a trilinear interpolation algorithm, and the label value of the current voxel is determined according to the interpolation result, so as to ensure that the label information remains continuous after deformation. Finally, the label values of all voxels are integrated to generate the real-time 3D safety map. In the map, the positions of the forbidden areas such as urethra and nerve and blood vessel bundle have been updated according to the intraoperative deformation, and the current position of the target lesion is also accurately mapped to the intraoperative coordinate system, which fully reflects the safety distribution of the current anatomical structure.

[0035] In the aforementioned deep learning-based robotic path planning method for precise prostate puncture, step S3 involves performing path deformation perception and risk assessment on the initial path map based on a 3D deformation field and a real-time 3D safety map to obtain segmented risk score vectors and affected sub-maps. It should be understood that the real-time 3D safety map has updated the distribution of currently restricted areas. Without targeted deformation perception and risk assessment of the initial path map, it is impossible to determine which path segments become dangerous due to the non-rigid deformation of the prostate during surgery, potentially leading to collisions between the puncture path and restricted areas. Furthermore, global assessment significantly increases computational load and affects real-time performance during surgery. Therefore, this application further performs path deformation perception and risk assessment on the initial path map based on a 3D deformation field and a real-time 3D safety map to quantify the deformation degree and safety risk of each path segment, generate segmented risk score vectors, and accurately identify affected sub-maps requiring correction. This approach enables a refined assessment of path risks, focusing only on high-risk local path segments. It provides clear targets for subsequent local rerouting, significantly reducing computational overhead while ensuring assessment accuracy, meeting intraoperative real-time requirements, and ensuring the dynamic safety of the puncture path.

[0036] In particular, in one specific embodiment, Figure 6 This is a flowchart of step S3 in the deep learning-based robotic path planning method for precise prostate puncture according to an embodiment of this application. Figure 6 As shown, based on a three-dimensional deformation field and a real-time 3D safety map, path deformation perception and risk assessment are performed on the initial path map to obtain segmented risk score vectors and affected sub-maps, including: S31, based on the three-dimensional deformation field and the real-time 3D safety map, multi-dimensional feature extraction is performed on each path segment in the initial path map to obtain a segmented feature matrix; S32, the segmented feature matrix is ​​input into the path deformation perception module to obtain a segmented risk score vector; S33, based on the segmented risk score vector, the affected sub-map is identified from the initial path map.

[0037] Specifically, in step S31, based on the three-dimensional deformation field and the real-time 3D safety map, multi-dimensional feature extraction is performed on each path segment in the initial path map to obtain a segment feature matrix. It should be understood that the safety of each path segment in the initial path map is closely related to the deformation amplitude and the distance from the restricted area. Single-dimensional features (such as deformation only or distance only) cannot fully reflect the actual risk of a path segment, directly affecting the accuracy of subsequent risk assessment. Therefore, this application performs multi-dimensional feature extraction on each path segment in the initial path map based on the three-dimensional deformation field and the real-time 3D safety map, thereby integrating deformation features and risk features to form a segment feature matrix that comprehensively describes the risk status of each path segment. This provides rich and accurate input data for path risk assessment, ensuring that the risk of each path segment can be characterized from multiple dimensions, laying the foundation for accurate identification of high-risk areas and improving the reliability of risk assessment.

[0038] In particular, in one specific embodiment, Figure 7 This is a flowchart of step S31 in the deep learning-based robotic path planning method for precise prostate puncture according to an embodiment of this application. Figure 7 As shown, based on a three-dimensional deformation field and a real-time 3D safety map, multi-dimensional feature extraction is performed on each path segment in the initial path map to obtain a segment feature matrix, including: S311, defining the sampling space of the first path segment to obtain multiple sampling points; S312, extracting the displacement vectors of each sampling point from the three-dimensional deformation field to obtain multiple sampling point position vectors; S313, calculating the average magnitude and standard deviation of the magnitudes of the multiple sampling point position vectors as the deformation features of the first path segment; S314, calculating the minimum Euclidean distance between the first path segment and the nearest restricted area using the real-time 3D safety map as the risk feature of the first path segment; S315, combining the risk feature and the deformation feature of the first path segment to obtain the feature vector of the first path segment.

[0039] More specifically, in step S311, a sampling space for the first path segment is defined to obtain multiple sampling points. It should be understood that since the first path segment is a continuous line segment, its overall deformation and risk status cannot be fully reflected by information from a single location. Relying solely on information from the path endpoints would result in the loss of details in the intermediate areas, leading to inaccurate subsequent feature extraction. Therefore, this application defines a sampling space for the first path segment to obtain multiple sampling points, thereby capturing information from different locations within the path segment through multi-point sampling, comprehensively reflecting its overall state. This provides a sufficient sample base for the subsequent extraction of deformation and risk features, ensuring that the features truly reflect the actual situation of the path segment and improving the comprehensiveness and accuracy of feature extraction.

[0040] More specifically, step S312, the displacement vector of each sampling point is extracted from the three-dimensional deformation field to obtain a plurality of sampling point position vectors. It should be understood that the position change of the sampling point in the operation cannot be directly obtained, and such displacement is the core basis for evaluating the deformation degree of the path segment, and the lack of displacement information will lead to the loss of data basis for deformation feature extraction. Therefore, the present application further extracts the displacement vector of each sampling point from the three-dimensional deformation field to obtain a plurality of sampling point position vectors, so as to obtain the specific displacement information of each sampling point through the deformation field and quantify its spatial offset. In this way, the original data can be provided for subsequent calculation of the deformation feature, so that the deformation feature can accurately reflect the offset amplitude and direction of each position of the path segment, and provide a reliable deformation basis for risk assessment, thereby improving the systematicness and comparability of deformation assessment.

[0041] In particular, in one possible embodiment, the implementation process of step S312 is as follows: first, the coordinates of each sampling point in the first path segment sampling space are converted to the preoperative MRI coordinate system, so as to ensure consistency with the coordinate system of the three-dimensional deformation field. Secondly, the three-dimensional deformation field is a dense vector field aligned with the preoperative MRI coordinate system, and each voxel position stores a three-dimensional displacement vector (Δx, Δy, Δz) representing the spatial offset of the position from preoperative to intraoperative. For each sampling point, its preoperative coordinates are located in the three-dimensional deformation field, and the displacement vector corresponding to the point is obtained through a trilinear interpolation algorithm, and if the sampling point is located in the interstitial space, the displacement vectors of the surrounding voxels are interpolated to obtain the accurate value. Finally, the displacement vector of each sampling point is taken as its position vector, which directly describes the spatial offset of the sampling point in the operation, and the position vectors of all sampling points together constitute a set of a plurality of sampling point position vectors.

[0042] More specifically, step S313, the average modulus length and the standard deviation of the modulus length of the plurality of sampling point position vectors are calculated as the deformation feature of the first path segment. It should be understood that since the plurality of sampling point position vectors can only reflect the local displacement of each point, they cannot comprehensively reflect the overall deformation degree and deformation uniformity of the first path segment, and single local displacement information is difficult to use for risk comparison between path segments. Therefore, the present application calculates the average modulus length and the standard deviation of the modulus length of the plurality of sampling point position vectors as the deformation feature of the first path segment, integrates the local displacement information through the statistical quantity, and obtains a quantitative index that can reflect the overall deformation state. In this way, the dispersed displacement data can be converted into intuitive deformation features, so that the deformation degrees of different path segments can be quantitatively compared, an explicit deformation basis is provided for subsequent risk assessment, and the systematicness and comparability of deformation assessment are improved.

[0043] More specifically, in step S314, the minimum Euclidean distance between the first path segment and the nearest forbidden zone is calculated by the real-time 3D safety map as the risk feature of the first path segment. It should be understood that since the safety of the first path segment depends on its distance from the forbidden zone, the closer the distance, the higher the risk of collision, but if only visual judgment cannot quantify this risk, it is difficult to make an accurate risk comparison. Therefore, the present application further calculates the minimum Euclidean distance between the first path segment and the nearest forbidden zone by the real-time 3D safety map as the risk feature of the first path segment, so as to convert the spatial relationship between the path segment and the forbidden zone into a quantitative distance index, and intuitively reflect its safety margin. In this way, a clear quantitative basis can be provided for risk assessment, so that the safety levels of different path segments can be directly compared, and high-risk segments close to the forbidden zone can be quickly identified, thereby improving the accuracy of risk assessment.

[0044] In particular, in one possible embodiment, the implementation process of step S314 is as follows: first, the various types of forbidden zones are marked in the real-time 3D safety map, such as urethral markers marked as 2 and nerve and vascular bundles marked as 3, and binary processing is performed thereon, with the voxel values of all forbidden zones set to 1 and the remaining areas set to 0 to obtain a forbidden zone mask. Second, three-dimensional distance transformation is performed on the forbidden zone mask to generate a distance transformation map, in which the value of each voxel represents the Euclidean distance from the voxel to the nearest forbidden zone voxel. Third, the distance values of the path segment are sampled. A plurality of points are uniformly selected on the first path segment, consistent with the sampling points of the sampling space, and the distance values corresponding to these points are extracted from the distance transformation map to obtain a plurality of distance samples. Finally, the minimum value is selected from the plurality of distance samples, which is the minimum Euclidean distance between the first path segment and the nearest forbidden zone, serving as the risk feature thereof.

[0045] More specifically, in step S315, the risk feature of the first path segment and the deformation feature of the first path segment are combined to obtain a first path segment feature vector. It should be understood that since the risk feature only reflects the static distance of the path segment from the forbidden zone and the deformation feature only reflects the dynamic offset state thereof, a single feature cannot fully reflect the comprehensive risk of the path segment, which may lead to one-sided risk assessment. Therefore, the present application arranges the features in the order of deformation feature first and risk feature second, i.e., the average module length and the module length standard deviation are placed first, and then the minimum Euclidean distance is placed, forming a three-dimensional feature vector, and finally, the above three feature values are combined in order to obtain a first path segment feature vector, so as to integrate the static safety distance and the dynamic deformation information, and form a comprehensive feature that can fully describe the risk state of the path segment. In this way, a multi-dimensional information-containing input can be provided for the path deformation perception module, so that the module can more comprehensively assess the risk, improve the accuracy and reliability of the risk score, and provide a more comprehensive basis for the subsequent identification of affected subgraphs.

[0046] Specifically, step S32, the segmented feature matrix is input into the path deformation perception module to obtain a segmented risk score vector. It should be understood that since the segmented feature matrix only contains multi-dimensional feature data of each path segment, the correlation between segments, such as the mutual influence of the risks of adjacent segments, is not embodied, and direct use of these features cannot obtain a quantitative score that reflects the overall risk distribution, making it difficult to accurately distinguish high-risk and low-risk segments. Therefore, the segmented feature matrix is input into the path deformation perception module, so as to fuse the features and mutual correlation of each path segment through the context perception ability of the path deformation perception module, and output a segmented risk score vector that can quantify the comprehensive risk of each segment. In this way, the multi-dimensional features can be converted into intuitive risk scores, so that the risk of each path segment can be quantitatively compared, facilitating the rapid identification of segments with higher risks, providing clear basis for the subsequent identification of affected subgraphs, and improving the efficiency and accuracy of risk assessment.

[0047] In particular, in one possible embodiment, the implementation process of step S32 is as follows: first, the segmented feature matrix containing multi-dimensional features of each path segment is input into the path deformation perception module, the core of which is a pre-trained small Transformer encoder model. The training of the model is completed on a large number of offline simulation data sets. In the training, the model receives a simulated path segment feature matrix as input and is required to output a risk score vector. The vector is compared with a pre-calculated true risk label, for example, a label generated based on whether the segment collides with the forbidden area after simulation deformation, and the model parameters are continuously optimized through supervised learning, such as using mean square error or cross-entropy loss function, so that it learns to accurately predict risks from multi-dimensional features. Secondly, the self-attention mechanism of the Transformer encoder calculates the correlation between each path segment feature vector and all other segment feature vectors. For example, if the adjacent segments of a path segment deform dramatically, the self-attention mechanism will increase the attention to it, so that the encoded features can reflect the associated risks between segments. Thirdly, the encoded feature matrix is processed by a feedforward neural network, and each path segment feature is mapped to a scalar value in the interval [0, 1] through a Sigmoid activation function. The value is the comprehensive risk score of the path segment, and the higher the score, the higher the risk. Finally, the risk scores of all path segments are arranged in the order of their positions in the initial path graph to form a segmented risk score vector, and each element of the vector corresponds to the risk level of a path segment.

[0048] In particular, when the segmented feature matrix is input into the path deformation perception module to obtain the segmented risk score vector, it needs to be noted that the segmented feature vectors in the segmented feature matrix, such as row vectors, are not permutation invariant, that is, the correlation between the row vectors depends not only on the feature distribution itself, but also needs to consider the sequence order or spatial topological relationship between the row vectors, for example, in the path planning scene, segments E1 and E2 are adjacent, while E1 and E9 are remote, and such topological relationship contains important physical information. Moreover, in the segmentation risk scoring, the propagation of deformation and the correlation of risk are also related to the distance in physics, that is, the influence of the severe deformation of a segment on its directly adjacent segments is much greater than that on the end segments of the path.

[0049] Therefore, in another possible preferred embodiment, the segmented feature matrix needs to be geometrically perceived relative position coding to obtain an updated segmented feature matrix before being input into the path deformation perception module, that is, the relative geometric relationship information between path segments is introduced. First, the geometric distance between path segments is defined, and the distance matrix is calculated according to the geometric distance, so as to accurately depict the connection relationship and influence range of each segment in the physical space, so that the distance matrix reflects the adjacency degree between the path segments in the operation, provides a basis for subsequent fusion of spatial information to evaluate the risk, and ensures that the risk evaluation can be fitted to the actual interaction of the path segments when the prostate deforms. Specifically, assuming is the Euclidean length of the path segment , is the midpoint of the path segment , the geometric distance between the path segment and the path segment is defined as the cumulative length along the path from to : ; wherein represents the calculation of the Euclidean length , represents the path segment , represents the path segment , is the summation index, that is, the path segment from to , represents the path segment , represents the geometric distance between the path segment and the path segment , This represents the actual physical distance that risk or deformation information needs to travel along the path, and it is expected to have higher resolution for short distances: ;in, This is the converted distance value. It is the floor symbol. yes The value, It is the resolution step size at close range. It focuses on distance; anything beyond this distance is considered very far. For example, setting... It is 1mm. The value is 16mm. The resulting distance matrix is... ,in Representative segment The distance vector, whose vector value is... The values ​​are inversely correlated, that is... The smaller the numerical value, the larger the vector value.

[0050] Then, based on the segmented feature matrix and the distance matrix, the weighted score of each row vector of the segmented feature matrix is ​​calculated, which includes the content score and the geometric position bias: ;in, Indicates segmentation The weighted score, Indicates segmentation For segmentation The weighted score, Indicates the first The feature vector of the segment, Indicates the first The feature vector of the segment, This represents the transpose operator. Representing vectors sum vector European distance, Indicates segmentation The distance vector. Here, for each row vector of the segmented feature matrix, it is transformed by linear interpolation into a distance vector. Same length, and therefore, Represents the state of being Path segmentation It should be for a distance from it. Segmentation The degree of innate attention bias is assigned to clarify the weight of each segment's impact on the current segment's risk. For example, when... When smaller, The corresponding embedding vector value is large, making the bias term positive and large, thus forcing the model to pay close attention to neighboring segments. And when... When the bias term is larger, the value of the bias term is smaller or even close to zero, indicating that the information of the remote segment is ignored. In this way, the weighted score can accurately reflect the superimposed influence of the adjacent high deformation segment on the segment close to the forbidden zone, so that the risk assessment is more in line with the actual situation of the intraoperative path affected by the dynamic deformation of the prostate.

[0051] Then, each row vector of the segment feature matrix is weighted based on the weighted score of each row vector of the segment feature matrix to obtain an updated segment feature matrix, that is, each row vector of the segment feature matrix is multiplied by the corresponding weighted score to obtain an updated segment feature matrix, so that the feature matrix contains the risk correlation information of the whole path, and the updated feature matrix can accurately reflect the risk transmission of each segment of the intraoperative path due to the deformation of the prostate, for example, the characteristics of a segment are integrated into the influence of its adjacent high-risk segment, providing a comprehensive basis for accurate risk assessment.

[0052] Finally, the updated segment feature matrix is input into the path deformation perception module to obtain a segment risk score vector, so that the comprehensive characteristics are converted into quantifiable risk values, and the robot can quickly locate the high-risk path segment caused by the dynamic deformation of the prostate during the operation, for example, a significant increase in the score of a segment indicates that it is in a dangerous state due to the combined effect of adjacent segments and its own deformation, providing a clear target for subsequent local rerouting.

[0053] In this way, the geometric topological structure (adjacency relationship, segment length) of the path is directly and explicitly encoded into the weighted score, that is, the cumulative length along the path is used as a measure of relative position, which can better reflect the propagation of deformation and stress along the medium and provide more rich and accurate prior knowledge for the model. At the same time, since the geometric position bias depends on the state of the current segment, the model can learn more complex rules, and the risk assessment of the model will have more context awareness. For example, if both neighbors of a segment are in a high-risk state, the model can perceive this situation through the geometric bias term and appropriately increase the risk score of the segment.

[0054] Specifically, in step S33, the affected subgraph is identified from the initial path graph based on the segment risk score vector. Specifically, the affected subgraph is identified from the initial path graph based on the segment risk score vector, so that the high-risk segment and its associated adjacent segments are integrated into a local subgraph, and the range that needs to be corrected is clearly defined. In this way, it can be ensured that the correction range contains not only the high-risk area but also the continuity of the path, so that the subsequent local rerouting can generate a new path that smoothly connects with the surrounding path, while avoiding the high time consumption of global processing, meeting the real-time requirements during the operation, and improving the effectiveness and efficiency of path correction.

[0055] In particular, in one possible embodiment, the implementation of step S33 is as follows: first, a risk threshold is set, such as 0.7, and the segmented risk score vector is traversed, and the path segments with scores higher than the threshold are marked as core high-risk segments, which are the areas in the path most likely to have safety hazards. Second, the path points connected at both ends of the core high-risk segments are marked as core high-risk nodes, which are the key endpoints of the high-risk segments. Third, to ensure the smoothness of the modified path, the path points directly adjacent to the core high-risk nodes are included in the affected node set, and the path points directly adjacent to the core high-risk nodes are connected by a path segment, ensuring that the correction range covers the high-risk area and the surrounding parts that need to be connected. Finally, all nodes in the affected node set are extracted from the initial path graph, as well as all path segments connecting these nodes, including core high-risk segments and adjacent segments, which together with the nodes constitute the affected subgraph, which is the object of subsequent local rerouting.

[0056] In the above-mentioned prostate precise puncture robot path planning method based on deep learning, step S4, local path rerouting based on graph neural network is performed on the affected subgraph to obtain a new local path. It should be understood that, since the affected subgraph contains high-risk path segments and nodes in the initial path graph, the original path may become unsafe due to prostate deformation, and directly following or simple adjustment cannot ensure the safety and smoothness of the new path, while global path re-planning will bring too high a computational load, making it difficult to meet the real-time requirements during surgery. Therefore, the application further performs local path rerouting based on graph neural network on the affected subgraph, so as to utilize the efficient processing capability of graph neural network for graph structure data to generate a new local path that avoids forbidden areas and smoothly connects with the surrounding path in the affected area. In this way, only the high-risk area can be accurately corrected, greatly reducing the computational load to ensure real-time performance, and through the learning ability of the graph neural network, the safety and continuity of the new path are ensured, laying a foundation for subsequent path integration.

[0057] In particular, in one possible embodiment, the implementation of step S4 is as follows: first, the affected subgraph and its node features are input into a pre-trained graph neural network model, such as the graph attention network GAT. The nodes of the subgraph are divided into two categories: internal nodes and boundary nodes, i.e., nodes connected to the unaffected part of the initial path graph. The features of each node at least include its original three-dimensional coordinates, the risk score from the path deformation perception and risk assessment module, and the distance to the nearest no-go zone in the real-time 3D safety map. Second, the graph neural network model performs a forward propagation calculation on the input subgraph through its pre-trained weights. In the training stage, the model has learned how to modify the dangerous path to a safe path on a large amount of simulation data, and its training target is designed to optimize three aspects at the same time: maximizing the distance between the new path and the no-go zone, minimizing the total length of the new path, and ensuring the smoothness of the path, i.e., minimizing the curvature change, especially ensuring smooth connection with the boundary nodes. Third, the model directly predicts a new three-dimensional coordinate for each internal node in the affected subgraph. The coordinates of the boundary nodes remain unchanged to ensure that the newly generated local path can be seamlessly spliced back to the unaffected part of the original path graph. Finally, the newly generated internal node coordinates and fixed boundary node coordinates are connected in order to form a new local path that is both safe and smooth.

[0058] In the above deep learning-based prostate precise puncture robot path planning method, step S5, the initial path graph is integrated based on the new local path to obtain a modified path graph. It should be understood that, since the affected subgraph is a local area in the initial path graph that has safety risks due to prostate deformation, its original path cannot adapt to the real-time anatomical structure during the operation, and the new local path is a safe and optimized path generated for this area. If the two are not integrated, the overall path will be broken or not smoothly connected, which cannot form a complete path that can be directly used for robot puncture, thereby affecting the accuracy and safety of the operation. In one specific example of the present application, the initial path graph is integrated based on the new local path to obtain a modified path graph: the new local path is used to replace the affected subgraph in the initial path graph to obtain the modified path graph. That is, the present application further integrates the initial path graph based on the new local path, seamlessly fuses the safe local modified path with the unaffected part of the initial path graph by replacing the affected subgraph with the new local path, and forms a modified path graph that eliminates risks while maintaining overall coherence. In this way, the modified path graph can completely cover the entire process from the puncture entry to the target lesion, and each path segment meets the real-time safety constraints, while avoiding repeated calculation of the unaffected area, significantly improving the efficiency of intraoperative path correction, providing precise and continuous execution basis for the robot, and ensuring the smoothness and safety of the operation.

[0059] In particular, in one possible embodiment, the implementation of step S5 is as follows: first, locate the specific range of the affected subgraph. In the initial path graph, all nodes (such as path points P5 to P10) and edges (such as path segments E5 to E9) contained in the affected subgraph are explicitly identified, and the three-dimensional coordinates, numbers of the nodes, and connection relationships of the edges are recorded to ensure that the replacement area is accurate and does not involve unaffected path parts. Second, process the boundary nodes to ensure the stability of the connection. Identify the boundary nodes of the affected subgraph, i.e., the nodes directly connected to the non-affected part in the initial path graph, such as P5 and P10, which are the key to the connection between the new local path and the original path; adjust the coordinates of the corresponding boundary nodes in the new local path to be exactly the same as those in the initial path graph, ensuring that there is no spatial deviation at the connection, and avoiding jumps or angle mutations in the path. Third, replace the nodes and edges in the affected area. Replace the internal nodes (such as P6', P7', P8', P9') and edges (such as E5', E6', E7', E8') in the new local path one by one to replace the corresponding nodes (P6, P7, P8, P9) and edges (E5, E6, E7, E8) in the affected subgraph of the initial path graph; for the nodes in the initial path graph that are not affected, such as P1 to P4, P11 and subsequent nodes, and the edges, such as E1 to E4, E9 and subsequent edges, keep their coordinates and connection relationships unchanged to ensure the stability of the path in the non-affected area. Finally, verify the continuity and safety of the corrected path graph. Check the smoothness of the edges at the connection between the new local path and the non-affected path, such as E4 and E5', E8' and E9, by calculating the included angle (e.g., ensuring that it does not exceed a predetermined threshold, such as 8°) and the curvature change (ensuring that it is within a predetermined safe range) of the connecting edges to confirm that the path transitions naturally. At the same time, verify whether all the replaced edges are completely within the safety area without any overlap with the forbidden area by combining the real-time 3D safety map. After the above steps, the initial path graph is updated to the corrected path graph, which can be directly used as the execution path for the robot to puncture.

[0060] In summary, the prostate precise puncture robot path planning method based on deep learning according to the embodiments of the present application is illustrated as follows. Firstly, global path planning is performed based on MRI data before surgery, and the puncture environment and the planned path are abstracted into a graph structure model. In surgery, the dynamic deformation of the prostate is perceived through real-time ultrasound images, and the affected local dangerous area in the initial path graph is accurately identified based on the deformation information. Subsequently, the graph neural network technology is introduced, and the affected local subgraph is quickly and intelligently rerouted to generate a new safe local path. Finally, the new path is seamlessly integrated into the original global path graph, and the efficient correction of the overall path is completed. In this way, the time-consuming bottleneck of global calculation can be avoided, thereby solving the real-time problem of intraoperative path dynamic correction under the premise of ensuring high precision, and significantly improving the path planning capability of the robot in a dynamic environment.

[0061] As described above, the prostate precise puncture robot path planning method based on deep learning according to the embodiments of the present application can be implemented in various wireless terminals. In a possible implementation, the prostate precise puncture robot path planning method based on deep learning according to the embodiments of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the prostate precise puncture robot path planning method based on deep learning can be a software module in the operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the prostate precise puncture robot path planning method based on deep learning can also be one of the many hardware modules of the wireless terminal.

[0062] Alternatively, in another example, the prostate precise puncture robot path planning method based on deep learning can also be a separate device from the wireless terminal, and the prostate precise puncture robot path planning method based on deep learning can be connected to the wireless terminal through a wired and / or wireless network, and transmit interactive information in an agreed data format.

Claims

1. A deep learning-based method for robotic path planning in precise prostate puncture, characterized in that, include: Preoperative global path generation and graph structuring were performed on preoperative MRI scan data to obtain an initial path map and an initial 3D safety map. The initial 3D safety map is updated in real time based on the captured current ultrasound frame to obtain a real-time 3D safety map and a three-dimensional deformation field. Based on the three-dimensional deformation field and the real-time 3D safety map, path deformation perception and risk assessment are performed on the initial path map to obtain segmented risk score vectors and affected sub-maps. Local path rerouting based on graph neural networks is performed on the affected sub-maps to obtain new local paths. Based on the new local paths, the initial path map is integrated to obtain a corrected path map. Specifically, based on a three-dimensional deformation field and a real-time 3D safety map, path deformation perception and risk assessment are performed on the initial path map to obtain segmented risk score vectors and affected sub-maps. This includes: extracting multi-dimensional features from each path segment in the initial path map based on the three-dimensional deformation field and the real-time 3D safety map to obtain a segmented feature matrix; inputting the segmented feature matrix into the path deformation perception module to obtain a segmented risk score vector; and identifying the affected sub-maps from the initial path map based on the segmented risk score vectors. Specifically, based on a three-dimensional deformation field and a real-time 3D safety map, multi-dimensional feature extraction is performed on each path segment in the initial path map to obtain a segment feature matrix. This includes: defining a sampling space for the first path segment to obtain multiple sampling points; extracting the displacement vectors of each sampling point from the three-dimensional deformation field to obtain multiple sampling point position vectors; calculating the average magnitude and standard deviation of the magnitudes of the multiple sampling point position vectors as the deformation features of the first path segment; calculating the minimum Euclidean distance between the first path segment and the nearest restricted area using the real-time 3D safety map as the risk feature of the first path segment; and combining the risk feature and the deformation feature of the first path segment to obtain the feature vector of the first path segment.

2. The deep learning-based robotic path planning method for precise prostate puncture as described in claim 1, characterized in that, Preoperative global path generation and graph structuring of preoperative MRI scan data to obtain an initial path map and an initial 3D safety map include: performing semantic segmentation on preoperative MRI scan data based on a pre-trained 3D U-Net model to obtain an initial 3D safety map; inputting the initial 3D safety map into an optimal path planning module to obtain the original optimal path; and converting the original optimal path into an initial path map.

3. The deep learning-based robotic path planning method for precise prostate puncture as described in claim 2, characterized in that, Transforming the original optimal path into an initial path graph includes: discretizing the original optimal path into a series of path points; using the path points as nodes of the initial path graph, and using the connections between adjacent path points as edges of the initial path graph.

4. The deep learning-based robotic path planning method for precise prostate puncture as described in claim 1, characterized in that, The process of updating the initial 3D safety map in real time based on the captured current ultrasound frame to obtain the real-time 3D safety map includes: inputting the current ultrasound frame and preoperative MRI scan data into a trained unsupervised registration network to obtain a three-dimensional deformation field; and applying the three-dimensional deformation field to the initial 3D safety map to obtain the real-time 3D safety map.

5. The deep learning-based robotic path planning method for precise prostate puncture as described in claim 1, characterized in that, The process of integrating the initial path graph with the new local path to obtain the corrected path graph includes: replacing the affected subgraphs in the initial path graph with the new local path to obtain the corrected path graph.

Images