Ai visualized boundary navigation system for dynamic resection of brain tumors

Abstract

The application relates to the technical field of medical operation navigation, and particularly discloses an AI visualized boundary navigation system for dynamic resection of brain tumors, which comprises a positioning and collecting module, which is used for collecting space position data of a surgical instrument, space coordinates of patient head markers, real-time images of brain tissues during operation and discrete tissue feature point coordinates, and obtaining a preoperative planning model containing an initial virtual tumor boundary, key anatomical structures and a surgical path. The technical scheme of the application physically separates a master tracking thread responsible for rendering of a surgical instrument cursor and a deformation calculation thread responsible for complex calculation in priority. In actual surgical resection operation, the architecture guarantees smooth operation of the instrument tracking picture and is not disturbed by the process of space deformation operation.

Description

Technical Field

[0001] This invention relates to the field of medical surgical navigation technology, specifically to an AI-based visual boundary navigation system for dynamic resection of intracranial tumors. Background Technology

[0002] During craniocerebral tumor resection surgery, the brain undergoes nonlinear spatial deformation due to brain tissue removal and cerebrospinal fluid loss. Existing surgical navigation systems primarily rely on preoperatively planned 3D images for spatial positioning, failing to synchronously reflect intraoperative tissue deformation, resulting in discrepancies between the virtual boundaries displayed by the system and the actual anatomical structures. Some existing technologies attempt to introduce deformation calculation models to correct this spatial deviation, but face limitations in the underlying architecture's processing logic during actual operation.

[0003] In specific intraoperative navigation scenarios, the spatial positioning and tracking of surgical instruments requires high-frequency refreshes to maintain the continuity of the display interface, while the calculation process for three-dimensional deformation compensation of brain tissue requires a large amount of computing resources and has a long computation cycle. Existing single-processor architectures cannot handle these two tasks simultaneously, resulting in delays in the tracking of surgical instruments during deformation calculation.

[0004] Furthermore, existing deformation prediction technologies lack objective reliability assessment mechanisms, are prone to outputting prediction results that deviate from reality in areas with complex deformation, and fail to demonstrate the reliability of the prediction results to the operator in an intuitive interface. Summary of the Invention

[0005] This invention aims to at least partially address one of the technical problems in related technologies. Therefore, the objective of this invention is to propose an AI-based visual boundary navigation system for dynamic resection of intracranial tumors, thereby reducing the risk of anatomical structure localization due to tissue deformation.

[0006] To achieve the above objectives, a first aspect of the present invention provides an AI-based visual boundary navigation system for dynamic resection of intracranial tumors, comprising:

[0007] The positioning and acquisition module is used to acquire in real time the spatial pose data of surgical instruments, the spatial coordinates of patient head markers, real-time images of intraoperative brain tissue and coordinates of discrete tissue feature points, as well as to obtain a preoperative planning model containing the initial virtual tumor boundary, key anatomical structures and surgical path.

[0008] The navigation controller includes a main control tracking thread configured with a first preset priority, and an AI deformation calculation thread that runs asynchronously with the main control tracking thread. The main control tracking thread is used to render surgical instrument cursors. The AI ​​deformation calculation thread is used to calculate global three-dimensional nonlinear offsets to generate a spatial elastic deformation compensation field matrix, calculate deformation prediction confidence through multimodal cross-validation to generate a corresponding confidence field matrix, and detect target update regions with deformation prediction confidence lower than a first confidence threshold based on the confidence field matrix, so as to perform local deformation field recalculation.

[0009] The coordinate mapping module is used to perform local interpolation remapping based on the spatial pose data of the surgical instruments and the spatial elastic deformation compensation field matrix, so as to dynamically update the initial virtual tumor boundary and generate the current virtual tumor boundary.

[0010] The intelligent AR display module is used to perform hierarchical visualization rendering of the current virtual tumor boundary based on the confidence field matrix.

[0011] To achieve the above objectives, a second aspect of the present invention proposes an AI-based visual boundary navigation method for dynamic resection of intracranial tumors, comprising:

[0012] Real-time acquisition of spatial pose data of surgical instruments, spatial coordinates of patient head markers, real-time images of intraoperative brain tissue and coordinates of discrete tissue feature points, as well as acquisition of a preoperative planning model including initial virtual tumor boundaries, key anatomical structures and surgical paths.

[0013] A master tracking thread configured with a first preset priority is run to render surgical instrument cursors, and an AI deformation solving thread that is asynchronous with the master tracking thread is run.

[0014] The AI ​​deformation calculation thread calculates the global three-dimensional nonlinear offset to generate a spatial elastic deformation compensation field matrix. The deformation prediction confidence is calculated through multimodal cross-validation to generate a corresponding confidence field matrix. Based on the confidence field matrix, the target update region with a deformation prediction confidence lower than the first confidence threshold is detected to perform local deformation field recalculation.

[0015] Based on the spatial pose data of the surgical instruments, local interpolation and remapping are performed in conjunction with the spatial elastic deformation compensation field matrix to dynamically update the initial virtual tumor boundary and generate the current virtual tumor boundary.

[0016] The current virtual tumor boundary is hierarchically visualized and rendered based on the confidence field matrix.

[0017] To achieve the above objectives, a third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described AI-based visual boundary navigation method for dynamic resection of intracranial tumors.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] The technical solution of this invention constructs a master-slave asynchronous dual-track operation architecture, physically separating the main control tracking thread responsible for rendering surgical instrument cursors from the deformation calculation thread responsible for complex calculations in terms of priority. In actual surgical resection operations, this architecture ensures the smooth operation of the instrument tracking screen, unaffected by the process of spatial deformation calculations.

[0020] Meanwhile, this scheme generates a confidence field matrix through multimodal cross-validation, quantitatively evaluates the reliability of deformation prediction results, and automatically triggers local recalculation of the target update region accordingly. Subsequently, it dynamically updates the virtual tumor boundary using coordinate mapping technology, and finally performs hierarchical visualization rendering. This allows operators to objectively distinguish the reliability of navigation boundaries in different regions through the visual display features of the augmented reality interface, and to refer to the locally corrected virtual tumor boundary during operation, reducing the risk of anatomical structure localization caused by tissue deformation. Attached Figure Description

[0021] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0022] Figure 1 This is a schematic diagram illustrating the implementation of the AI-based visual boundary navigation system for dynamic resection of intracranial tumors provided by the present invention.

[0023] Figure 2 This is an adaptive frequency time series diagram of the master-slave asynchronous dual-track operation architecture in the AI ​​visualization boundary navigation system for dynamic resection of intracranial tumors provided by this invention;

[0024] Figure 3 This is a schematic diagram of the three-dimensional distribution of the grid density of the four-layer spatial elastic deformation compensation field matrix in the AI ​​visualization boundary navigation system for dynamic resection of intracranial tumors provided by the present invention.

[0025] Figure 4 The present invention provides a nonlinear convergence curve and Pareto front plot of the weighted constraint energy function for the recalculation of the local deformation field in the AI ​​visualization boundary navigation system for dynamic resection of intracranial tumors.

[0026] Figure 5 This invention provides a three-dimensional reconstructed bitmap of the absolute geometric boundary Poisson surface of a cavity based on instrument trajectory in an AI-visualized boundary navigation system for dynamic resection of intracranial tumors.

[0027] Figure 6 This invention provides a combined diagram of the centripetal collapse nonlinear response surface and three-dimensional vector field in the AI ​​visualization boundary navigation system for dynamic resection of intracranial tumors.

[0028] Figure 7 This is a time series curve of the evolution of two-stage physical compression deformation of brain tissue in the AI ​​visualization boundary navigation system for dynamic resection of intracranial tumors provided by the present invention.

[0029] Figure 8 This is a flowchart illustrating the AI-based visual boundary navigation method for dynamic resection of intracranial tumors provided by the present invention.

[0030] Figure 9 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0031] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0032] The following description, with reference to the accompanying drawings, describes an AI-based visual boundary navigation method, system, and electronic device for dynamic resection of intracranial tumors according to embodiments of the present invention.

[0033] Example 1:

[0034] This embodiment provides an AI-based visual boundary navigation system for dynamic resection of intracranial tumors. This system is mainly used in microsurgical neurosurgery and aims to solve the problem of spatial displacement of brain tissue caused by cerebrospinal fluid loss, tumor tissue removal, and gravity during surgical resection.

[0035] The system provided in this embodiment relies on an industrial-grade computing hardware platform and is equipped with a deeply customized real-time operating system kernel to ensure low latency and deterministic response in underlying data processing. Specifically, the system in this embodiment includes the following:

[0036] For example, the system provided in this embodiment includes a positioning and acquisition module. This positioning and acquisition module is the data input source of the entire navigation system, responsible for establishing a coordinate mapping bridge between the physical surgical space and the virtual computing space. In the actual physical hardware deployment, this module integrates a near-infrared binocular stereo vision tracking device, an ultrasound image acquisition interface, and a medical probe interface. The positioning and acquisition module is used to acquire the spatial pose data of surgical instruments in real time. This spatial pose data is represented by a six-degree-of-freedom vector, containing three translational coordinate components and three rotational attitude components, to ensure that the absolute position and orientation of the surgical instrument tip in physical space can be accurately restored. At the same time, this module is also used to acquire the spatial coordinates of the patient's head marker in real time. The patient's head marker is usually fixed on a skull fixation frame, serving as a dynamic reference benchmark for the entire surgical tracking process, used to eliminate the interference of slight changes in patient position on tracking accuracy.

[0037] Optionally, to achieve intraoperative fusion of multimodal data, the localization and acquisition module is also used to acquire real-time images and discrete tissue feature point coordinates of the brain tissue during surgery. Real-time images of the brain tissue are typically acquired by an intraoperative ultrasound probe calibrated in spatial location, reflecting the real-time echogenic texture characteristics within the brain tissue. Discrete tissue feature point coordinates are obtained by the surgeon using a cortical surface tracking probe with an optical reflector ball, performing contact sampling on the exposed surface of the cerebral cortex or the cavity wall after tumor resection.

[0038] In addition, the localization and acquisition module also has data parsing and import functions to obtain a preoperative planning model that includes the initial virtual tumor boundary, key anatomical structures, and surgical path. This preoperative planning model is a static digital twin model extracted from the patient's preoperative MRI or computed tomography data through 3D reconstruction and semantic segmentation algorithms. The key anatomical structures include major intracranial arteries, venous sinuses, and important motor nerve fiber bundles.

[0039] Specifically, the core of this embodiment is the navigation controller, which undertakes the entire scheduling task from data parsing and state tracking to complex deformation calculation. To solve the hardware resource contention problem between high-frequency optical tracking and computationally intensive deformation calculation, the navigation controller is built with a master-slave asynchronous dual-track architecture at the underlying level. This architecture includes two independent threads that interact with each other through shared memory: the master tracking thread and the AI ​​deformation calculation thread.

[0040] It is important to note that the main control tracing thread is configured with the first preset priority, which is set as the highest level of real-time interrupt response priority in the operating system kernel's scheduling policy. The core responsibility of the main control tracing thread is to ensure the high-frequency smoothness of visual navigation; it only performs lightweight coordinate transformation and rendering command issuance tasks for rendering surgical instrument cursors.

[0041] In actual operation, the main control tracing thread runs at a first preset frequency, which is typically set to be greater than or equal to 120 Hz to match the smooth perception limit of human vision. The main control tracing thread reads the latest spatial elastic deformation compensation field matrix and confidence field matrix from the shared memory at the operating system's underlying layer. This shared memory employs a lock-free concurrent read mechanism, avoiding read-write lock waiting between threads through a circular buffer. After reading the matrix data, the main control tracing thread uses rigid matrix multiplication to convert the physical coordinates of the surgical instrument tip into virtual scene coordinates with deformation compensation, thereby completing the real-time rendering and positioning of the cursor.

[0042] For example, the AI ​​deformation solving thread runs asynchronously with the main tracking thread. This thread is configured with normal computation priority and is specifically responsible for handling complex nonlinear partial differential equations and tensor operations. The core function of the AI ​​deformation solving thread is to solve the global three-dimensional nonlinear offset to generate a spatial elastic deformation compensation field matrix. This matrix records the vector data of the offset of each discrete voxel node in the virtual space from its preoperative static position to its intraoperative dynamic position. Since deformation solving is a dynamic and continuous process, the AI ​​deformation solving thread adopts an adaptive frequency operation mechanism. During the normal surgical microscope observation phase, due to the slow tissue deformation, the thread runs at a lower base frequency; when the system detects that the rate of change of the spatial pose data of the surgical instruments exceeds the first rate of change threshold, it indicates that the surgeon is performing intensive tissue resection or traction operations. At this time, the AI ​​deformation solving thread will automatically increase its running frequency to a second preset frequency.

[0043] like Figure 2 This visually demonstrates the underlying thread scheduling response and computing power allocation process of the navigation controller during dynamic resection. The horizontal axis of the figure represents the operation time in seconds, the left vertical axis represents the operating frequency in Hertz, and the right vertical axis represents the rate of change of spatial pose data in millimeters per second.

[0044] The blue curve in the figure represents the frequency of the main control tracking thread, which remains slightly fluctuating around 120 Hz throughout the entire 60-second monitoring cycle. This high-frequency stable state, unaffected by the complex deformation calculation process, ensures the smoothness of the surgical instrument cursor rendering.

[0045] The green curve represents the rate of change of spatial pose data acquired by the optical positioning module. During the period from 20 to 40 seconds, due to the operator's intensive tissue removal and instrument traction movements, the green curve shows a significant surge peak, with a large number of consecutive data points exceeding the first rate of change threshold of 15 mm / s.

[0046] At this point, the red curve representing the AI ​​deformation calculation thread frequency, after detecting that the rate of change exceeded the limit in the background, adaptively stepped up from the original 1 Hz normal operating frequency to around 2 Hz, thereby providing higher density data refresh and matrix calculation support in the resection stage where tissue deformation is most severe.

[0047] Based on the simulation waveforms in the actual R&D environment, it can be seen that the red curve has a slight physical scheduling delay during the switching phase of frequency increase and decrease, and the blue high-frequency curve shows slight glitches and jitters in the concurrent operation range of high system load.

[0048] This indicates that although the master-slave asynchronous dual-track operating architecture has effectively isolated rendering and computing resources, there is still room for further optimization in terms of kernel interrupt response mechanism and thread wake-up sensitivity when multimodal data is concurrently flooded in.

[0049] It is important to emphasize that the second preset frequency is significantly higher than the normal operating frequency, thereby providing higher density data refresh support during the most intense stage of tissue deformation and ensuring the real-time approximation capability of deformation field calculation.

[0050] Optionally, to maximize solution accuracy with limited hardware computing power, the aforementioned spatial elastic deformation compensation field matrix is ​​not constructed using a uniform spatial grid, but is designed as a non-uniform spatial data structure containing four grid resolution levels. These four levels are distributed in a gradient according to the importance of the anatomical structure and its distance from the resection center.

[0051] Specifically, the four grid resolution levels consist of a critical structure-specific layer, a core layer, a transition layer, and an outer layer. The critical structure-specific layer covers the key anatomical structures described in the preoperative planning model. For these highly displacement-sensitive nerves and blood vessels, this layer is configured with the densest grid resolution in the entire matrix; that is, the grid resolution of the critical structure-specific layer is significantly higher than that of the core layer. The core layer is a spatial region radiating outwards from the geometric centroid of the current virtual tumor boundary to a first radius; this region is where surgical manipulation is most concentrated. The spatial region extending outwards from the outer edge of the core layer to a second radius is defined as the transition layer, whose grid resolution becomes increasingly sparse with distance. The outer layer covers all remaining surgical field spatial regions not encompassed by the above three layers and employs the sparsest grid resolution. This four-layer design ensures the accuracy of critical site calculations while significantly reducing the overall matrix's floating-point computation.

[0052] like Figure 3 This visually illustrates the non-uniform spatial data structure design that maximizes solution accuracy under limited hardware computing power. In the figure, the horizontal extension direction corresponds to the horizontal axis spatial coordinate, the depth extension direction corresponds to the vertical axis spatial coordinate, and the height extension direction corresponds to the vertical axis spatial coordinate; all units are in millimeters.

[0053] The distribution of three-dimensional scatter points of different colors in the figure clearly shows that the matrix is ​​distributed in four levels according to the importance of anatomical structures and the distance from the resection center.

[0054] The dense cluster of dark red dots in the core area of ​​the diagram represents the critical structure-specific layer, which covers areas such as nerves and blood vessels that are extremely sensitive to displacement. The point cloud distribution in this area is the densest, corresponding to an extremely high grid resolution of 0.5 mm.

[0055] The orange-yellow spherical dot cloud clusters surrounding it represent the core layer radiating outward from the geometric centroid of the virtual tumor boundary to a radius of 20 millimeters, corresponding to a grid resolution of 1 millimeter. This area is where surgical operations are most concentrated.

[0056] The outward-expanding green scattering represents a transition layer with a radius of 20 mm to 50 mm, corresponding to a grid resolution of 3 mm. Its point cloud density shows a significant sparsification evolution as the spatial distance increases.

[0057] The outermost blue sparse dots represent the outer layer covering the remaining surgical area, using the lowest grid resolution of 10 mm. This spatial gradient distribution of color from warm to cool and dot density from dense to sparse verifies the effectiveness of the system's technology in allocating computing resources to high-risk clinical critical areas, effectively reducing the overall floating-point computation of the matrix.

[0058] Furthermore, by combining the boundary details of the simulated images with real data, it can be seen that in the boundary region between the core layer and the transition layer, due to the 3D voxelization stitching of grids with different resolutions and local spatial deformation disturbances, green and orange-yellow scattered points exhibit a certain degree of discrete, jagged, and overlapping phenomenon on the spatial transition interface. This objectively indicates that the current system still has certain engineering optimization space in terms of smooth interpolation across grid nodes and the continuity of spatial mechanics transition. In the future, the rigid truncation error at the boundary can be further eliminated by introducing a higher-order smooth transition function.

[0059] Specifically, in this embodiment, the AI ​​deformation calculation thread not only outputs the deformation displacement, but also calculates the deformation prediction confidence of the output deformation result through a multimodal cross-validation mechanism, and generates a confidence field matrix that completely corresponds to the spatial elastic deformation compensation field matrix in spatial topology. The calculation of deformation prediction confidence does not rely on a single data source, but is obtained by weighted fusion of scores normalized from three independent dimensions.

[0060] First, there's the model uncertainty dimension, which primarily assesses the predictive stability of the AI ​​neural network itself. Its mechanism is based on multiple Monte Carlo inferences. A random deactivation layer is introduced into the neural network, and the same input data is passed forward multiple times. The standard deviation of the multiple predictions is calculated, and this standard deviation is then mapped back to the first dimension score. The smaller the standard deviation, the more stable the model's prediction for that region, and the higher the first dimension score.

[0061] Secondly, there is the multimodal consistency dimension, which is responsible for verifying the consistency between the AI ​​prediction results and the actual physical measurement data. The system uses a mutual information algorithm to calculate the degree of matching between the real-time ultrasound image features of the brain tissue during surgery and the coordinates of discrete tissue feature points, and the preoperative images after deformation driven by the AI-predicted deformation field. This matching degree is then linearly mapped to the second dimension score. The higher the matching degree, the closer the predicted deformation field is to the actual physical feedback.

[0062] Finally, there is the clinical prior constraint dimension, which verifies the physiological rationality of the deformation results based on a prior knowledge base of cranial anatomy. Any predicted displacement that violates basic biomechanical principles of brain tissue will be penalized, and the calculated rationality index is mapped to the third dimension score.

[0063] It should be noted that, in order to rigorously quantify the above fusion process, the system uses the following confidence calculation formula to obtain the final deformation prediction confidence:

[0064] ;

[0065] in, This indicates the reliability of the deformation prediction ultimately generated by the spatial vertices; Represents the fixed weight coefficients for the pre-configured model uncertainty dimension; This represents the score for the first dimension calculated by the system; Represents the fixed weight coefficients of the pre-configured multimodal consistency dimension; This represents the score for the second dimension calculated by the system; Represents the fixed weight coefficients of the pre-configured clinical prior constraint dimensions; This represents the score for the third dimension calculated by the system.

[0066] Through this weighted summation, each node in the confidence field matrix is ​​assigned an objective quantitative evaluation index.

[0067] For example, based on the generated confidence field matrix, the AI ​​deformation calculation thread can achieve proactive error correction and secondary optimization. The system will detect the spatial distribution state in real time based on the confidence field matrix. When specific triggering conditions are met simultaneously, the system will detect the target update region and automatically perform local deformation field recalculation. These automatic triggering conditions are extremely stringent, requiring the simultaneous fulfillment of three logical judgments:

[0068] First, the system detects regions where the deformation prediction confidence level is below the second confidence threshold, and the volume of these continuously clustered regions in three-dimensional space is greater than a first preset volume threshold. This eliminates false triggers caused by isolated noise points. Second, this continuous region must be located on a pre-planned surgical path, or the three-dimensional Euclidean distance between this region and key anatomical structures must be less than a first preset safety distance threshold. This ensures that computational power is only allocated to high-risk clinically critical areas. At this point, the system determines that the original deformation prediction results for this region no longer have navigation reference value and designates it as the target update region.

[0069] Specifically, after determining the target update region, the AI ​​deformation calculation thread, when performing local deformation field recalculation, abandons the conventional end-to-end neural network inference and instead adopts a more rigorous mathematical optimization method based on physical mechanics. The system uses a weighted constraint energy function as the optimization objective, which consists of two mathematical terms: a data term and a smoothing term. The data term is used to narrow the distance between the predicted displacement and the actual measured displacement, i.e., calculating the weighted sum of squares of the difference between the actual measured displacement and the predicted displacement of discrete tissue feature points within the target update region. The smoothing term is used to maintain the continuous mechanical medium properties of the brain tissue, i.e., calculating the sum of squares of the difference between the displacements to be solved at adjacent spatial vertices.

[0070] It should also be noted that the specific mathematical model of this weighted constrained energy function is as follows:

[0071] ;

[0072] in, This represents the total energy value of the constructed weighted constrained energy function; Represents the global balance coefficient configured for the data item; This represents the total number of valid discrete tissue feature points extracted within the target update region; Indicates that for the first Local trust weights assigned to each feature point; This indicates the number of predictions made by the current model. Spatial displacement vector of each feature point; This represents the first value obtained through actual measurement using physical instruments. Spatial displacement vector of each feature point; Represented as the global elastic constraint coefficient configured for the smoothing term; This represents the set of all spatial vertices within the target update region; Represents the set of elements. The set of all first-order adjacent vertices of a given spatial vertex; Indicates the first The displacement vectors of the spatial vertices to be solved; Indicates its adjacent first The displacement vectors of the vertices in space to be solved.

[0073] The system uses the Gauss-Newton optimization algorithm in the nonlinear least squares solver to iteratively calculate the Jacobian matrix and continuously minimize the weighted constraint energy function. This continues until the energy converges, thus completing the recalculation of the local deformation field in the target update region.

[0074] like Figure 4 This visually illustrates the underlying mathematical optimization process and multi-objective game relationship during local recalculation of the target update region. In the left subgraph, the horizontal axis represents the number of iterations, and the vertical axis represents the total energy value, which is dimensionless.

[0075] The blue optimization convergence curve in the left subplot reflects the execution trajectory of the Gauss-Newton optimization algorithm. In the first 20 iterations, the total energy value drops rapidly from an initial high of nearly 1000 to around 50. This indicates that the algorithm can quickly narrow down the predicted displacement and the actual measured displacement of discrete tissue feature points within the target area.

[0076] However, after 30 iterations, the blue curve did not remain a horizontal straight line, but instead showed slight up-and-down oscillations. This objectively reflects the numerical instability caused by the local complex topology of brain tissue under stress deformation, indicating that the solver still has room for further smoothing optimization in eliminating high-frequency residual stress.

[0077] In the right subplot of the figure, the horizontal axis represents the energy value of the smoothing term, in dimensionless units, and the vertical axis represents the energy value of the data term, in dimensionless units. The red Pareto front curve in the right subplot illustrates the constraint between the two conflicting optimization objectives of maintaining tissue physical continuity and fitting true displacement data.

[0078] The gray scatter dots in the figure represent sampling combinations with different weighting coefficient configurations, while the green dots indicate the optimal balance point currently selected by the system. This green selection point is set in the inflection point region of the red curve, achieving a reasonable trade-off between data fitting accuracy and grid smoothness.

[0079] However, it can be observed that the green selection point does not completely fit the bottom edge of the red theoretical extreme value curve. This means that the global static balance coefficient used in the current system still has compromises in local overfitting or underfitting when dealing with extreme nonlinear collapse. In the future, this technical defect can be made up for by introducing a dynamic adaptive weight allocation mechanism.

[0080] Optionally, after the underlying data is recalculated, the system uses a coordinate mapping module to reflect the mesh displacement of the underlying layer onto the surface navigation screen. The coordinate mapping module performs local interpolation remapping in three-dimensional space based on the spatial pose data of the surgical instruments and the updated spatial elastic deformation compensation field matrix. The purpose of this process is to dynamically update the initial virtual tumor boundary, ensuring it deforms along with the real tissue, ultimately generating the current virtual tumor boundary. To prevent tearing or abrupt changes in the surface rendering due to jumps in the underlying data, the coordinate mapping module uses a confidence-guided progressive smooth interpolation function for boundary updates.

[0081] Specifically, the final updated displacement of the spatial vertices is not a direct replacement, but rather a weighted calculation of the displacement before the update and the newly calculated displacement based on interpolation coefficients. The calculation formula corresponding to this interpolation logic is as follows:

[0082] ;

[0083] in, This represents the final updated displacement vector output after smoothing calculation of the vertex in the space; Represents the interpolation coefficients dynamically allocated by the system; This represents the pre-update displacement vector stored at this spatial vertex before the current update cycle. This represents the newly calculated displacement vector just output by the AI ​​deformation calculation thread.

[0084] It is important to note that the interpolation coefficients The value of is not a fixed constant, but rather has a strict positive correlation with the deformation prediction reliability of that spatial vertex. The higher the deformation prediction reliability, The closer the value is to one, the more the system trusts the newly calculated data; however, when the system detects that the deformation prediction confidence is lower than the aforementioned second confidence threshold, in order to prevent deformation data that may contain serious errors from polluting the navigation screen, the system forcibly... Set it to zero. At this point, in the formula... When the weight of a vertex is zero, the system will pause the displacement update of that vertex and maintain its original shape until the local deformation field is recalculated and higher-confidence data is output before the update resumes.

[0085] For example, in order to transform complex confidence data into visual language that doctors can intuitively understand, the intelligent AR display module in this embodiment plays a crucial visualization role. The intelligent AR display module is used to perform hierarchical visualization rendering of the current virtual tumor boundary based on the confidence field matrix. Specifically, this module integrates a multi-level visualization encoding scheme into the graphics rendering pipeline. This scheme queries the deformation prediction confidence of each spatial vertex constituting the boundary in real time, and dynamically modifies the visual attributes of the boundary at the shader level according to the numerical range of the confidence level, specifically matching the corresponding boundary line type, line color, transparency, and line width.

[0086] Furthermore, the multi-level visualization encoding scheme ensures the objectivity and cautionary nature of the navigation information presentation. For example, the boundary of a high-confidence zone may be rendered as a continuous thin solid line, indicating extremely accurate navigation; however, when the system detects that the deformation prediction confidence of the currently rendered boundary has fallen below the second confidence threshold, the system determines that the navigation data in that area is unreliable. At this time, in addition to changing the color and line type, the intelligent AR display module will also activate the rendering effect in the time dimension, configuring the boundary of the zone with a dynamic display feature of a first preset flashing frequency, directly conveying the warning message "for reference only, proceed with caution" to the surgeon through visual flashing.

[0087] It's important to note that such passive visual warnings can be overlooked in complex surgical environments. Therefore, the intelligent AR display module is also equipped with a tiered dynamic risk warning mechanism that is deeply tied to spatial distance. During surgery, the system frequently calculates the spatial distance between the virtual coordinates of the surgical instruments and various surrounding boundaries. When the system detects that the shortest Euclidean distance between the surgical instruments and a boundary where the deformation prediction confidence level is below a first confidence threshold is less than a second preset safety distance threshold, it triggers an audio prompt at a first preset audio playback frequency. This audio prompt is typically a slow, ticking sound, intended to remind the surgeon that they are approaching an area with relatively weak navigation data.

[0088] As the device advances further, and the spatial distance shrinks further, even falling below the third preset safety distance threshold, the system determines that it has entered a critical state of extremely high risk. At this point, the system immediately escalates the audible alarm, triggering an audible prompt at a second preset audio playback frequency higher than the first preset audio playback frequency, for example, as a rapid, continuous beeping sound. Simultaneously, on the visual display side, the system increases the flashing frequency of the corresponding boundary to a second preset flashing frequency higher than the first preset flashing frequency, forming a multi-sensory high-energy alarm with sound and light linkage. In this logic, the system strictly limits the third preset safety distance threshold to be less than the second preset safety distance threshold, thus constituting a multi-level dynamic spatial risk warning range.

[0089] Optionally, the underlying basis for all the clinical prior constraints involved in the aforementioned confidence assessment systems comes from a pre-constructed prior knowledge base of cranial anatomy. This knowledge base is not a simple text database, but a structured parameter matrix storing quantified physiological deformation constraint rules. These rules transform biomechanical experience in neurosurgery into computer-executable mathematical boundaries. The physiological deformation constraint rules specifically include hard constraints in four core dimensions:

[0090] First, for structures such as the inner table of the skull and the falx cerebri, a maximum permissible deformation displacement threshold is stored; any predicted displacement exceeding this threshold will be considered a solution failure. Second, an upper limit threshold for the relative positional changes of major intracranial blood vessels is stored, because vascular bundles are typically tightly wrapped by connective tissues such as the arachnoid membrane, and their relative distance to other adjacent structures is unlikely to change drastically in a short period. Third, an upper limit threshold for the changes in the directional angle of nerve fiber bundles used to maintain the integrity of nerve function is stored.

[0091] Specifically, to verify whether the angular changes of nerve fiber bundles conform to the physiological deformation constraint rules, the system calculates the angular change value by comparing the spatial tangent vectors of the nerve fiber before and after deformation. The specific geometric angle verification formula is as follows:

[0092] ;

[0093] in, This indicates the calculated angle of deflection of the nerve fiber bundle in the spatial orientation before and after deformation; This represents the three-dimensional spatial orientation tangent vector of a point on the nerve fiber bundle in the preoperative planning model. This represents the three-dimensional tangent vector of the corresponding point in the current virtual computational space after deformation derivation using the spatial elastic deformation compensation field matrix. The inverse cosine value is obtained by calculating the ratio of the dot product to the modulus product. The system requires the calculated inverse cosine to be... The value must not exceed the upper limit threshold of the angle change of the nerve fiber bundle preset in the knowledge base; otherwise, the score of the third dimension of that area will be significantly reduced.

[0094] Finally, the physiological deformation constraint rules also include constraints to maintain the order of interconnections in the basic topological relationships between anatomical structures. This is a graph-based verification mechanism that ensures that deformation calculations do not lead to incorrect physical penetration between two originally non-intersecting gyri in virtual space, or cause the breakage of originally connected vascular branches.

[0095] In general, existing conventional navigation systems often adopt a single-threaded synchronous processing architecture, which is prone to causing the entire interface to lag when faced with massive three-dimensional deformation matrix calculations. This can lead to serious medical accidents caused by visual feedback delays when doctors are performing delicate dissection operations. At the same time, existing systems lack a self-doubt and verification mechanism for the prediction results output by AI, often displaying incorrect deformation boundaries as definite solid lines, causing doctors to blindly trust and remove tissues that should not be removed.

[0096] This embodiment, through its highly innovative master-slave asynchronous dual-track operating architecture, effectively decouples the resource conflicts between image tracking and underlying recalculation, ensuring low-latency cursor following. More importantly, by constructing a multi-dimensional confidence assessment system and hierarchical visual coding, the system achieves objective confidence assessment of its own prediction results. It no longer provides doctors with a single, fixed static boundary, but dynamically displays the data reliability of different regions, and automatically initiates high-precision local recalculation for self-correction when low-confidence areas are detected. This comprehensive technical solution, combining multi-level physical space computing power allocation, mathematical optimization, and multi-sensory early warning, greatly bridges the information gap between digital virtual space and dynamic physical space, providing solid engineering and technical support for the precise and safe resection of brain tumors.

[0097] Example 2:

[0098] Based on Example 1, this embodiment provides an in-depth technical solution to address the core pain point in the process of intracranial tumor resection—namely, the complex cavity collapse problem caused by the substantial removal of tumor material.

[0099] Specifically, the AI ​​deformation calculation thread also includes a cavity dynamics contraction compensation module; during the surgical instrument's resection operation, the cavity dynamics contraction compensation module dynamically captures and records the spatial position sequence of the continuous operation of the surgical instrument tip to reconstruct the three-dimensional spatial geometric contour of the resection area, and defines the three-dimensional spatial geometric contour as the absolute cavity geometric boundary; the cavity dynamics contraction compensation module has a built-in fluid-elasticity coupled dynamics model, which is used to simulate the physical deformation trend of the residual tumor boundary and surrounding brain tissue collapsing towards the inside of the cavity according to the local stress release state shaped by the absolute cavity geometric boundary, and calculate and generate an active contraction displacement field for the current virtual tumor boundary;

[0100] The AI ​​deformation calculation thread uses the absolute cavity geometric boundary as the spatial rigid barrier boundary for deformation field calculation, and injects the active contraction displacement field as a dynamic load condition into the spatial elastic deformation compensation field matrix to perform dynamic iterative calculation of boundary conditions, so as to correct the local contraction and collapse displacement of the current virtual tumor boundary during the resection process in real time.

[0101] For example, in actual microsurgical neurosurgery, surgeons typically use ultrasonic surgical aspirators or bipolar electrocoagulation forceps to fragment and remove tumor tissue. Traditional navigation systems can only perform intermittent passive scanning via intraoperative ultrasound, failing to capture the centripetal collapse of brain tissue caused by material loss and cerebrospinal fluid drainage at the moment of resection. The cavity dynamics contraction compensation module in this embodiment is designed to fill this monitoring blind spot during dynamic resection. This module does not introduce additional invasive detection equipment; instead, it cleverly uses the surgical instruments performing the resection as dynamic probes, thereby achieving high-frequency active sensing of the resected cavity.

[0102] It is important to note that, in order to achieve precise capture of the resection action, the positioning and acquisition module acquires the three-dimensional spatial coordinates of the surgical instrument tip with optical markers in real time at an extremely high sampling frequency. The cavity dynamics contraction compensation module continuously monitors these coordinate data in the background and dynamically captures and records the spatial position sequence of the surgical instrument tip during continuous operation.

[0103] In practical applications, because the movement trajectory of surgical instruments includes a large number of invalid hovering, exit clearing, and high-frequency jitter, the system does not directly record all trajectory points. Instead, this module integrates a downsampling and denoising algorithm based on a spatial voxel hash map. Only when the surgical instrument is in an effective resection state and the tip coordinates undergo substantial displacement in a new unrecorded voxel space will the coordinates be recorded in the spatial position sequence, thereby ensuring that the recorded data can truly reflect the physical elimination boundary of tumor material.

[0104] Optionally, after acquiring the purified spatial position sequence, the cavity dynamics contraction compensation module needs to perform a computationally intensive geometric modeling task to reconstruct the three-dimensional spatial geometric contour of the resection area. This reconstruction process relies on the 3D Delaunay Triangulation and Poisson Surface Reconstruction algorithms. The system first converts the continuous spatial position sequence into a dense directed point cloud, and then uses the Poisson Surface Reconstruction algorithm to extract a watertight three-dimensional surface. This three-dimensional surface, composed of countless triangular facets, precisely encloses all spatial regions traversed by the surgical instrument trajectory. The module then defines the three-dimensional spatial geometric contour as the absolute cavity geometric boundary. In the system's virtual computational space, this boundary has extremely high physical significance: the interior of the boundary is marked as the "vacuum / fluid phase," meaning that the tumor entity in this area has been completely removed and no longer provides any mechanical support; the exterior of the boundary is marked as the "porous elastic phase," representing residual tumor tissue or normal brain parenchyma that still possesses mechanical strength.

[0105] like Figure 5 This demonstrates the engineering implementation process of the cavity dynamics contraction compensation module, which transforms discrete surgical manipulation trajectories into effective physical and mechanical boundaries. The units for the horizontal, vertical, and axial spatial coordinates in the figure are all millimeters.

[0106] The widely distributed gray scattering in the figure represents the spatial position sequence of continuous operation of the surgical instrument tip captured and recorded by the system at high frequency. These discrete point clouds delineate an irregular concave region with a depth of about 25 mm in the three-dimensional coordinate system, objectively reflecting the real spatial range of the physical removal of tumor material.

[0107] The orange-red semi-transparent surface covering the gray point cloud represents the three-dimensional spatial geometric contour extracted and generated by the system from the discrete trajectory point cloud using the surface reconstruction algorithm. This watertight surface is strictly defined in the virtual space as an impenetrable absolute cavity geometric boundary, providing a basic geometric constraint barrier for mechanical calculations of subsequent simulations of brain tissue collapsing into the cavity.

[0108] By combining the surface texture details of the simulation images, it is clearly observed that the top opening edge of the orange-red curved surface exhibits irregular jagged protrusions, and local mesh depressions and surface roughness appear in the relatively sparse point cloud distribution area at the bottom of the surface. This realistically reflects the problem of uneven spatial sampling density caused by differences in operation dwell time when relying solely on the irregular motion trajectory of surgical instruments as the spatial sampling data source. This indicates that the current geometric reconstruction algorithm has limitations in handling the surface smoothing tension of extremely sparse trajectory point clouds, and there is an urgent need to introduce curvature-aware adaptive point cloud upsampling interpolation technology in the future to further improve the physical smoothness of geometric boundaries.

[0109] Specifically, in order to accurately predict the deformation behavior of the residual tissue after the loss of mechanical support, the cavity dynamics contraction compensation module incorporates a fluid-elastic coupling dynamics model. Since brain tissue is widely regarded in biomechanics as a saturated porous elastic material filled with cerebrospinal fluid, a simple pure elastic model cannot accurately describe its delayed response and large deformation characteristics after being subjected to force.

[0110] In this embodiment, the model treats brain tissue as a two-phase coupled continuous medium consisting of a solid glial cytoskeleton and a liquid interstitial fluid. When the tumor is removed, forming a cavity, the local high-pressure stress field maintained by the tumor volume is instantly lost. Cerebrospinal fluid seeps into the lower-pressure cavity, causing the solid cytoskeleton to contract. This model is used to simulate the physical deformation trend of the residual tumor boundary and surrounding brain tissue collapsing inward into the cavity, based on the local stress release state shaped by the absolute cavity's geometric boundary.

[0111] For example, to quantify this complex physical deformation trend within a computer, a fluid-elasticity coupled dynamics model relies on a rigorous set of partial differential equations to calculate the active contraction displacement field for the current virtual tumor boundary. Assuming that the tiny brain tissue units are in quasi-static equilibrium, the system uses an extended form based on the Navier-Cauchy equations to solve for the displacement distribution. The core partial differential equations solved by the system at each time step are expressed as follows:

[0112]

[0113] in, This represents a vector differential operator used to perform gradient or divergence operations on spatial variables; This indicates the macroscopic shear modulus of brain tissue inferred by the system based on preoperative image density; This represents the active contraction displacement vector field that the system needs to solve, which describes the direction and distance of each point in space to collapse into the cavity; The macroscopic volumetric modulus of brain tissue, representing the brain tissue's ability to withstand pressure. This represents the volumetric stress release driving force vector triggered by the resection action. The value of this vector is non-linearly positively correlated with the surface area and depth of the absolute cavity geometric boundary.

[0114] It is also important to note that the practical application of the above formula lies in transforming extremely complex biomechanical problems into large-scale sparse linear equation systems that can be solved rapidly by computers through finite element analysis (FEM) or the meshless Galerkin method. By solving these equations, the system can clearly understand the dynamics of the cavity under the current cavity morphology. Driven by this, how far will each remaining millimeter of brain tissue collapse inward? Solve for the output... After discretization, an active contraction displacement field is formed, providing a direct mathematical basis for subsequent boundary correction.

[0115] like Figure 6 This diagram illustrates the local stress release state shaped by the system according to the cavity's geometric boundaries, as well as the physical deformation trend of the residual brain tissue. The horizontal and vertical spatial coordinates in the figure are both in millimeters, collectively defining a surgical calculation core region with a side length of 60 millimeters. The vertical spatial coordinate, also in millimeters, represents the amplitude of active contraction and collapse displacement of the residual tissue.

[0116] The three-dimensional nonlinear response surface in the figure, which gradually transitions from cool to warm tones, reflects the nonlinear spatial distribution characteristics of stress release caused by the loss of resected material. In the region near the resection center, the surface exhibits a warm tone and the displacement amplitude reaches its maximum value of approximately 15 mm. In the peripheral region far from the center, the displacement amplitude gradually converges exponentially to a cool tone with a smooth state approaching zero. This objectively reflects the nonlinear stress decay characteristics of the porous elastic medium of brain tissue.

[0117] The red three-dimensional vector arrows superimposed on the curved surface represent the generated active contraction displacement field, which densely points from the outer region to the collapse center, clearly indicating the specific direction and driving intensity of the residual brain tissue collapsing into the cavity.

[0118] Based on simulation data from a real R&D environment, it can be observed that in the region of abrupt slope change near the peak displacement center of the response surface, the surface mesh exhibits a certain degree of local irregular oscillation and surface burr protrusions. Furthermore, the corresponding red vector arrows show some directional deviations and scattered, intersecting phenomena. This indicates that in the dynamic iterative solution process handling collisions and truncation at rigid spatial boundaries due to extreme nonlinear deformation, the current system's local organization flexibility mapping and relaxation iteration step size still introduce certain numerical truncation errors when balancing abrupt changes in high-gradient stress. Future research can further eliminate numerical oscillation defects in high-stress gradient regions by introducing curvature-sensing mesh re-division techniques with finer-grained adaptive adjustment capabilities, thereby improving the physical continuity and smoothness of complex rheological calculations.

[0119] Optionally, after obtaining the theoretically calculated active contraction displacement field, the common model error penetration problem in numerical simulations must be considered. Simply superimposing the displacement field may result in virtual mesh nodes of residual tissue crossing the cavity boundary, or even serious topological errors such as self-intersection. To prevent this physically unrealistic phenomenon, the AI ​​deformation calculation thread uses the absolute cavity geometric boundary as a rigid barrier constraint for the deformation field calculation. At the algorithm level, the system introduces a collision detection mechanism based on the signed distance field (SDF). The system calculates the shortest distance from all spatial vertices to the absolute cavity geometric boundary in real time. When any spatial vertex attempts to enter the interior of the absolute cavity geometric boundary under the drive of the active contraction displacement field, the collision detection mechanism is triggered, and the normal component of the vertex's displacement is forcibly truncated, causing it to always slide along the cavity surface or remain stationary on the boundary, thus ensuring the topological consistency and physical rigor of the virtual anatomical structure.

[0120] Specifically, with the application of a rigid spatial boundary, the system must unify the local collapse deformation caused by resection with the global deformation caused by overall gravity or cerebrospinal fluid loss from the surgical area. To this end, the AI ​​deformation calculation thread simultaneously injects the active contraction displacement field as a dynamic load condition into the spatial elastic deformation compensation field matrix for dynamic iterative calculation of the boundary conditions. In Example 1, the spatial elastic deformation compensation field matrix primarily handles deformation observed based on ultrasound and probes; while in Example 2, this matrix is ​​endowed with the ability to handle deformation driven by active physics. This injection is not a simple data overlay, but rather uses a high-order numerical iterative method to find the lowest energy balance point between local resection stress and global deformation trend.

[0121] For example, to ensure the numerical stability of the dynamic iterative solution process and avoid the collapse of the entire matrix due to drastic changes in local stress, the system employs a relaxation iterative update algorithm that incorporates local tissue stiffness mapping. Within each solution time step, the update of the spatial elastic deformation compensation field matrix follows the iterative formula:

[0122] ;

[0123] in, This represents the spatial elastic deformation compensation field matrix for the next iteration step, obtained after this round of calculations. This represents the spatial elastic deformation compensation field matrix used at the beginning of the current calculation cycle for the current iteration step; This represents the numerical relaxation step size used to control the magnitude of matrix updates, which is dynamically and adaptively adjusted with the cutoff rate; This represents the Hadamard product operator, which performs element-wise multiplication on corresponding elements of a matrix. This represents the local organizational flexibility tensor that reflects the ability of the current spatial location to undergo deformation, i.e., the normalized inverse matrix of the stiffness tensor; This represents the absolute cavity geometric boundary constraint operator used to perform anti-penetration collision detection and displacement truncation. It outputs a truncation matrix when it touches the boundary, and an identity matrix otherwise.

[0124] It is also important to note that the core function of this formula in practical applications is to achieve smooth fusion of deformation data. Through and Under the combined regulation of these factors, in regions close to the resected cavity and where the tissue is relatively soft (with a small stiffness tensor value), the matrix rapidly absorbs the active contraction displacement field. The effect of the matrix is ​​significant, resulting in collapse compensation. However, when the distance from the resected cavity is far or when encountering rigid structures such as blood vessels and nerves (whose local tissue flexibility tensor values ​​are close to zero), the update amplitude of the matrix will be greatly suppressed, thus accurately simulating the nonlinear attenuation and transmission process of stress in non-uniform brain tissue.

[0125] Optionally, through the aforementioned in-depth physical and mathematical solutions, the system ultimately achieves the technical objective of real-time correction of the local contraction, collapse, and displacement of the current virtual tumor boundary during the resection process. In the front-end rendering of the intelligent AR display module, as the surgeon continuously removes the tumor with instruments, the doctor can intuitively observe that the current virtual tumor boundary, originally outlined in the field of view, not only does not appear to lag due to tissue removal, but also dynamically and precisely follows the rhythm of the instruments' resection, contracting inward and adjusting its shape. When the instruments rapidly empty the tumor, forming a cavity, the virtual boundary exhibits an elastic collapse trend towards the inside of the cavity. This real-time, interactive visual feedback greatly eliminates the "spatial cognitive dissonance" often experienced by doctors under traditional static navigation.

[0126] In summary, considering the limitations of existing technologies, when dealing with intraoperative brain displacement, existing technologies rely heavily on intermittent external imaging equipment (such as intraoperative MRI or intraoperative ultrasound) to passively refresh data. This not only severely disrupts the continuous rhythm of the surgery, but also leaves the navigation data in a state of complete blind trust failure during the window between two imaging scans.

[0127] The technical solution disclosed in this embodiment fully reuses the conventional operating trajectory of surgical instruments, transforming it into first-hand physical data for constructing the absolute cavity boundary. By further introducing a fluid-elasticity coupled dynamics model and performing rigorous partial differential equation solving and numerical iteration, the system can actively predict and compensate for cavity collapse caused by material removal. This overall solution enables the navigation system to possess advanced intelligence of "perceiving the resection action" and "reasoning about the physical consequences," effectively solving the most dangerous problem of severe local nonlinear brain displacement during deep tumor resection, significantly improving the real-time realism of surgical navigation, and greatly reducing the clinical risk of cutting critical blood vessels or damaging important nerve fibers due to navigation boundary lag.

[0128] Example 3:

[0129] Based on Examples 1 and 2, this embodiment further provides an active physical compression compensation scheme based on tissue rheological properties for microsurgical procedures involving deep brain tumors, such as the forced instrument traction operations unique to transsulcal and transcortical approaches.

[0130] Specifically, the positioning and acquisition module is also used for real-time spatial positioning of the brain retractor with optical markers. The AI ​​deformation calculation thread also includes a device contact mechanics and viscoelastic creep assessment module. During the operation of the brain retractor intervening in normal brain tissue to establish a deep surgical channel, the device contact mechanics and viscoelastic creep assessment module constructs a dynamic, tissue-impermeable three-dimensional spatial repulsion entity boundary in virtual space based on the real-time spatial pose data of the brain retractor and the preset three-dimensional solid contour model of the brain retractor. The device contact mechanics and viscoelastic creep assessment module is configured with brain tissue viscoelastic mechanics parameters based on time decay factors, which are used to perform two-stage deformation compensation calculations for the spatial vertices around the three-dimensional spatial repulsion entity boundary. The two-stage deformation compensation calculations include a transient elastic displacement calculation stage and a steady-state compressive creep calculation stage.

[0131] In the transient elastic displacement calculation stage, when spatial displacement of the brain retractor is detected and the boundary of the three-dimensional spatial repulsion entity invades the virtual space occupied by the original brain tissue, the instantaneous mechanical compression displacement required for the spatial vertex in the invading area to move out of the repulsion area along the normal direction of the boundary of the three-dimensional spatial repulsion entity is calculated; in the steady-state compression creep calculation stage, when the displacement change rate of the brain retractor is detected to approach zero and it is in a static traction holding state, the secondary collapse extension displacement of the spatial vertex under constant pressure is calculated based on the viscoelastic mechanical parameters of the brain tissue and the continuous compression time variable; the AI ​​deformation calculation thread superimposes the instantaneous mechanical compression displacement and the secondary collapse extension displacement in real time to the corresponding position of the spatial elastic deformation compensation field matrix to dynamically correct the spatial coordinates of the current virtual tumor boundary under forced traction.

[0132] For example, when removing deep-seated brain tumors, such as those in the basal ganglia or ventricles, the surgeon cannot directly observe the lesion through the cortical surface. In such cases, a brain retractor (usually composed of a malleable metal plate and a fixed serpentine arm) must be used to continuously push aside the normal brain parenchyma (such as gyri) blocking the surgical path, thereby artificially exposing a surgical channel that leads directly to the deep part of the tumor.

[0133] Traditional navigation systems based on preoperative images present a static virtual image of the brain tissue. When a real brain retractor is inserted into the brain tissue, a serious model error occurs in the navigation image – the virtual brain tissue is covered by the real instrument, causing doctors to lose spatial judgment of the actual location of deep blood vessels and nerves in the compressed area.

[0134] To address this issue, the positioning and acquisition module in this embodiment is also used for real-time spatial positioning of the brain retractor with optical markers. Before actual operation, the doctor rigidly fixes a standard infrared reflective ball array to the handle base of the brain retractor. The system, through a preoperative pivot calibration algorithm, can accurately calculate the absolute spatial coordinates of the retractor's pressure plate tip and the entire metal plate surface.

[0135] It is important to note that simply obtaining the spatial coordinates of the brain retractor is insufficient for mechanical compensation. The AI ​​deformation calculation thread also includes a module for evaluating instrument contact mechanics and viscoelastic creep. This module primarily addresses the physical occupancy issue in virtual space. During the process of the brain retractor intervening in normal brain tissue to establish a deep surgical channel, the module constructs a dynamic, tissue-impenetrable 3D spatial repulsion boundary in virtual space based on the real-time spatial pose data of the brain retractor and a pre-set 3D solid contour model of the brain retractor. This pre-set 3D solid contour model is a digital geometric copy of the instrument pre-entered into the system using a high-precision 3D scanner. As the retractor moves in the physical world, this 3D spatial repulsion boundary moves synchronously in virtual space. This boundary is assigned an infinitely large normal stiffness property, meaning that in the physics calculation engine, no virtual space vertex representing soft tissue is allowed to exist inside this boundary.

[0136] Optionally, brain tissue is not a simple linear elastomer, but a complex biorheological medium rich in intracellular fluid and extracellular interstitial fluid, exhibiting significant viscoelastic characteristics. This means that the deformation response of brain tissue to physical compression is not instantaneous, but highly correlated with the time of force application. Therefore, the device contact mechanics and viscoelastic creep assessment module is equipped with brain tissue viscoelastic mechanics parameters based on a time decay factor. These parameters can typically be individually extracted from preoperative magnetic resonance elastography (MR elastography) data, or assigned values ​​based on a standard biomechanical database according to the patient's age. Based on these parameters, the module performs a two-stage deformation compensation calculation for the spatial vertices surrounding the boundary of the three-dimensional spatial repulsion entity. The surrounding area is typically defined as the deformation influence zone within a certain distance (e.g., three centimeters) from the surface of the repulsion boundary. The two-stage deformation compensation calculation includes a transient elastic displacement calculation stage and a steady-state compression creep calculation stage, which correspond to two distinct clinical operating states of the retractor: "dynamic push-pull" and "static hold," respectively.

[0137] Specifically, the system first enters the transient elastic displacement calculation stage. In this stage, when the system's high-frequency positioning module detects that the brain retractor has undergone spatial displacement and the three-dimensional spatial repulsion entity boundary has invaded the virtual space originally occupied by the brain tissue, this simulates the dynamic process of a doctor forcefully prying open the brain tissue. At this time, the system uses the Bounding Volume Hierarchy (BVH) algorithm to perform high-frequency collision detection, filtering out all spatial vertices "covered" by the three-dimensional spatial repulsion entity boundary. For these vertices that have experienced physical interference, the system must forcibly calculate the instantaneous mechanical compression displacement required for the spatial vertex within the intrusion area to move out of the repulsion area along the normal direction of the three-dimensional spatial repulsion entity boundary.

[0138] For example, to accurately quantify the cascading displacement effect of this instantaneous pushing action on deep brain tissue, the system needs to calculate not only the displacement of the contact surface apex but also the gradient deformation caused by stress propagation into the deeper layers. Here, the system uses the following spatial attenuation pushing algorithm to calculate the instantaneous mechanical compression displacement:

[0139] ;

[0140] in, This represents the instantaneous mechanical compression displacement vector generated by a single spatial vertex at the moment it is moved by the puller, calculated by the system. This represents the absolute geometric intrusion depth scalar of the current intrusion into the original soft tissue virtual space by the boundary of the repulsive entity in three-dimensional space. This represents the unit normal vector at the nearest point where the boundary of a repulsive entity in three-dimensional space collides with a vertex in that space, and is used to indicate the direction in which the organization is forcibly pushed away; This represents the shortest three-dimensional Euclidean distance from the vertex of the space to the geometric surface of the boundary of the three-dimensional repulsive entity in an unforced state. This represents the pre-configured elastic deformation space transmission attenuation constant, which characterizes the dissipation rate of mechanical stress during transmission in a porous elastic medium such as brain tissue.

[0141] It should be noted that the engineering application significance of the above formula lies in: through The product of these factors ensures that brain tissue adhering to the retractor surface is strictly expelled according to the direction and distance of the device's movement, thus resolving the model's erroneous penetration problem; simultaneously, through the natural exponential term... The introduction of this feature allows the system to realistically simulate how the pushing force on deep tissues far from the retractor changes with distance. The increase in retraction exhibits a physical phenomenon of exponential and rapid decay. This calculation process ensures that within seconds of the doctor inserting the retractor, the relevant brain regions in the navigation screen can undergo instantaneous elastic avoidance consistent with hydrodynamic properties.

[0142] Alternatively, once the surgeon has pushed the retractor to the desired position, it is usually locked and secured to the skull fixation frame using a serpentine robotic arm. At this point, the retractor remains absolutely stationary in space, continuously widening the surgical channel. However, the actual physical deformation does not stop. Due to the continuous and constant compression of the brain tissue, the interstitial fluid inside is slowly squeezed out, and the solid framework of the tissue undergoes stress relaxation and rheology, causing the pushed-out brain tissue to slowly collapse further outward over time.

[0143] Without compensating for this delayed deformation, the deviation between the navigation boundary and the actual tissue will be significantly amplified after half an hour of surgery. At this point, the system automatically switches to the steady-state compression creep calculation stage. In this stage, when the displacement rate of the brain retractor approaches zero (i.e., the velocity is below the set static threshold) and is in a static traction holding state, the system stops calculating elastic displacement and instead initiates rheological calculations based on time integration. Based on the viscoelastic mechanical parameters of the brain tissue and the continuous compression time variable, the system calculates the secondary collapse extension displacement of the spatial vertex under constant pressure over time.

[0144] Specifically, in order to accurately calculate this time-dependent tissue creep within discrete time steps, the system employs a time-domain integration algorithm based on the generalized Kelvin-Voigt rheological model, and its specific formula for the evolution of secondary extension displacement is as follows:

[0145] ;

[0146] in, This represents the vector of secondary collapse and extension displacement of the spatial vertices due to prolonged constant pressure, obtained from the system solution. This represents the instantaneous mechanical compression displacement vector of the vertex recorded by the system at the last moment before the retractor is locked and stationary, which serves as the stress background input for creep calculation; The dimensionless proportionality coefficient representing the maximum creep elongation of a material in the viscoelastic mechanical parameters of brain tissue; This represents the accumulated continuous compression time variable of the retractor since it enters the static tension holding state, which is generated in real time by a high-precision timer inside the system. The characteristic relaxation time constant in the viscoelastic mechanical parameters of brain tissue reflects the time span required for the drainage of fluid from the brain tissue to reach a new stress equilibrium.

[0147] For example, the key role of this formula in the actual system operation is that it establishes a compensation vector that grows non-linearly over time. In the initial stage after the puller has just locked ( (Smaller), brain tissue undergoes noticeable creep displacement due to rapid fluid infiltration; with increasing pressure time... The continuous increase and far exceed the characteristic relaxation time constant When the exponential term approaches zero, the secondary collapse extension displacement... Gradually approaching The system determines the maximum physical limit at which the brain tissue reaches a new state of static equilibrium. It accurately captures and outputs this slow-growing, elongated displacement.

[0148] like Figure 7 This figure objectively demonstrates the dynamic displacement evolution of brain tissue caused by continuous compression from a brain retractor during the establishment of a deep surgical channel. The horizontal axis represents the duration of continuous compression in seconds, and the vertical axis represents the total deformation displacement in millimeters.

[0149] The blue curve in the figure represents the deformation data during the transient elastic displacement calculation stage. In the initial stage from 0 to 20 seconds, as the doctor advances the brain retractor, the blue curve shows a steeply rising waveform with an extremely high slope, and the total deformation displacement rapidly increases to nearly 12 millimeters. This reflects the rapid physical displacement effect generated by the brain tissue when subjected to instantaneous mechanical compression by the device, proving that the system can capture changes in device pose at high frequency and output the instantaneous mechanical compression displacement with low latency.

[0150] Once the time exceeds 20 seconds and the brain retractor remains in a static traction state, the curve changes from the blue transient shift to the red steady-state compressive creep calculation stage. The red curve no longer shows a steep rise, but instead transforms into a gentle nonlinear growth trend. In the following few hundred seconds, due to the slow outflow of interstitial fluid from the brain tissue, stress relaxation of the solid skeleton is triggered, and the secondary collapse extension displacement gradually accumulates, causing the total deformation displacement to slowly climb from 12 mm and approach the physical limit of 18 mm.

[0151] This continuous waveform transformation from blue to red and from steep to gentle demonstrates that the two-stage deformation compensation algorithm used in this scheme can realistically simulate the viscoelastic collapse of brain tissue over time, effectively compensating for the lag defects of static navigation.

[0152] However, observing the data pattern in the latter part of the red curve reveals that the curve exhibits significant low-frequency oscillations and irregular sensor noise fluctuations as it approaches the horizontal asymptote, failing to converge to an ideal, smooth hydrostatic equilibrium state. This indicates that the adaptive decay adjustment mechanism of the time integral parameter in the current system's built-in rheological model still exhibits some damping hysteresis when dealing with complex nonlinear cerebrospinal fluid seepage over long time scales. In the future, introducing a higher-order fractional rheological model can further improve the smoothness and disturbance resistance of the data under long-term constant pressure.

[0153] It is also important to note that after acquiring the deformation data representing transient push-pull and steady-state maintenance, the AI ​​deformation calculation thread superimposes the instantaneous mechanical compression displacement and the secondary collapse extension displacement onto the corresponding positions of the spatial elastic deformation compensation field matrix in real time. Since the spatial elastic deformation compensation field matrix is ​​the sole reference for global coordinate mapping, this superposition operation is equivalent to locally injecting a high-energy forced displacement field directly driven by the instrument onto the original basic deformation field. The system fuses the two through matrix addition, thereby dynamically correcting the spatial coordinates of the current virtual tumor boundary under forced traction with extremely high precision.

[0154] In summary, considering the performance of current technologies in deep craniocerebral surgery, most current navigation systems can only handle static brain collapse caused by resection, or require the surgeon to frequently stop the operation to operate the intraoperative ultrasound probe for time-consuming image registration. This not only increases the risk of surgical infection but also fails to reflect the forced displacement caused by the retractor in real time. When the surgeon blindly operates deep in the traction area while looking at a severely deviated navigation image, there is a high risk of tearing deep veins or critical nerve bundles that have shifted in position due to traction.

[0155] The two-stage deformation compensation scheme disclosed in this embodiment transforms the surgical instrument itself into a physical pressure boundary source and introduces an advanced biorheological model. The system can provide instantaneous elastic avoidance with no delay when the surgeon moves the retractor, and accurately simulates the subtle viscoelastic collapse of brain tissue over time when the retractor is stationary. This technical solution eliminates the need for any complex intraoperative scanning procedures; relying solely on instrument trajectory tracking and physical evolution algorithms, it provides a truly biomechanically sensitive and dynamically evolving precise navigation boundary for complex surgeries such as deep sulci access, greatly improving the positioning safety of deep, hidden anatomical structures in the surgical area and the overall efficiency of the surgical operation.

[0156] Example 4:

[0157] like Figure 8 As shown, this embodiment provides a method embodiment directly corresponding to the aforementioned system. Addressing the technical problems mentioned in the background art, such as the inability of existing surgical navigation systems to simultaneously handle high-frequency instrument tracking and high-computational-power deformation calculation when faced with nonlinear spatial deformation of brain tissue during surgery, and the lack of an objective reliability assessment mechanism, this embodiment provides an AI-based visualized boundary navigation method for dynamic resection of intracranial tumors.

[0158] In practical applications, this method typically relies on a medical neuronavigation workstation equipped with high-performance heterogeneous computing resources (such as a multi-core CPU paired with a high-performance graphics accelerator card) for execution. The method specifically includes the following steps:

[0159] Step S1: Real-time acquisition of spatial pose data of surgical instruments, spatial coordinates of patient head markers, real-time images of intraoperative brain tissue and coordinates of discrete tissue feature points, and acquisition of a preoperative planning model including initial virtual tumor boundaries, key anatomical structures and surgical paths.

[0160] In actual clinical hardware configurations, spatial data acquisition primarily relies on a binocular infrared optical tracking camera system mounted at the head of the operating table. After anesthesia, the patient's head is fixed to a rigid head frame (such as a Mayfield head frame), and the system uses a passive optical reflective sphere array mounted on the rigid head frame as a marker for the patient's head to establish a dynamic reference benchmark for the patient's physical space.

[0161] Surgical instruments (such as suction devices and bipolar electrocoagulation forceps) are also equipped with specific optical reflective spheres at their tails. The optical tracking camera system captures the three-dimensional physical coordinates of these reflective spheres at a high frame rate, and then calculates the six-degree-of-freedom spatial pose data of the surgical instruments.

[0162] For real-time intraoperative imaging of brain tissue, the system uses an intraoperative ultrasound probe with a spatial positioning base to scan before and after the dura mater is cut, acquiring two-dimensional or three-dimensional B-mode ultrasound images. The coordinates of discrete tissue feature points are obtained by the surgeon using a spatially calibrated cortical surface tracking probe to make contact-based point markings on the exposed lesion surface. Furthermore, the workstation imports the patient's MRI or diffusion tensor imaging (DTI) data from the hospital's Picture Archiving and Communication System (PACS) before surgery. After three-dimensional reconstruction and semantic segmentation using medical image processing software, a preoperative planning model is generated. This model includes static initial virtual tumor boundaries and key anatomical structures such as nerve fiber bundles.

[0163] Step S2: Run the master tracking thread configured with the first preset priority to render the surgical instrument cursor, and run the AI ​​deformation solving thread asynchronous with the master tracking thread.

[0164] To address the hardware resource contention between high-frequency optical tracking and high-load deformation calculations, the underlying real-time operating system (RTOS) of the neural navigation workstation employs a physical core isolation and thread decoupling strategy. The system binds the main tracking thread to a specific priority CPU core and assigns it a first preset priority (i.e., the highest level of real-time interrupt response priority). This thread's sole responsibility is to read positioning data and perform lightweight rigid coordinate transformation matrix calculations, thereby rendering a crosshair representing the position of the surgical instrument tip within the monitor or microscope's field of view with extremely low latency, ensuring the continuity of the surgeon's hand-eye coordination.

[0165] Meanwhile, the system runs an AI deformation calculation thread on the graphics acceleration computing card. This thread has a lifecycle independent of the main control tracing thread. The two only exchange necessary data through a lock-free shared memory area, thereby ensuring that frame rate fluctuations caused by complex AI tensor operations will not be propagated up to the cursor rendering pipeline.

[0166] Step S3: The global three-dimensional nonlinear offset is calculated through the AI ​​deformation calculation thread to generate a spatial elastic deformation compensation field matrix. The deformation prediction confidence is calculated through multimodal cross-validation to generate a corresponding confidence field matrix. Based on the confidence field matrix, the target update region with a deformation prediction confidence lower than the first confidence threshold is detected to perform local deformation field recalculation.

[0167] During the actual surgical phase, due to cerebrospinal fluid release and brain tissue collapse, the physical space deviates from the preoperative imaging space. At this time, the AI ​​deformation calculation thread runs a three-dimensional deformation neural network in the background, outputting a global three-dimensional nonlinear offset covering the entire surgical field space, and discretizing it into a spatial elastic deformation compensation field matrix containing multi-level grid resolution. To objectively verify the physical reliability of this matrix, the system performs multimodal cross-validation: the spatial similarity and Euclidean distance of the deformation boundary predicted by the AI ​​are compared with the actual echo boundary of intraoperative ultrasound obtained in step S1 and the physical contact points collected by the cortical surface tracking probe. Combined with the uncertainty variance of the model itself, the deformation prediction confidence of each grid node in the space is calculated, and a confidence field matrix is ​​generated.

[0168] During continuous monitoring, if the system finds that the deformation prediction confidence level for a certain continuous spatial region is lower than the preset first confidence threshold, it determines that the deformation data output by the current AI for that region deviates from the actual physical feedback. The system then marks that region as the target update region, suspends the direct output for that region, and instead calls a nonlinear least squares optimization algorithm based on physical constraints to recalculate the local deformation field. It then uses actual measured ultrasonic and probe data to forcibly fit the displacement field of that region until its confidence level recovers to a safe range.

[0169] Step S4: Based on the spatial pose data of the surgical instruments, local interpolation remapping is performed in conjunction with the spatial elastic deformation compensation field matrix to dynamically update the initial virtual tumor boundary and generate the current virtual tumor boundary.

[0170] After completing the underlying matrix calculation and security verification, the system enters the coordinate mapping and mesh reconstruction stage. The workstation uses the spatial pose data of the surgical instruments as an index to find the corresponding displacement vector in the spatial elastic deformation compensation field matrix. Through three-dimensional linear interpolation or spline interpolation algorithms, the system applies the corresponding displacement vector to the vertices of the surface mesh of the initial virtual tumor boundary generated before surgery. This process transforms the originally static and rigid preoperative image boundary into the current virtual tumor boundary that can undergo nonlinear distortion and retreat synchronously with the collapse of real brain tissue and the traction of brain sulci. The local interpolation remapping mechanism ensures the smoothness and continuity of the boundary surface during the deformation process, avoiding the tearing or self-intersection of polygonal meshes.

[0171] Step S5: Perform hierarchical visualization rendering of the current virtual tumor boundary based on the confidence field matrix.

[0172] The final visual feedback is primarily presented through an augmented reality (AR) surgical microscope (such as a surgical microscope eyepiece with a holographic projection light path) connected to the workstation's video interface or a separate high-definition navigation display. Instead of rendering the entire tumor outline as a single solid line, the system introduces a graphics shader engine bound to a confidence field matrix. The rendering engine reads the deformation prediction confidence of each part of the current virtual tumor boundary and performs hierarchical visualization rendering. For example, in areas with confidence scores above a safety threshold, the boundary is rendered as a stable solid line with high contrast; while in boundary portions near or within the target update area, the rendering engine automatically replaces them with dynamic lines featuring gradient transparency, different warning colors (such as yellow dashed lines), or even specific flashing frequencies. In this way, the complex mathematical statistical confidence scores in the background are directly transformed into intuitive visual warning information in the surgical field.

[0173] In summary, the method provided in this embodiment deeply integrates physical tracking of hardware devices, multi-threaded scheduling of the operating system, underlying AI deformation calculation, and front-end AR visualization rendering. The beneficial effects of this method in practical surgical applications are as follows: its master-slave asynchronous dual-track operation architecture effectively ensures the smooth tracking of surgical instruments in any complex computing environment, avoiding the screen stuttering risks common in conventional single-threaded systems; simultaneously, through the confidence assessment system generated by multimodal cross-validation and hierarchical visualization rendering technology, the system can not only actively correct navigation deviations caused by brain tissue collapse, but also intuitively demonstrate the reliability of the current virtual boundary to the surgeon.

[0174] This method breaks through the limitations of traditional static navigation systems that blindly rely on preoperative data, reduces anatomical positioning deviations caused by severe tissue deformation, and can more objectively and accurately guide surgical operations when removing deep tumors, thereby reducing the potential clinical risk of accidentally damaging key nerves and blood vessels.

[0175] Example 5:

[0176] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0177] like Figure 9 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0178] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0179] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 9 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0180] The memory 103 is used to store a computer program corresponding to the AI ​​visualization boundary navigation method for dynamic resection of intracranial tumors according to the above embodiments of the present invention. This computer program is controlled and executed by the processor 101. The processor 101 is used to execute the computer program stored in the memory 103 to implement the content shown in the foregoing method embodiments.

[0181] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 9 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0182] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An AI-powered visual boundary navigation system for dynamic resection of intracranial tumors, characterized in that, include: The positioning and acquisition module is used to acquire in real time the spatial pose data of surgical instruments, the spatial coordinates of patient head markers, real-time images of intraoperative brain tissue and coordinates of discrete tissue feature points, as well as to obtain a preoperative planning model containing the initial virtual tumor boundary, key anatomical structures and surgical path. The navigation controller includes a main control tracking thread configured with a first preset priority, and an AI deformation calculation thread that runs asynchronously with the main control tracking thread. The main control tracking thread is used to render surgical instrument cursors. The AI ​​deformation calculation thread is used to calculate global three-dimensional nonlinear offsets to generate a spatial elastic deformation compensation field matrix, calculate deformation prediction confidence through multimodal cross-validation to generate a corresponding confidence field matrix, and detect target update regions with deformation prediction confidence lower than a first confidence threshold based on the confidence field matrix, so as to perform local deformation field recalculation. The coordinate mapping module is used to perform local interpolation remapping based on the spatial pose data of the surgical instruments and the spatial elastic deformation compensation field matrix, so as to dynamically update the initial virtual tumor boundary and generate the current virtual tumor boundary. The intelligent AR display module is used to perform hierarchical visualization rendering of the current virtual tumor boundary based on the confidence field matrix.

2. The AI-based visual boundary navigation system for dynamic resection of intracranial tumors according to claim 1, characterized in that, The main control tracking thread runs at a first preset frequency to read the latest spatial elastic deformation compensation field matrix and the confidence field matrix from shared memory, and convert the physical coordinates of the surgical instrument tip into virtual scene coordinates; The AI ​​deformation calculation thread runs at an adaptive frequency. When it detects that the rate of change of the spatial pose data of the surgical instrument exceeds the first rate of change threshold, the AI ​​deformation calculation thread automatically increases its running frequency to a second preset frequency, which is higher than the normal running frequency.

3. The AI-based visual boundary navigation system for dynamic resection of intracranial tumors according to claim 1, characterized in that, The spatial elastic deformation compensation field matrix contains four grid resolution levels; The four grid resolution levels sequentially include: a key structure-specific layer covering the key anatomical structures, a core layer with a first radius range set centered on the current virtual tumor boundary, a transition layer extending outward from the core layer to a second radius range, and an outer layer covering the remaining surgical area. The mesh resolution of the critical structure-specific layer is higher than that of the core layer.

4. The AI-based visual boundary navigation system for dynamic resection of intracranial tumors according to claim 1, characterized in that, The AI ​​deformation calculation thread calculates the deformation prediction reliability by weighted fusion of scores normalized to three independent dimensions: Model uncertainty dimension: The standard deviation of the prediction results under the same input is calculated based on multiple Monte Carlo inferences and mapped to the first dimension score; Multimodal consistency dimension: Calculate the degree of matching between the real-time images of the intraoperative brain tissue and the coordinates of the discrete tissue feature points and the AI-predicted deformation field, and map it into a second dimension score; Clinical prior constraint dimension: The physiological rationality of the deformation results is verified based on the prior knowledge base of cranial anatomy and mapped to the third dimension score; The deformation prediction reliability is the sum of the products of the fixed weight coefficients corresponding to the three dimensions and the corresponding dimension scores.

5. The AI-based visual boundary navigation system for dynamic resection of intracranial tumors according to claim 1, characterized in that, The automatic triggering conditions for recalculating the local deformation field in the target update region include: Based on the confidence field matrix, a region with a continuous spatial volume greater than a first preset volume threshold and a deformation prediction confidence lower than a second confidence threshold is detected, and this region is located on the surgical path or is less than a first preset safety distance threshold from the key anatomical structure. The second confidence threshold is less than the first confidence threshold.

6. The AI-based visual boundary navigation system for dynamic resection of intracranial tumors according to claim 1, characterized in that, When the AI ​​deformation calculation thread performs the recalculation of the local deformation field, it uses a weighted constraint energy function as the optimization objective. The weighted constraint energy function includes a data term and a smoothing term. The data term is the weighted sum of squares of the difference between the actual displacement measurement value and the predicted displacement value of the discrete tissue feature point in the target update region. The smoothing term is the sum of squares of the difference between the displacement values ​​to be solved for adjacent spatial vertices. The weighted constraint energy function is minimized using the Gauss-Newton optimization algorithm to complete the recalculation of the local deformation field.

7. The AI-based visual boundary navigation system for dynamic resection of intracranial tumors according to claim 5, characterized in that, The coordinate mapping module uses a confidence-guided progressive smooth interpolation function for boundary updates; The final updated displacement of the spatial vertex is calculated by weighting the displacement before update and the new calculated displacement according to the interpolation coefficients; The magnitude of the interpolation coefficient is positively correlated with the deformation prediction confidence of the spatial vertex. When the deformation prediction confidence is lower than the second confidence threshold, the interpolation coefficient is set to zero and the displacement update of the spatial vertex is paused until the local deformation field is recalculated.

8. The AI-based visual boundary navigation system for dynamic resection of intracranial tumors according to claim 5, characterized in that, The intelligent AR display module uses a multi-level visualization encoding scheme to render the current virtual tumor boundary; The multi-level visualization coding scheme matches the corresponding boundary line type, color, transparency, and line width according to the numerical range to which the deformation prediction confidence belongs. When the deformation prediction confidence is lower than the second confidence threshold, the boundary of the range is configured with a dynamic display feature of a first preset flashing frequency.

9. The AI-based visual boundary navigation system for dynamic resection of intracranial tumors according to claim 8, characterized in that, The intelligent AR display module is also equipped with a hierarchical dynamic risk warning mechanism; When the distance between the surgical instrument and the boundary where the deformation prediction confidence is lower than the first confidence threshold is less than the second preset safety distance threshold, a sound prompt at the first preset audio playback frequency is triggered. When the distance is further less than the third preset safe distance threshold, a sound prompt is triggered at a second preset audio playback frequency higher than the first preset audio playback frequency, and the flashing frequency of the corresponding boundary is increased to a second preset flashing frequency higher than the first preset flashing frequency. The third preset safety distance threshold is less than the second preset safety distance threshold.

10. The AI-based visual boundary navigation system for dynamic resection of intracranial tumors according to claim 4, characterized in that, The clinical prior constraint dimension relies on a cranial anatomy prior knowledge base that stores quantified physiological deformation constraint rules. The physiological deformation constraint rules include: the upper limit threshold for the maximum allowable deformation displacement of rigid structures, the upper limit threshold for the relative positional change of intracranial major blood vessels, the upper limit threshold for the change in the orientation angle of nerve fiber bundles, and constraint relationships used to maintain the connection order of the basic topological relationships between various anatomical structures without change.

Images