Spinal anesthesia puncture positioning guide system and method based on deep learning

Abstract

The application discloses a kind of intravertebral anesthesia puncture positioning guide system and method based on deep learning, it is related to medical ultrasonic image analysis field.The method utilizes planning ultrasound sequence, planning probe pose and planning back surface data to construct bone surface coupling graph and planning puncture channel;Current back surface data and bone surface coupling graph are input into morphing decoupling model, obtain soft tissue continuous morphing field, bone unit segmental transformation and bone anchor point priori uncertainty;Determine the verification position according to priori uncertainty, update segmental transformation and posteriori uncertainty by verification ultrasound data;Migrate planning puncture channel and judge its effectiveness, generate needle insertion area, angle range and expected depth interval when effective, generate rescan instruction when invalid.

Description

Technical Field

[0001] This invention relates to the field of medical ultrasound image analysis technology, and in particular to a deep learning-based spinal anesthesia puncture positioning and guidance system and method. Background Technology

[0002] Spinal anesthesia puncture typically requires determining the needle insertion area, angle, and expected depth based on bony structures such as the spinous process, laminae, and intervertebral spaces. Ultrasound imaging can visualize the bony echogenic boundaries of the lumbar vertebrae before puncture; however, after the planning scan, the patient may change position due to reflexion, tilting, breathing, or pain response. This results in continuous non-rigid deformation of the back skin and subcutaneous soft tissue, and relative rotation and translation between adjacent vertebrae may also occur, altering the correspondence between the back surface position, deep intervertebral spaces, and planned puncture direction established during the planning scan phase.

[0003] Current positioning methods typically use skin markers to maintain the planned position or migrate deep targets based on the overall registration results between the current back surface and the planned back surface. Skin markers cannot characterize segmental motion of deep vertebrae; simple surface registration easily mixes soft tissue compression, local curvature changes, and relative vertebral motion into the same deformation. When the planned position differs significantly from the current position, the migrated puncture channel may deviate from the original target intervertebral space or approach bony boundaries such as the spinous process, laminae, and facet joints.

[0004] Furthermore, repeatedly scanning all lumbar vertebrae increases operation time, while directly outputting localization results based solely on surface changes lacks a verification process for high-uncertainty bony structures and a technical means to determine whether the migration puncture channel remains effective. Therefore, a technical solution is needed that can establish a patient-specific relationship between bony anchor points and the back surface, distinguish between continuous soft tissue deformation and bony segmental movement, select verification locations based on prediction uncertainty, and determine whether to output localization information or re-scan instructions after verification. Summary of the Invention

[0005] This invention provides a deep learning-based spinal anesthesia puncture positioning and guidance system and method to solve the problems that when the planned position is inconsistent with the current position, changes on the back surface cannot accurately represent the movement of deep lumbar vertebral segments, and there is a lack of verification for uncertain bony structures and judgment of the effectiveness of migration puncture channels.

[0006] To achieve the aforementioned objectives, the present invention employs the following technical solution: a deep learning-based method for guiding spinal anesthesia puncture localization, comprising:

[0007] S1. Receive the planned ultrasound sequence, the planned probe pose, and the planned back surface data, construct a bone surface coupling map and plan the puncture channel. The bone surface coupling map represents the spatial relationship between bony anchor points and the mapping relationship between bony anchor points and the back surface.

[0008] S2. Receive the current back surface data, input it and the bone surface coupling map into the deformation decoupling model to obtain the soft tissue continuous deformation field, the segmental transformation of the bony unit and the prior uncertainty of the bony anchor point;

[0009] S3. Determine the verification position based on the prior uncertainty of the bony anchor point and generate a probe adjustment command, and receive the verification ultrasound data of the verification position and the current probe pose;

[0010] S4. Extract the current bony anchor point from the verification ultrasound data, match it with the bone surface coupling map, and update the segmental transformation of the bony unit and the posterior uncertainty of the bony anchor point;

[0011] S5. Based on the continuous deformation field of the soft tissue and the updated segmental transformation of the bony unit, the planned puncture channel is migrated to obtain the migrated puncture channel and the uncertainty of the channel position.

[0012] S6. Determine the effectiveness of the migration puncture channel based on the bone surface coupling diagram, the posterior uncertainty of the bone anchor point, and the channel position uncertainty;

[0013] S7. When the migration puncture channel is valid, generate the needle insertion area, angle range and expected depth range; when the migration puncture channel is invalid, generate a re-scan command.

[0014] Furthermore, step S1 specifically includes:

[0015] Based on the planned probe pose, the planned ultrasound sequence is transformed to the planned spatial coordinate system, and the planned bony anchor points corresponding to the spinous process, lamina, articular process, sacral continuous surface and intervertebral space are extracted from the transformed planned ultrasound sequence.

[0016] The planned bony anchor points are used as bony nodes, the corresponding surface projection points in the planned back surface data are used as surface nodes, and the anatomical adjacency relationship between bony nodes and the spatial mapping relationship between bony nodes and surface nodes are used as edges to construct the bone-surface coupling diagram.

[0017] A planned puncture channel that avoids bony boundaries is generated based on the interlaminar window of the target intervertebral space and the corresponding surface nodes.

[0018] Furthermore, the process of constructing the bone surface coupling map in step S1 also includes: receiving the planned contact pressure and the current contact pressure, querying the pre-calibrated monotonic mapping relationship between the contact pressure and the soft tissue compression amount based on the pressure difference between the two, and obtaining the soft tissue thickness compensation amount.

[0019] The distance between bony nodes and surface nodes in the bone-surface coupling diagram is corrected using the soft tissue thickness compensation amount, and the corrected distance is input into the deformation decoupling model.

[0020] When the current contact pressure exceeds the calibration range of the monotonic mapping relationship, the rescan command is generated.

[0021] Furthermore, in step S2, the deformation decoupling model includes a surface feature encoding branch, a bone map encoding branch, a coupling transformation decoding branch, and an uncertainty output branch;

[0022] The surface feature encoding branch extracts the deformation features between the planned back surface data and the current back surface data. The bony map encoding branch extracts the node and edge features of the bone surface coupling map. Under the constraints of continuous soft tissue deformation, rigidity of each bony unit, and unchanged adjacency relationship of bony nodes, the coupling transformation decoding branch outputs the continuous soft tissue deformation field and the segmental transformation of the bony unit. The uncertainty output branch outputs the prior uncertainty of the bony anchor point.

[0023] Further, step S3 specifically includes: generating a set of candidate verification locations on the current back surface data. Each candidate verification location includes the probe position and probe orientation, and is arranged according to...

[0024]

[0025] Determine the verification location ;in, For the number of bony anchor points, For the first The prior uncertainty of a bony anchor point To be at the candidate validation position The expected posterior uncertainty after verification is completed. For the expected visibility, The cost of the probe's availability, For bony acoustic shadowing amount, , and The weights are non-negative.

[0026] Further, step S4 specifically includes: transforming the verification ultrasound data to the current spatial coordinate system according to the current probe pose, and extracting the current bony anchor point and the confidence level of the current bony anchor point from the transformed verification ultrasound data;

[0027] Based on the anatomical hierarchy and adjacency of the bony nodes, the current bony anchor point is graph matched with the planned bony anchor point in the bone surface coupling graph to obtain bony topology consistency.

[0028] Anchor point matching residuals are generated based on the matching position difference, matching direction difference, and the current bony anchor point confidence. The pose of the corresponding bony unit and the posterior uncertainty of the bony anchor point are then updated using the anchor point matching residuals.

[0029] Further, step S6 specifically includes: generating an expanded bony boundary along the normal of the bony boundary according to the expansion distance positively correlated with the posterior uncertainty of the bony anchor point;

[0030] Detect whether the migration puncture channel passes through the interlaminar window corresponding to the target intervertebral space and whether it intersects with the expanded bony boundary;

[0031] The migration puncture channel is deemed valid when it passes through the interlaminar window, does not intersect the expanded bony boundary, and the bony topology consistency, the anchor point matching residual, and the channel position uncertainty all meet the corresponding thresholds.

[0032] Further, step S7 specifically includes: determining the overlapping area of ​​the migration puncture channel and the current back surface data as the needle insertion area; determining the angle range based on the central axis of the migration puncture channel and the channel position uncertainty; and determining the expected depth range based on the path length from the needle insertion area to the target intervertebral space and the soft tissue thickness compensation amount.

[0033] The system continuously compares subsequent back surface data with the current back surface data. When the surface displacement or surface curvature change exceeds the corresponding threshold, the needle insertion area, the angle range, and the expected depth range are cancelled, and the rescan instruction is generated.

[0034] Another objective of this invention is to provide a deep learning-based spinal anesthesia puncture localization and guidance system, the system comprising:

[0035] The mapping module constructs a bone-surface coupling map and a planned puncture channel based on the planned ultrasound sequence, planned probe pose, and planned back surface data. The bone-surface coupling map represents the spatial relationship between bony anchor points and the mapping relationship between bony anchor points and the back surface.

[0036] The deformation decoupling module generates a continuous soft tissue deformation field, segmental transformation of bony units, and prior uncertainty of bony anchor points based on the current back surface data and the bone surface coupling diagram.

[0037] The verification module determines the verification position based on the prior uncertainty of the bony anchor point and generates a probe adjustment command, and receives the verification ultrasound data of the verification position and the current probe pose.

[0038] The correction module extracts the current bony anchor point from the verification ultrasound data, matches it with the bone surface coupling map, and updates the segmental transformation of the bony unit and the posterior uncertainty of the bony anchor point.

[0039] The migration module migrates the planned puncture channel based on the continuous deformation field of the soft tissue and the updated segmental transformation of the bony unit, generating the migration puncture channel and the channel position uncertainty.

[0040] The judgment module determines the validity of the migration puncture channel based on the bone surface coupling diagram, the posterior uncertainty of the bone anchor point, and the channel position uncertainty.

[0041] The output module generates the needle insertion area, angle range, and expected depth range when the migration puncture channel is valid, and generates a re-scan command when the migration puncture channel is invalid.

[0042] Furthermore, it also includes an ultrasonic device, a surface acquisition device, a pose acquisition device, a pressure acquisition device, and a coordinate transformation module;

[0043] The pose acquisition device includes an optical tracking marker or electromagnetic tracking sensor disposed on the ultrasound probe, and the pressure acquisition device is disposed between the grip part of the ultrasound probe and the transducer; the coordinate transformation module, based on the calibration parameters from the ultrasound image coordinate system, the probe coordinate system, and the surface acquisition device coordinate system to the tracking coordinate system, transforms the planned ultrasound sequence, the verification ultrasound data, the planned back surface data, and the current back surface data to the same spatial coordinate system, and corrects the distance between the bony anchor point and the back surface according to the contact pressure output by the pressure acquisition device.

[0044] The beneficial effects of this invention are as follows:

[0045] 1. This invention constructs a bone-surface coupling map by planning ultrasound sequences, planning probe poses, and planning back surface data, so that the anatomical spatial relationship between planned bony anchors and the mapping relationship between planned bony anchors and the back surface are stored in the same data structure, providing individualized constraints for patients to migrate deep intervertebral spaces and plan puncture channels under the current body position;

[0046] 2. The deformation decoupling model of the present invention decomposes back changes into a continuous deformation field of soft tissue and segmental transformation of bony units, avoiding the direct equivalent of continuous non-rigid deformation of skin, fat and muscle to vertebral movement, so that the surface side part and the bony target side part of the planned puncture channel can migrate according to different physical properties.

[0047] 3. This invention determines the verification location based on the prior uncertainty of the bony anchor point, and uses the verification ultrasound data to update the segmental transformation of the bony unit and the posterior uncertainty of the bony anchor point, so that the limited current ultrasound observations are concentrated on the bony anchor points that have a greater impact on the migration results, forming a data closed loop consisting of model prediction, directional observation and posterior correction.

[0048] 4. This invention uses bone surface coupling diagrams, posterior uncertainty of bone anchor points, and channel position uncertainty to jointly determine the validity of the migration puncture channel. It generates the needle insertion area, angle range, and expected depth interval only when the channel is valid, and generates a re-scan instruction when the channel is invalid. This reduces the possibility of continuing to use the failed planning results after a change in body position. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the overall process of the deep learning-based spinal anesthesia puncture localization and guidance method according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the bone surface coupling diagram, the continuous deformation field of soft tissue, the segmental transformation of bone units, and the migration relationship of the puncture channel in an embodiment of the present invention.

[0051] Figure 3 This is a flowchart of the candidate verification location selection, current bony anchor point extraction, graph matching, and posterior uncertainty update in an embodiment of the present invention.

[0052] Figure 4 This is a structural block diagram of a deep learning-based spinal anesthesia puncture positioning and guidance system according to an embodiment of the present invention.

[0053] Figure 5 This is a schematic diagram illustrating the puncture positioning and guidance principle in an embodiment of the present invention.

[0054] In the diagram: 100, Processor; 110, Mapping Module; 120, Deformation Decoupling Module; 130, Verification Module; 140, Correction Module; 150, Migration Module; 160, Judgment Module; 170, Output Module; 180, Coordinate Transformation Module; 200, Ultrasonic Device; 210, Ultrasonic Probe; 300, Surface Acquisition Device; 400, Pose Acquisition Device; 410, Optical Tracking Mark; 420, Electromagnetic Tracking Sensor; 500, Pressure Acquisition Device; 600, Display Device; 810, Back Surface; 820, Bony Boundary; 830, Interlaminar Window; 840, Needle Insertion Area; 850, Migration Puncture Channel; 860, Target Area. Detailed Implementation

[0055] To make the content of this invention easier to understand, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings.

[0056] like Figure 4 As shown, the present invention provides a deep learning-based spinal anesthesia puncture positioning and guidance system, including a mapping module, a deformation decoupling module, a verification module, a correction module, a transfer module, a judgment module, and an output module.

[0057] The mapping module constructs a bone-surface coupling map and a planned puncture channel based on the planned ultrasound sequence, planned probe pose, and planned back surface data. The bone-surface coupling map represents the spatial relationship between bony anchor points and the mapping relationship between bony anchor points and the back surface.

[0058] The deformation decoupling module generates a continuous soft tissue deformation field, segmental transformation of bony units, and prior uncertainty of bony anchor points based on the current back surface data and bone surface coupling diagram.

[0059] The verification module determines the verification position based on the prior uncertainty of the bony anchor point and generates probe adjustment instructions, and receives the verification ultrasound data of the verification position and the current probe pose.

[0060] The correction module extracts the current bony anchor point from the validation ultrasound data, matches it with the bone surface coupling map, and updates the segmental transformation of the bony unit and the posterior uncertainty of the bony anchor point.

[0061] The migration module plans the puncture channel based on the continuous deformation field of soft tissue and the updated segmental transformation of bone units, and generates the migration puncture channel and the channel position uncertainty.

[0062] The judgment module determines the validity of the migration puncture channel based on the bone surface coupling diagram, the posterior uncertainty of the bone anchor point, and the uncertainty of the channel position.

[0063] The output module generates the needle insertion area, angle range, and expected depth range when the migration puncture channel is valid, and generates a re-scan instruction when the migration puncture channel is invalid.

[0064] Specifically, such as Figure 4 As shown, the puncture positioning guidance system includes a processor 100, a mapping module 110, a deformation decoupling module 120, a verification module 130, a correction module 140, a migration module 150, a judgment module 160, and an output module 170. The mapping module 110 is connected to the ultrasound device 200, the pose acquisition device 400, and the surface acquisition device 300. Each module then transmits data with a unified coordinate system and timestamps in sequence according to the method steps.

[0065] The output of the mapping module 110 is connected to the deformation decoupling module 120 and the migration module 150; the output of the deformation decoupling module 120 is connected to the verification module 130, the migration module 150, and the judgment module 160, respectively; the correction module 140 connects the updated segmental transformation and posterior uncertainty to the migration module 150 and the judgment module 160, respectively. The valid or invalid state of the judgment module 160 is connected to the output module 170, and the display device 600 presents positioning guidance information or a rescan command.

[0066] Furthermore, the system also includes an ultrasound device, a surface acquisition device, a pose acquisition device, a pressure acquisition device, and a coordinate transformation module. The pose acquisition device includes an optical tracking mark or an electromagnetic tracking sensor mounted on the ultrasound probe, and the pressure acquisition device is mounted between the grip of the ultrasound probe and the transducer. The coordinate transformation module, based on calibration parameters from the ultrasound image coordinate system, the probe coordinate system, and the surface acquisition device coordinate system to the tracking coordinate system, transforms the planned ultrasound sequence, verified ultrasound data, planned back surface data, and current back surface data to the same spatial coordinate system, and corrects the distance between the bony anchor point and the back surface based on the contact pressure output by the pressure acquisition device.

[0067] Specifically, such as Figure 4 and Figure 5 As shown, the ultrasound device 200 includes an ultrasound main unit and a low-frequency convex array ultrasound probe 210. The surface acquisition device 300 is a depth camera whose field of view covers the target lumbar spine region. The pose acquisition device 400 uses an optical tracking marker 410 or an electromagnetic tracking sensor 420, and performs calibration in the corresponding tracking space.

[0068] The pressure acquisition device 500 includes a force sensor, a signal conditioning circuit, and a sampling interface. The force sensor is located between the grip and the transducer, with its sensing direction aligned with the center of the sound beam; the normal force divided by the calibrated effective contact area forms the contact pressure. The pressure sampling frequency is not lower than the ultrasound frame rate, and the sampling results are associated with the corresponding ultrasound frames according to timestamps. The display device 600 is connected to the output module 170 and is used to display probe adjustment commands, needle insertion area 840, angle range, expected depth range, and rescan commands.

[0069] The coordinate transformation module 180 obtains the transformations from the ultrasound image coordinate system to the probe coordinate system, from the probe coordinate system to the tracking coordinate system, and from the surface acquisition device coordinate system to the tracking coordinate system during the system calibration phase. For homogeneous coordinates in the ultrasound image coordinate system... Its coordinates in a unified spatial coordinate system It can be represented as:

[0070]

[0071] in, This is the homogeneous transformation matrix from the ultrasound image coordinate system to the probe coordinate system. This is a homogeneous transformation matrix from the probe coordinate system to a unified spatial coordinate system. The planned back surface data and the current back surface data are transformed using this homogeneous transformation matrix from the surface acquisition device coordinate system to the unified spatial coordinate system. Each transformation matrix consists of rigid body rotation and translation, and after calibration, error verification is performed using a positioning body with known geometric dimensions.

[0072] like Figure 5 As shown, the puncture needle travels along the migration puncture channel 850 from the insertion area 840 toward the target area 860. The system does not require a rigid connection between the ultrasound probe 210 and the puncture needle. An isolation sleeve is provided on the outer surface of the ultrasound probe 210 and the pressure acquisition device 500 that come into contact with the patient; the zero-point pressure is recalibrated after the isolation sleeve is installed.

[0073] Based on the above system, this invention provides a deep learning-based method for guiding spinal anesthesia puncture localization, such as... Figure 1As shown, the procedure includes: Step S1: Receiving the planned ultrasound sequence, planned probe pose, and planned back surface data; constructing a bone-surface coupling map and planning the puncture channel; the bone-surface coupling map characterizes the spatial relationship between bony anchor points and the mapping relationship between bony anchor points and the back surface; Step S2: Receiving the current back surface data and inputting it and the bone-surface coupling map into the deformation decoupling model to obtain the soft tissue continuous deformation field, the segmental transformation of bony units, and the prior uncertainty of bony anchor points; Step S3: Determining the verification position based on the prior uncertainty of bony anchor points and generating probe adjustment instructions; receiving the verification ultrasound data of the verification position and the current probe pose; Step S... 4. Extract the current bony anchor point from the verification ultrasound data, match it with the bone surface coupling map, and update the segmental transformation of the bony unit and the posterior uncertainty of the bony anchor point; Step S5. Based on the continuous deformation field of the soft tissue and the updated segmental transformation of the bony unit, migrate and plan the puncture channel to obtain the migration puncture channel and the channel position uncertainty; Step S6. Determine the validity of the migration puncture channel based on the bone surface coupling map, the posterior uncertainty of the bony anchor point, and the channel position uncertainty; Step S7. When the migration puncture channel is valid, generate the needle insertion area, angle range, and expected depth interval; when the migration puncture channel is invalid, generate a re-scan command.

[0074] Specifically, such as Figure 1 As shown, step S1 is the planning data modeling stage. The ultrasound device 200, the pose acquisition device 400, and the surface acquisition device 300 respectively acquire the planned ultrasound sequence, the planned probe pose, and the planned back surface data. The coordinate transformation module 180 aligns the data according to the timestamp and calibration parameters, and the mapping module 110 generates a bone-surface coupling map and a planned puncture channel that avoids the bony boundary 820.

[0075] In step S2, the deformation decoupling module 120 decomposes the current back surface change into a three-dimensional vector field characterizing the continuous displacement of soft tissue, as well as the rigid body rotation and translation of each vertebra or bony node group, and outputs the prior uncertainty of the bony anchor point without correction by the current ultrasound observation.

[0076] Steps S3 and S4 constitute the orientation verification step. Verification module 130 selects a verification location that significantly contributes to uncertainty reduction and is reachable by the probe. Correction module 140 updates the segmental transformation of the bony unit and the posterior uncertainty using the current bony anchor point in the verification ultrasound data. When the image recognition confidence is low or the matching relationship is unstable, the posterior uncertainty is not unreasonably reduced.

[0077] In step S5, the migration module 150 interpolates to form a migration puncture channel 850 based on the surface end and target end of the migration channel according to the continuous deformation field of soft tissue and the updated segmental transformation of bone unit, and propagates the posterior uncertainty, surface registration error and deformation estimation error to form the channel position uncertainty.

[0078] Steps S6 and S7 constitute the judgment output stage. The judgment module 160 detects the target layer, bony boundary interval, and topological consistency of the migration puncture channel 850. When all conditions are met, the output module 170 generates the needle insertion area 840, angle range, and expected depth interval; if any necessary condition is not met, a re-scan instruction is generated.

[0079] In one embodiment, step S1 specifically includes: transforming the planned ultrasound sequence to a planned spatial coordinate system according to the planned probe pose; extracting the planned bony anchor points corresponding to the spinous process, lamina, articular process, sacral continuous surface, and intervertebral space from the transformed planned ultrasound sequence; using the planned bony anchor points as bony nodes, and the corresponding surface projection points in the planned back surface data as surface nodes, constructing a bone-surface coupling map using the anatomical adjacency relationship between bony nodes and the spatial mapping relationship between bony nodes and surface nodes as edges; and generating a planned puncture channel that avoids bony boundaries based on the interlaminar window of the target intervertebral space and the corresponding surface nodes.

[0080] Specifically, such as Figure 2 As shown, after coordinate transformation, the planned ultrasound sequence outputs pixel regions of the continuous surfaces of the spinous process, lamina, articular processes, and sacrum from the lumbar ultrasound segmentation network. These regions are then used to form planned bony anchor points based on contours, curvature extrema, and relationships with adjacent structures. The intervertebral space center is determined by the passable area between the edges of adjacent laminas. For each bony node, a surface projection point is located in the planned back surface data along the back surface normal or the direction of the ultrasound beam. When holes exist in the point cloud, local patches around the projection point are used as surface nodes.

[0081] Anatomical adjacency edges between bony nodes record intra-layer, adjacent-layer, and left-right relationships, while spatial mapping edges between bony nodes and surface nodes record planned distances, directions, and imaging visibility. The planned puncture channel shares nodes with the bone surface coupling map, and the interlaminar window 830 and needle insertion area 840 are updated synchronously when nodes are updated.

[0082] Furthermore, during the construction of the bone-surface coupling map, the planned contact pressure and the current contact pressure are received. Based on the pressure difference between the two, the pre-calibrated monotonic mapping relationship between the contact pressure and the soft tissue compression is queried to obtain the soft tissue thickness compensation amount. The soft tissue thickness compensation amount is used to correct the distance between the bony nodes and the surface nodes in the bone-surface coupling map, and the corrected distance is input into the deformation decoupling model. When the current contact pressure exceeds the calibration range of the monotonic mapping relationship, a rescan command is generated.

[0083] Specifically, such as Figure 5As shown, the pressure acquisition device 500 is positioned between the grip of the ultrasonic probe 210 and the transducer to measure the normal contact pressure applied by the transducer to the back surface 810. The planned contact pressure is acquired synchronously with the planned ultrasonic sequence, while the current contact pressure is acquired synchronously with the current back surface data or verification ultrasonic data. Both types of pressure are zero-point calibrated and low-pass filtered before being used in the calculation to avoid changes in the compensation results due to instantaneous peak values ​​caused by high-frequency hand tremors.

[0084] Before the system is put into use, calibration samples are established for the applicable probe contact area and back soft tissue thickness range. Using an unpressurized state or a specified baseline pressure state as a reference, the distance from the skin to the bone surface is measured by ultrasound, and the contact pressure is obtained. Corresponding compression amount Piecewise linear interpolation or monotonic-preserving cubic interpolation is used to establish a monotonic mapping relationship, and extrapolation is not performed outside the coverage area of ​​the calibration data. The planned contact pressure is... The current contact pressure is At that time, the soft tissue thickness compensation amount is:

[0085]

[0086] when When positive, the compression caused by the current contact pressure is greater than the compression caused by the planned contact pressure. For the distance from the bony node to the surface node along the local surface normal in the bone-surface coupling diagram... The correction distance can be expressed as The correction distance is used as the side feature input of the bony map coding branch in the deformation decoupling model, separating the surface subsidence caused by pressure from the movement of bony units caused by body position. Missing pressure data, continuous fluctuations exceeding a preset stability threshold, or... If the value is less than the lower limit of the calibration or greater than the upper limit of the calibration, the unverified compression amount will not be used; instead, a rescan command will be generated.

[0087] In one embodiment, in step S2, the deformation decoupling model includes a surface feature encoding branch, a bony map encoding branch, a coupling transformation decoding branch, and an uncertainty output branch. The surface feature encoding branch extracts the deformation features between the planned back surface data and the current back surface data. The bony map encoding branch extracts the node and edge features of the bone-surface coupling map. Under the constraints of continuous soft tissue deformation, rigidity of each bony unit, and unchanged adjacency relationship of bony nodes, the coupling transformation decoding branch outputs the continuous soft tissue deformation field and the segmental transformation of bony units. The uncertainty output branch outputs the prior uncertainty of the bony anchor point.

[0088] Specifically, such as Figure 2As shown, the surface feature encoding branch reads the planned back surface point cloud and the current back surface point cloud. Point features include 3D coordinates, surface normals, and local curvature. The point cloud is cropped to the target region and resampled to 4096 points. The number of output channels for the four-level point cloud encoding structure are 64, 128, 256, and 256, respectively.

[0089] The bony map encoding branch reads the bone-surface coupling map. Bony node features include 3D coordinates, anchor point category, vertebral body level, and planning recognition confidence. Surface node features include 3D coordinates, normal vector, and local curvature. Edge features include inter-node distance, direction vector, edge type, and tissue thickness. The bony map encoding branch employs a three-layer graph attention network, with each layer having a feature dimension of 128.

[0090] Two encoding branches are connected to the coupled transform decoding branch via a cross-attention layer. The cross-attention layer uses bony node features as queries and local features of the planned dorsal surface and the current dorsal surface as keys and values. The coupled transform decoding branch includes a soft tissue decoding head and a bony unit decoding head. The soft tissue decoding head outputs a 3D displacement vector, and the bony unit decoding head outputs rotation and translation parameters for each bony unit.

[0091] For the Nodes in a skeletal unit The segmental transformation is expressed as:

[0092]

[0093] in, Let be a rotation matrix. It is a translation vector. Let these be the predicted node coordinates. The rotation matrix satisfies:

[0094]

[0095] in, The identity matrix is ​​used. The bone unit decoder outputs a six-dimensional continuously rotated representation, which is then orthogonalized to obtain... This allows individual bony units to undergo only rotation and translation.

[0096] The uncertainty output branch is connected in parallel with the coupled transformation decoding branch, and outputs the logarithmic variance in three coordinate directions for each bony anchor point. After exponential transformation, the non-negative variance and the prior uncertainty of the bony anchor point are formed.

[0097] The training data consisted of paired data from the same subject in multiple positions. Each sample included planned back surface data, current back surface data, planned bone-surface coupling map, current position validation ultrasound data, and corresponding annotations for the same bony anchor points. The training, validation, and test sets were divided according to the subjects at a ratio of 70%, 15%, and 15%, respectively. Ultrasound images were normalized to 512×512 pixels, and data augmentation included speckle noise reduction, local acoustic shadowing, random point cloud point loss, and small-amplitude rotation and translation.

[0098] The total loss during model training is expressed as:

[0099]

[0100] in, The distance between the point set of the transformed planned back surface and the current back surface. To predict the positional error between the bony anchor points and the labeled bony anchor points, Penalty for changes in node distances within bony units. Penalty for changes in the order and connection relationships of skeletal nodes. This is a Gaussian negative log-likelihood loss based on the prediction variance. The loss weights can be set sequentially as follows: , , , and Training was performed using the AdamW optimizer with an initial learning rate of... The weight decays to The batch size is eight, the maximum number of training rounds is two hundred, and training stops when the aggregate loss of the verification set does not decrease for twenty consecutive rounds.

[0101] To ensure interpretability of the uncertainty, after training, the validation set is divided into intervals based on the predicted standard deviation, and the root mean square value of the bony anchor point error is calculated for each interval. When the predicted standard deviation deviates from the corresponding error, temperature scaling is used to calibrate the standard deviation; the temperature parameter is determined only by the validation set. Therefore, the output bony anchor point prior uncertainty corresponds to the standard deviation of the spatial location, rather than an arbitrary model score, and can be directly used to validate location selection and bony boundary expansion.

[0102] In one embodiment, step S3 specifically includes: generating a set of candidate verification locations on the current back surface data. Each candidate verification location includes the probe position and probe orientation, and the verification location is determined according to the following formula. :

[0103]

[0104] in, For the number of bony anchor points, For the first The prior uncertainty of a bony anchor point To be at the candidate validation position The expected posterior uncertainty after verification is completed. For the expected visibility, The cost of the probe's availability, For bony acoustic shadowing amount, , and The weights are non-negative.

[0105] Specifically, such as Figure 3 As shown, the set of candidate verification locations Generating within the target lumbar spine region on the current back surface (810°). Candidate surface points are first selected at 5-10 mm intervals. Then, the initial probe orientation is constructed based on the surface normal, and lateral and longitudinal tilt angles are discretely set near the initial orientation. Candidates falling outside the body surface area, spatially interfering with the operating table or existing instruments, or exceeding the reachable angle of the ultrasound probe (210°) are discarded. Each retained candidate contains a three-dimensional position and three-dimensional orientation, which can be directly converted into probe adjustment commands.

[0106] Prior uncertainty and expected posterior uncertainty All values ​​were calculated using the square root of the location covariance trace and normalized within the same case. The validation prediction network was supervised by actual validation scans, taking as input candidate probe poses and local sub-images of the bone surface coupling map, and outputting the predicted localization variance of each bony anchor point after scanning. Predicted visibility... A value between zero and one represents the probability that the bony anchor point enters the sound beam and forms a recognizable hyperechoic boundary.

[0107] Probe achievable cost The amount of bony acoustic shadowing is formed by weighting translation, rotation, and surface normal deviation. It is formed by the intersection length of the sound beam propagation path and the identified bone surface, and both are normalized to zero to one. Assign higher values ​​to bony anchor points near the channel. and Adjust the probe movement cost and bone occlusion cost separately; all terms in the formula are dimensionless.

[0108] Verification module 130 selects the module with the largest objective function. If the highest objective function value is not greater than the minimum information gain, or if the expected visibility of all candidates is lower than the threshold, the verification module 130 outputs a rescan command. Otherwise, the display device 600 displays the displacement and attitude adjustment amount from the current probe pose to the verification position; after the probe is in position, the system receives the verification ultrasound data and the current probe pose.

[0109] In one embodiment, step S4 specifically includes: transforming the verification ultrasound data to the current spatial coordinate system according to the current probe pose; extracting the current bony anchor point and the current bony anchor point confidence from the transformed verification ultrasound data; performing graph matching between the current bony anchor point and the planned bony anchor point in the bone surface coupling graph according to the anatomical hierarchy and adjacency relationship of the bony nodes to obtain bony topological consistency; forming anchor point matching residuals based on the matching position difference, matching direction difference and the current bony anchor point confidence; and using the anchor point matching residuals to update the pose of the corresponding bony unit and the posterior uncertainty of the bony anchor point.

[0110] Specifically, the current probe pose is synchronized with each frame of verification ultrasound data according to the timestamp. The coordinate transformation module 180 uses calibration parameters between the coordinate systems of the ultrasound image, probe, tracking, and surface acquisition device to map the verification ultrasound pixels to the current spatial coordinate system. Frames with timestamp differences exceeding the synchronization tolerance are not included in anchor point extraction.

[0111] The bony anchor point extraction network employs a four-level convolutional encoder and a four-level decoder, with same-scale skip connections between the encoder and decoder. The input is a 512×512 pixel validation ultrasound image after intensity normalization. The output includes class probability maps and pixel-level confidence maps for the spinous process, lamina, articular processes, continuous sacral surfaces, and intervertebral spaces. Training annotations include various hyperechoic bone surface segments, acoustic shadowing initiation boundaries, and invisible regions. The loss function consists of class cross-entropy, boundary distance loss, and confidence calibration loss. The current bony anchor point is extracted from the connected regions of the class probability map, and the confidence of the current bony anchor point is formed by the mean of the confidence map within the corresponding connected regions.

[0112] Image matching first uses the continuous surface of the sacrum or the spinous process with the highest confidence as the reference node, and arranges candidate nodes in the head-to-tail direction; then it checks whether the edges between adjacent spinous processes, between spinous processes and the same layer of lamina, and between adjacent laminas are coupled to the bone surface. Figure 1 The skeletal topological consistency can be taken as the ratio of the number of edges that satisfy the expected category, order, and adjacency relationship to the total number of edges participating in the matching. This value is used for subsequent validity judgment and does not directly replace the spatial residuals.

[0113] For the first match For each bony anchor point, the matching position difference is calculated as the Euclidean distance between the current coordinates and the transformed planned coordinates, and the matching direction difference is calculated as the angle between the current bone surface normal and the predicted bone surface normal. These two are weighted to form the anchor point matching residual, with low-confidence observations having a smaller weight. The correction module 140 uses this residual as the observation innovation quantity, updates the rotation and translation through weighted rigid body registration, and then determines the observation covariance using image resolution, extraction confidence, and calibration error. The posterior uncertainty of the bony anchor points is obtained through extended Kalman update or Gaussian information fusion. If the matching cost is ambiguous, a rescan instruction is generated.

[0114] In one embodiment, step S6 specifically includes: generating an expanded bony boundary along the normal of the bony boundary according to the expansion distance positively correlated with the posterior uncertainty of the bony anchor point; detecting whether the migration puncture channel passes through the interlaminar window corresponding to the target intervertebral space and whether it intersects with the expanded bony boundary; when the migration puncture channel passes through the interlaminar window, does not intersect with the expanded bony boundary, and the bony topology consistency, anchor point matching residual, and channel position uncertainty all meet the corresponding thresholds, the migration puncture channel is determined to be valid.

[0115] Specifically, such as Figure 2 and Figure 5 As shown, the migration puncture channel 850 is a three-dimensional channel defined by the cross-section of the needle insertion end, the cross-section of the target end, and the central axis connecting the two. The migration module 150 first uses a continuous soft tissue deformation field to migrate the surface end of the planned channel, and then uses an updated bone unit segmental transformation to migrate the channel end near the bone surface and the target region 860, connecting the two ends through continuous interpolation. The channel position uncertainty is obtained from the soft tissue deformation prediction covariance, the bone unit pose covariance, the coordinate calibration covariance, and the propagation of the current probe pose error.

[0116] The bony boundary 820 is composed of the surfaces of the lamina, spinous process, and articular processes adjacent to the planned puncture channel. For the first... The posterior covariance of the bony anchor points is projected along the cross-sectional direction of the channel from each sampling point to obtain the positional standard deviation. The expansion distance can be expressed as:

[0117]

[0118] in, The basic distance is formed by the instrument radius, segmentation error, and reserved gap. These are non-negative confidence coefficients. When using an approximate Gaussian error, It can be set to two, so that the expansion distance covers about 95% of the unidirectional position fluctuations. and Formulas with the same length dimension maintain dimensional consistency.

[0119] The judgment module 160 establishes an expanded bony boundary along the normal of the bony boundary and performs three geometric checks on the migration puncture channel. The first check checks whether the central axis or channel section enters and passes through the interlaminar window 830 of the target intervertebral space; the second check checks whether the channel body intersects with the expanded bony boundary; and the third check checks whether the target end of the channel is located in the predetermined target area. The geometric check results are then compared with bony topology consistency, anchor point matching residuals, and channel position uncertainty. Only when all conditions meet their respective thresholds is a valid judgment output.

[0120] Each threshold is determined by an independent validation set and fixed during the deployment phase. The bony topology consistency threshold is used to exclude erroneous intervertebral levels, the anchor point matching residual threshold is taken as the high quantile of the correctly matched samples, and the channel position uncertainty threshold is set according to the allowable lateral channel error. If any of these conditions are not met, the judgment module 160 outputs an invalid status, and the output module 170 generates a rescan instruction.

[0121] In one embodiment, step S7 specifically includes: determining the overlapping area between the migration puncture channel and the current back surface data as the needle insertion area; determining the angle range based on the central axis of the migration puncture channel and the channel position uncertainty; determining the expected depth range based on the path length from the needle insertion area to the target intervertebral space and the soft tissue thickness compensation amount; continuously comparing subsequent back surface data with the current back surface data; canceling the needle insertion area, angle range, and expected depth range when the surface displacement or surface curvature change exceeds the corresponding threshold, and generating a re-scan instruction.

[0122] Specifically, such as Figure 5 As shown, the needle insertion area 840 is located at the intersection of the migration puncture channel 850 and the back surface 810. Since the migration puncture channel has a non-zero cross-sectional area, the overlap result is a region on the body surface rather than a single pixel. The output module 170 transforms the overlap region into the surface view of the display device 600 and preserves the three-dimensional boundary of the overlap region in the current spatial coordinate system for subsequent surface change monitoring.

[0123] The angle range is based on the angle between the central axis of the migration puncture channel and the local normal and cephalo-coccygeal direction of the current back surface. Output module 170 projects the channel position uncertainty onto the target end cross-section and calculates the allowable axis deflection without intersecting the expanded bony boundary, thus obtaining the upper and lower limits of the angle in the sagittal and transverse directions, respectively. The predicted depth range is based on the path length of the central axis from the needle insertion area to the target area 860 within the interlaminar window 830, plus soft tissue thickness compensation and position uncertainty in the path direction. The soft tissue thickness compensation only adjusts the length of the soft tissue segment and does not change the position of the bony target.

[0124] After the guidance result is output, the surface acquisition device 300 continues to acquire subsequent back surface data. Each frame of subsequent back surface data undergoes rigid registration to eliminate overall drift of the acquisition device, and is then compared with the current back surface data used as a reference. Surface displacement is taken as the high quantile of the displacement of the corresponding point within an 840-degree neighborhood of the needle insertion area, and surface curvature change is taken as the high quantile of the principal curvature difference within the same region. When a specified number of frames are reached consecutively and any value exceeds the corresponding threshold, it is determined that the patient's position or local soft tissue state has deviated from the state at the time of verification completion.

[0125] Upon receiving an out-of-limit status, the output module 170 cancels the needle insertion area, angle range, and expected depth interval, and generates a rescan command. The puncture needle is only used for positioning and guidance information; the system does not control needle advancement or drug injection.

[0126] Specific application examples:

[0127] Example 1: This example uses the third to fourth lumbar intervertebral space in a lateral decubitus position. The ultrasound device 200 acquires the planned ultrasound sequence, the pose acquisition device 400 simultaneously acquires the planned probe pose, and the surface acquisition device 300 acquires the planned back surface data. The mapping module 110 identifies adjacent spinous processes and lamina using the continuous surface of the sacrum as the hierarchical starting point, establishing a bone-surface coupling map and a planned puncture channel passing through the target interlaminar window 830.

[0128] Planned contact pressure is The current contact pressure is The pre-defined monotonic mapping relationship gives , The soft tissue thickness compensation amount is:

[0129]

[0130] The mapping module 110 reduces the distance between bony nodes and surface nodes in the bone-surface coupling map. The corrected edge features and current back surface data are then input into the deformation decoupling module 120. The deformation decoupling module 120 outputs the continuous soft tissue deformation field, the segmental transformation of each lumbar vertebral bony unit, and the prior uncertainty of the bony anchor point. The prior uncertainty of the lower lamina anchor point of the third lumbar vertebra is... The prior uncertainty of the anchor point of the superior lamina of the fourth lumbar vertebra is The uncertainty of both is higher than that of the adjacent spur anchor points.

[0131] The verification module 130 generates a set of candidate verification locations on the current back surface 810 at eight-millimeter intervals and calculates the expected uncertainty reduction, expected visibility, probe reachability cost, and bony acoustic shadowing for each candidate. The verification location with the highest objective function is located on the mid-lateral side of the third to fourth lumbar intervertebral space. At this point, the long axis of the probe is deflected six degrees relative to the head-to-tail direction. The display device 600 outputs the corresponding probe adjustment command. After the operator completes the adjustment, the system receives twenty frames of verification ultrasound data and the corresponding current probe position.

[0132] The correction module 140 extracts the current bony anchor points of the inferior lamina of the third lumbar vertebra and the superior lamina of the fourth lumbar vertebra. The bony topology consistency obtained by the image matching is... The weighted anchor point matching residual is After pose and covariance updates, the posterior uncertainty of the two anchor points was reduced to [values ​​to be filled in]. and The migration module 150 uses the updated segmental transformation migration planning to obtain the migration puncture channel 850, whose target cross-section channel position uncertainty is: .

[0133] base distance Set as Confidence coefficient Set to two. For the posterior standard deviation of... The bony boundary sampling points have an expansion distance of:

[0134]

[0135] Geometric analysis showed that the migration puncture channel 850 passed through the interlaminar window 830 and did not intersect with the expanded bony boundary. The minimum distance between the migration puncture channel and the expanded bony boundary was [missing information]. If the bony topology consistency is higher than the set lower limit, and the anchor point matching residual and channel position uncertainty are lower than the corresponding upper limit, the judgment module 160 determines that the migration puncture channel is valid.

[0136] The output module 170 generates an insertion area 840 of approximately 8 mm by 6 mm, generating an angle range of 8 to 13 degrees in the sagittal direction and -2 to 3 degrees in the transverse direction relative to the local normal of the back surface, and generating a predicted depth range of 47 mm to 54 mm. During the monitoring phase after output, the surface displacement of the neighborhood of the insertion area is... lower than The displacement threshold and the surface curvature change are both below the corresponding threshold, so the output module 170 maintains the positioning guidance information.

[0137] Example 2: This example uses the same data acquisition and mapping process as Example 1. After verification and output of positioning guidance information, the patient's back moves due to breathing and posture adjustments. The surface acquisition device 300 continuously acquires subsequent back surface data. After the processor 100 eliminates the overall drift of the acquisition device, it detects five consecutive frames of surface displacement in the 840-neighborhood of the needle insertion area. , , , and All exceeded The displacement threshold. Output module 170 cancels the needle insertion area, angle range, and expected depth interval, and generates a rescan instruction.

[0138] Under another abnormal condition, the current contact pressure reaches Exceeding the monotonic mapping relation to Calibration range. The system does not extrapolate the pressure, does not use the soft tissue thickness compensation amount generated by the pressure, and directly generates a rescan command. After the current contact pressure enters the calibration range and remains stable, the current back surface data and verification ultrasound data are reacquired, and then deformation decoupling, active verification, image matching, channel migration, and validity judgment are performed in sequence.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based method for guiding spinal anesthesia puncture localization, characterized in that, include: S1. Receive the planned ultrasound sequence, the planned probe pose, and the planned back surface data, construct a bone surface coupling map and plan the puncture channel. The bone surface coupling map represents the spatial relationship between bony anchor points and the mapping relationship between bony anchor points and the back surface. S2. Receive the current back surface data, input it and the bone surface coupling map into the deformation decoupling model to obtain the soft tissue continuous deformation field, the segmental transformation of the bony unit and the prior uncertainty of the bony anchor point; S3. Determine the verification position based on the prior uncertainty of the bony anchor point and generate a probe adjustment command, and receive the verification ultrasound data of the verification position and the current probe pose; S4. Extract the current bony anchor point from the verification ultrasound data, match it with the bone surface coupling map, and update the segmental transformation of the bony unit and the posterior uncertainty of the bony anchor point; S5. Based on the continuous deformation field of the soft tissue and the updated segmental transformation of the bony unit, the planned puncture channel is migrated to obtain the migrated puncture channel and the channel position uncertainty. S6. Determine the effectiveness of the migration puncture channel based on the bone surface coupling diagram, the posterior uncertainty of the bone anchor point, and the channel position uncertainty; S7. When the migration puncture channel is valid, generate the needle insertion area, angle range and expected depth range; when the migration puncture channel is invalid, generate a re-scan command.

2. The deep learning-based spinal anesthesia puncture localization and guidance method according to claim 1, characterized in that, Step S1 specifically includes: Based on the planned probe pose, the planned ultrasound sequence is transformed to the planned spatial coordinate system, and the planned bony anchor points corresponding to the spinous process, lamina, articular process, sacral continuous surface and intervertebral space are extracted from the transformed planned ultrasound sequence. The planned bony anchor points are used as bony nodes, the corresponding surface projection points in the planned back surface data are used as surface nodes, and the anatomical adjacency relationship between bony nodes and the spatial mapping relationship between bony nodes and surface nodes are used as edges to construct the bone-surface coupling diagram. A planned puncture channel that avoids bony boundaries is generated based on the interlaminar window of the target intervertebral space and the corresponding surface nodes.

3. The deep learning-based spinal anesthesia puncture localization and guidance method according to claim 2, characterized in that, The process of constructing the bone surface coupling map in step S1 also includes: receiving the planned contact pressure and the current contact pressure, querying the pre-calibrated monotonic mapping relationship between the contact pressure and the soft tissue compression amount based on the pressure difference between the two, and obtaining the soft tissue thickness compensation amount. The distance between bony nodes and surface nodes in the bone-surface coupling diagram is corrected using the soft tissue thickness compensation amount, and the corrected distance is input into the deformation decoupling model. When the current contact pressure exceeds the calibration range of the monotonic mapping relationship, the rescan command is generated.

4. The deep learning-based spinal anesthesia puncture localization and guidance method according to claim 1, characterized in that, In step S2, the deformation decoupling model includes a surface feature encoding branch, a bone map encoding branch, a coupling transformation decoding branch, and an uncertainty output branch; The surface feature encoding branch extracts the deformation features between the planned back surface data and the current back surface data. The bony map encoding branch extracts the node and edge features of the bone surface coupling map. Under the constraints of continuous soft tissue deformation, rigidity of each bony unit, and unchanged adjacency relationship of bony nodes, the coupling transformation decoding branch outputs the continuous soft tissue deformation field and the segmental transformation of the bony unit. The uncertainty output branch outputs the prior uncertainty of the bony anchor point.

5. The deep learning-based spinal anesthesia puncture localization and guidance method according to claim 1, characterized in that, Step S3 specifically includes: generating a set of candidate verification locations on the current back surface data. Each candidate verification location includes the probe position and probe orientation, and is arranged according to... Determine the verification location ;in, For the number of bony anchor points, For the first The prior uncertainty of a bony anchor point To be at the candidate validation position The expected posterior uncertainty after verification is completed. For the expected visibility, The cost of the probe's availability, For bony acoustic shadowing amount, , and The weights are non-negative.

6. The deep learning-based spinal anesthesia puncture localization and guidance method according to claim 1, characterized in that, Step S4 specifically includes: transforming the verification ultrasound data to the current spatial coordinate system according to the current probe pose, and extracting the current bony anchor point and the confidence level of the current bony anchor point from the transformed verification ultrasound data; Based on the anatomical hierarchy and adjacency of the bony nodes, the current bony anchor point is graph matched with the planned bony anchor point in the bone surface coupling graph to obtain bony topology consistency. Anchor point matching residuals are generated based on the matching position difference, matching direction difference, and the current bony anchor point confidence. The pose of the corresponding bony unit and the posterior uncertainty of the bony anchor point are then updated using the anchor point matching residuals.

7. The deep learning-based spinal anesthesia puncture localization and guidance method according to claim 6, characterized in that, Step S6 specifically includes: generating an expanded bony boundary along the normal of the bony boundary according to the expansion distance positively correlated with the posterior uncertainty of the bony anchor point; Detect whether the migration puncture channel passes through the interlaminar window corresponding to the target intervertebral space and whether it intersects with the expanded bony boundary; The migration puncture channel is deemed valid when it passes through the interlaminar window, does not intersect the expanded bony boundary, and the bony topology consistency, the anchor point matching residual, and the channel position uncertainty all meet the corresponding thresholds.

8. The deep learning-based spinal anesthesia puncture localization and guidance method according to claim 3, characterized in that, Step S7 specifically includes: determining the overlapping area of ​​the migration puncture channel and the current back surface data as the needle insertion area; determining the angle range based on the central axis of the migration puncture channel and the channel position uncertainty; and determining the expected depth range based on the path length from the needle insertion area to the target intervertebral space and the soft tissue thickness compensation amount. The system continuously compares subsequent back surface data with the current back surface data. When the surface displacement or surface curvature change exceeds the corresponding threshold, the needle insertion area, the angle range, and the expected depth range are cancelled, and the rescan instruction is generated.

9. A deep learning-based spinal anesthesia puncture positioning and guidance system, used to implement the deep learning-based spinal anesthesia puncture positioning and guidance method according to any one of claims 1 to 8, characterized in that, The system includes: The mapping module constructs a bone-surface coupling map and a planned puncture channel based on the planned ultrasound sequence, planned probe pose, and planned back surface data. The bone-surface coupling map represents the spatial relationship between bony anchor points and the mapping relationship between bony anchor points and the back surface. The deformation decoupling module generates a continuous soft tissue deformation field, segmental transformation of bony units, and prior uncertainty of bony anchor points based on the current back surface data and the bone surface coupling diagram. The verification module determines the verification position based on the prior uncertainty of the bony anchor point and generates a probe adjustment command, and receives the verification ultrasound data of the verification position and the current probe pose. The correction module extracts the current bony anchor point from the verification ultrasound data, matches it with the bone surface coupling map, and updates the segmental transformation of the bony unit and the posterior uncertainty of the bony anchor point. The migration module migrates the planned puncture channel based on the continuous deformation field of the soft tissue and the updated segmental transformation of the bony unit, generating the migration puncture channel and the channel position uncertainty. The judgment module determines the validity of the migration puncture channel based on the bone surface coupling diagram, the posterior uncertainty of the bone anchor point, and the channel position uncertainty. The output module generates the needle insertion area, angle range, and expected depth range when the migration puncture channel is valid, and generates a re-scan command when the migration puncture channel is invalid.

10. The deep learning-based spinal anesthesia puncture positioning and guidance system according to claim 9, characterized in that, It also includes an ultrasonic device, a surface acquisition device, a pose acquisition device, a pressure acquisition device, and a coordinate transformation module; The pose acquisition device includes an optical tracking marker or electromagnetic tracking sensor disposed on the ultrasound probe, and the pressure acquisition device is disposed between the grip part of the ultrasound probe and the transducer; the coordinate transformation module, based on the calibration parameters from the ultrasound image coordinate system, the probe coordinate system, and the surface acquisition device coordinate system to the tracking coordinate system, transforms the planned ultrasound sequence, the verification ultrasound data, the planned back surface data, and the current back surface data to the same spatial coordinate system, and corrects the distance between the bony anchor point and the back surface according to the contact pressure output by the pressure acquisition device.

Images