A method and system for locating anesthesia block targets based on multi-source data

Abstract

The present invention relates to the field of computer-assisted surgery technology, and specifically to a method and system for locating anesthesia block targets based on multi-source data. The method comprises: acquiring an MRI image of the target area and constructing a preoperative three-dimensional structural model; acquiring an intraoperative US image and performing multimodal registration with the MRI image; extracting the periodic displacement field of the anatomical structure based on a continuous US image frame sequence, constructing a dynamic displacement model, and performing temporal transformation on the registered image to generate a compensated multimodal image; identifying and segmenting key anatomical structures through a deep neural network to obtain structural mask information; screening multiple candidate injection points within the mask range and selecting the optimal block target based on an objective function score; superimposing the block target and the segmented structure results, matching them to the preoperative three-dimensional model, and constructing a three-dimensional reference model to achieve visual guidance during surgery.

Description

Technical Field

[0001] The present invention relates to the technical field of computer-assisted surgery, and in particular to a method and system for positioning an anesthesia block target based on multi-source data. Background Art

[0002] During clinical anesthesia, nerve block techniques are widely used in surgical procedures, pain management, and other fields. To improve surgical safety and effectiveness, doctors often use ultrasound guidance to perform nerve blocks. However, traditional ultrasound guidance methods suffer from limited image resolution, poor soft tissue contrast, and reliance on operator experience for anatomical structure identification. These issues can easily lead to complications such as inaccurate puncture, inadequate drug injection, and damage to nerves and blood vessels.

[0003] Although complex technologies such as "ultrasound guidance + neurostimulator" combined positioning have been introduced clinically in an effort to enhance positioning reliability by introducing electrophysiological feedback based on visual positioning, many limitations still exist. On the one hand, neurostimulators have individual differences in nerve sensitivity among different patients, and the feedback signal is unstable. On the other hand, ultrasound images themselves are still limited by operator experience and imaging angles, and structural recognition relies on subjective judgment. In addition, there is a lack of effective fusion of preoperative and intraoperative image data, making it difficult to provide stable and consistent spatial reference information. Dynamic structures such as the diaphragm and pleura shift under respiratory drive, making accurate target prediction even more difficult.

[0004] Therefore, a method and system for locating anesthesia block targets based on multi-source data were proposed. Summary of the Invention

[0005] This invention provides a multi-source data-based method and system for locating anesthesia block targets. By combining preoperative MRI images with intraoperative US images, and employing techniques such as image registration, dynamic modeling, deep learning segmentation, and intelligent assessment, this method accurately identifies key structures such as nerves, blood vessels, and fascia, and automatically selects the optimal block target location. This method fuses the recognition results with intraoperative images and matches them to a preoperative 3D model, constructing a 3D reference model. This provides an intuitive and accurate positioning reference for clinical anesthesia, improving the safety and effectiveness of block procedures.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A method for locating anesthesia block targets based on multi-source data, comprising:

[0008] Obtain MRI images of the target area, perform image segmentation and 3D modeling on the MRI images, extract boundary information of the anatomical structure, and construct a preoperative structural model;

[0009] Obtain a US image of the target area as an intraoperative guidance image. Based on the anatomical structure information between the MRI image and the US image, the mutual information maximization is used to spatially align the MRI image and the US image to generate a registered multimodal image.

[0010] Based on the displacement field of the anatomical structure's periodic motion identified in the intraoperative continuous US image frame sequence, a dynamic displacement model of the target structure is constructed, and the multimodal registration images are temporally transformed to generate compensated multimodal images.

[0011] Input the compensated multimodal image into the trained deep neural network model to identify and segment the anatomical structure and output the structural mask information;

[0012] Obtain multiple candidate injection points within the segmentation mask range, design an objective function to score each point, and select one or more points with the largest score as the blocking target;

[0013] The block target and the structural segmentation results are superimposed and displayed on the intraoperative US image, and the superimposed results are 3D matched to the preoperative structural model to generate a three-dimensional reference model to assist clinical anesthesia operations in achieving target positioning.

[0014] Furthermore, the steps for constructing the preoperative structural model are:

[0015] applying image enhancement and denoising algorithms to the MRI images to improve soft tissue edge clarity;

[0016] Perform semantic segmentation on MRI images based on a deep neural network model to extract boundary information of anatomical structures;

[0017] The segmentation results are input into the 3D reconstruction algorithm, and a 3D model of the preoperative anatomical structure is constructed through interpolation reconstruction and surface mesh generation.

[0018] Furthermore, the steps of generating the registered multimodal image are:

[0019] The MRI images were used as reference images, and the intraoperative US images were used as moving images;

[0020] Based on the spatial distribution relationship of key anatomical structures in MRI images and US images, the mutual information value between the two images is calculated;

[0021] A non-rigid registration algorithm based on mutual information maximization is used to perform elastic deformation on the US image to align the US image with the MRI image on key anatomical structures and generate a registered multimodal image.

[0022] Furthermore, the steps of generating the compensated multimodal image are:

[0023] Based on the continuously acquired intraoperative US image sequence, the target area is modeled in time series to identify the displacement field of the anatomical structure in multiple frames;

[0024] The dynamic deformation field of key anatomical structures is established by using a temporal convolutional neural network to transform the displacement field;

[0025] The dynamic deformation field is applied to the registered multimodal images, and the static structural model is transformed in a temporal spatial manner to generate compensated multimodal images that change dynamically with breathing.

[0026] Furthermore, the calculation formula of the objective function is:

[0027] ;

[0028] in, Indicates candidate injection points The comprehensive score of Indicates candidate points The minimum Euclidean distance to the target neuron, Indicates candidate points The minimum Euclidean distance to the blood vessel, Indicates candidate points The width of the space between the fascia layers, 、 and is the weighting coefficient.

[0029] Furthermore, the steps of generating a 3D reference model are:

[0030] The structural mask and the blocking target are converted into a pseudo-color layer, and the pseudo-color layer is superimposed with the intraoperative US image using gradient fusion technology to obtain a pseudo-color US image;

[0031] The two-dimensional pseudo-color US image is reconstructed into three dimensions using back-projection reconstruction technology, aligned with the three-dimensional preoperative structural model, and a three-dimensional reference model is generated.

[0032] The present invention also proposes an anesthesia block target positioning system based on multi-source data, comprising:

[0033] MRI image structure extraction module, used to obtain MRI images of the target area, perform image segmentation and 3D modeling on the MRI images, extract boundary information of anatomical structures, and construct a preoperative structural model;

[0034] The image registration module is used to obtain the US image of the target area as the intraoperative guidance image. Based on the anatomical structure information between the MRI image and the US image, the mutual information maximization is used to spatially align the MRI image and the US image to generate a registered multimodal image;

[0035] The image compensation module is used to construct a dynamic displacement model of the target structure based on the displacement field of the anatomical structure's periodic motion identified in the intraoperative continuous US image frame sequence, and to perform temporal transformation on the multimodal registration images to generate compensated multimodal images;

[0036] The mask acquisition module is used to input the compensated multimodal image into the trained deep neural network model to identify and segment the anatomical structure and output the structural mask information;

[0037] The blocking target positioning module is used to obtain multiple candidate injection points within the segmentation mask range, design an objective function to score each point, and select one or more points with the largest score as the blocking target;

[0038] The three-dimensional model reference module is used to superimpose the block target and the structural segmentation results on the intraoperative US image, and perform 3D matching of the superimposed results with the preoperative structural model to generate a three-dimensional reference model to assist clinical anesthesia operations in achieving target positioning.

[0039] The beneficial effects of the present invention are:

[0040] By fusing preoperative MRI images with intraoperative US images, multimodal structural information acquisition and registration of the patient's target area are achieved, overcoming the challenges of incomplete information and insufficient recognition accuracy associated with a single imaging modality. Time series modeling and optical flow estimation are performed on the intraoperative continuous US image sequence to identify periodic displacements of anatomical structures. A dynamic deformation field is established using a temporal convolutional neural network to effectively compensate for structural position shifts caused by respiratory motion, improving the spatiotemporal accuracy of target localization. A trained deep neural network is used to accurately identify and semantically segment key structures (such as nerves, blood vessels, and fascia). Based on this, a multi-factor comprehensive evaluation objective function is designed to automatically evaluate and select the optimal block targets for anatomical safety and anesthetic efficacy. Finally, through image fusion and 3D visualization, the structural segmentation results and the block targets are superimposed on the intraoperative US images and matched back to the preoperative 3D structural model to construct a reference model to assist in intraoperative navigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0042] Figure 1 This is a flow chart of a method for locating anesthesia block targets based on multi-source data provided by the present invention;

[0043] Figure 2 This is a structural diagram of a multi-source data-based anesthesia block target positioning system provided by the present invention. DETAILED DESCRIPTION

[0044] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0045] A method for locating block targets for anesthesia based on multi-source data, such as Figure 1 As shown, including:

[0046] S100: Acquire an MRI image of the target area, perform image segmentation and three-dimensional modeling on the MRI image, extract boundary information of the anatomical structure, and construct a preoperative structural model;

[0047] Furthermore, the steps for constructing the preoperative structural model are:

[0048] applying image enhancement and denoising algorithms to the MRI images to improve soft tissue edge clarity;

[0049] Perform semantic segmentation on MRI images based on a deep neural network model to extract boundary information of anatomical structures;

[0050] The segmentation results are input into the 3D reconstruction algorithm, and a 3D model of the preoperative anatomical structure is constructed through interpolation reconstruction and surface mesh generation.

[0051] Specifically, a 1.5T or 3.0T magnetic resonance imaging system is used to image the target area of ​​the patient and obtain high-resolution MRI images of soft tissue structures including nerves, blood vessels, muscles, fascia, etc. In order to improve the clarity and segmentation accuracy of soft tissue boundaries in the image, the following image preprocessing process is used:

[0052] Perform non-local mean denoising on MRI images to reduce background noise interference;

[0053] Use histogram equalization or adaptive contrast enhancement algorithms (such as CLAHE) to enhance grayscale contrast and highlight soft tissue boundaries;

[0054] Perform normalization processing (such as z-score normalization) on the enhanced image to adapt to the subsequent neural network model input requirements.

[0055] The preprocessed MRI images are fed into a pretrained deep learning model for anatomical segmentation. Preferably, the neural network is a multi-channel semantic segmentation network based on a U-Net architecture, and the training dataset is derived from expert-annotated public or clinical MRI images. The network can perform pixel-level segmentation of the following anatomical structures: peripheral nerve bundles (e.g., brachial plexus), vascular structures (arteries and veins), fascial layers (muscular septa, fascial bursae), bony landmarks, and muscle tissue. The multi-channel probability map output by the network is thresholded to generate a segmentation mask representing the boundary contours of each structure.

[0056] The segmentation mask is input into the 3D reconstruction module. The specific modeling process is as follows:

[0057] Interpolation reconstruction is performed between slices for the segmentation layers of each type of structure to improve the spatial continuity in the Z-axis direction;

[0058] Apply the Marching Cubes algorithm or Poisson surface reconstruction algorithm to extract isosurfaces and generate a three-dimensional mesh model;

[0059] Perform smoothing and topology optimization on 3D meshes to improve model continuity and visualization quality.

[0060] By introducing MRI images for preoperative high-precision three-dimensional modeling and using deep neural networks to achieve automatic segmentation and boundary extraction of key anatomical structures, the recognition accuracy of soft tissues such as nerves, blood vessels, and fascia is effectively improved; combining interpolation reconstruction and three-dimensional surface mesh generation technology, a spatially continuous and structurally clear preoperative structural model is constructed, providing an accurate anatomical basis for subsequent multimodal image registration and target positioning.

[0061] S200: Acquire a US image of the target area as an intraoperative guidance image, and based on the anatomical structure information between the MRI image and the US image, spatially align the MRI image and the US image by maximizing mutual information to generate a registered multimodal image;

[0062] Furthermore, the steps of generating the registered multimodal image are:

[0063] The MRI images were used as reference images, and the intraoperative US images were used as moving images;

[0064] Based on the spatial distribution relationship of key anatomical structures in MRI images and US images, the mutual information value between the two images is calculated;

[0065] A non-rigid registration algorithm based on mutual information maximization is used to perform elastic deformation on the US image to align the US image with the MRI image on key anatomical structures and generate a registered multimodal image.

[0066] Specifically, MRI images and intraoperative US images were preprocessed separately to improve the stability and accuracy of registration. MRI images were edge-enhanced and smoothed to highlight structures such as nerves, blood vessels, and fascia. US images were subjected to Speckle denoising and edge enhancement to reduce speckle noise. Both images were uniformly resampled and resized to ensure consistency of the registration input.

[0067] Based on the intraoperative US probe position and spatial markers in the MRI image (such as bony structures and fascial interfaces), the initial alignment area is determined, and key anatomical structures in the target area (anatomical structures closely related to the block target, such as the nerve plexus in neuroanesthesia or the fascia for intrafascial injection) are extracted to narrow the scope of the registration calculation.

[0068] Mutual information is selected as the registration evaluation index, and a joint histogram is constructed to calculate the mutual information value to measure the statistical correlation between the US image and the MRI image after transformation. The mutual information maximization is set as the objective function, and the spatial transformation parameters of the US image are globally optimized.

[0069] During the registration process, B-spline free deformation is used as a non-rigid transformation model. The US image is defined on a control grid, and the positions of the control points are optimized to achieve local elastic deformation. A gradient descent optimization algorithm is used to iteratively update the control point parameters to maximize the mutual information value. To avoid overfitting, a smoothing regularization term can be added to maintain the continuity and physiological rationality of the deformation field.

[0070] The final transformed US image is superimposed and fused with the MRI image to form a registered multimodal image with consistent structural alignment. The registered image will be used in subsequent target positioning and structure recognition steps.

[0071] Multimodal image registration methods can fully utilize the high soft tissue contrast advantage of MRI and the real-time dynamic characteristics of US. Through a mutual information-driven elastic alignment strategy, they can effectively solve the misalignment problem between different modal images caused by tissue deformation or probe posture differences, improve the accuracy of anatomical structure matching, and provide more accurate image guidance support for subsequent nerve block operations.

[0072] S300: Based on the displacement field of the anatomical structure's periodic motion identified in the intraoperative continuous US image frame sequence, a dynamic displacement model of the target structure is constructed, and the multimodal registration image is temporally transformed to generate a compensated multimodal image;

[0073] Furthermore, the steps of generating the compensated multimodal image are:

[0074] Based on the continuously acquired intraoperative US image sequence, the target area is modeled in time series to identify the displacement field of key anatomical structures in multiple frames;

[0075] The dynamic deformation field of key anatomical structures is established by using a temporal convolutional neural network to transform the displacement field;

[0076] The dynamic deformation field is applied to the registered multimodal images, and the static structural model is transformed in a temporal spatial manner to generate compensated multimodal images that change dynamically with breathing.

[0077] Specifically, during anesthesia surgery, an ultrasound imaging device is used to continuously acquire an intraoperative US image sequence of the target area, which is recorded as ,in Indicates time The collected Frame US image. Lucas optical flow estimation is performed on the key anatomical structures in the image sequence that represent the movement driven by breathing, and their displacement fields in consecutive frames are extracted. , for two consecutive frames of US images and , and process it through the Lucas optical flow algorithm to obtain the Frame displacement field .

[0078] The displacement field of key anatomical structures is modeled and predicted, and a temporal convolutional neural network is used to learn the deformation law of the target structure in the time dimension. The network input is continuous Displacement field sequence of target structure in the frame , network output for the future Predicted deformation field at the moment , that is, the predicted deformation of the target structure at a future time point; the network training goal is to minimize the L2 loss between the predicted deformation field and the true displacement field: it can be expressed as:

[0079] ;

[0080] in, represents the loss value, and Represents the spatial dimension of the displacement field.

[0081] The predicted dynamic deformation field Applied to the static multimodal image obtained previously, for each pixel Perform spatial transformation ,in is the transformed coordinate. The registered image is deformed by resampling to generate the Compensated multimodal images under .

[0082] By introducing temporal dynamic modeling of anatomical structures and temporal spatial compensation of registered images, the impact of tissue position changes caused by intraoperative respiratory motion on the accuracy of block target identification is effectively addressed. By constructing the displacement field of key structures using continuous ultrasound image sequences and extracting structural motion patterns with the help of a temporal convolutional neural network, a dynamic deformation field is generated. This then performs temporal transformations on static registered images, reflecting the actual position of anatomical structures at different respiratory phases in real time. This improves the timeliness and spatial consistency of image registration, contributing to the increased accuracy and safety of block target positioning.

[0083] S400: Inputting the compensated multimodal image into the trained deep neural network model to identify and segment the anatomical structure and output structure mask information;

[0084] Specifically, compensated multimodal images, that is, images that combine MRI static structure and US dynamic motion compensation information, are used as model input. The input image is a single frame or a multi-frame sequence, and the resolution is consistent with the original US image. The deep neural network model can be a medical image segmentation network based on an encoder-decoder structure, and nnU-Net is preferably used in this embodiment. The model is trained on preoperative or intraoperative datasets, using labeled MRI and US images and their corresponding structure manual segmentation labels as supervisory signals. The output is a structural mask image of multiple channels, each channel represents an anatomical structure (such as nerves, blood vessels, fascia, pleura, etc.), and the value of each pixel represents the probability or classification label of the corresponding position belonging to a certain structure.

[0085] S500: obtaining multiple candidate injection points within the segmentation mask range, designing an objective function to score each point, and selecting one or more points with the largest score as blocking targets;

[0086] Furthermore, the calculation formula of the objective function is:

[0087] ;

[0088] in, Indicates candidate injection points The comprehensive score of Indicates candidate points The minimum Euclidean distance to the target neuron, Indicates candidate points The minimum Euclidean distance to the blood vessel, Indicates candidate points The width of the space between the fascia layers, 、 and is a weighting coefficient, which can be given according to the guidance of an experienced physician or obtained by fitting through annotated historical data.

[0089] By acquiring multiple candidate injection points within the segmentation mask and scoring and screening them based on an objective function constructed from comprehensive anatomical features, the scientific and safe selection of target sites for blockade is effectively improved. This method quantitatively assesses key factors such as the relative position of each candidate point to the nerve, avoidance of important structures (such as blood vessels and pleura), and fascial space width, ensuring that the final injection point is selected while ensuring anesthesia and minimizing puncture risk.

[0090] S600: The block target and the structural segmentation result are superimposed and displayed on the intraoperative US image, and the superimposed result is 3D matched with the preoperative structural model to generate a three-dimensional reference model to assist clinical anesthesia operation in achieving target positioning.

[0091] Furthermore, the steps of generating a 3D reference model are:

[0092] The structural mask and the blocking target are converted into a pseudo-color layer, and the pseudo-color layer is superimposed with the intraoperative US image using gradient fusion technology to obtain a pseudo-color US image;

[0093] The two-dimensional pseudo-color US image is reconstructed into three dimensions using back-projection reconstruction technology, aligned with the three-dimensional preoperative structural model, and a three-dimensional reference model is generated.

[0094] Specifically, different pseudo-color codes are assigned to the blocking targets and each type of identified anatomical structure (e.g., nerves, blood vessels, pleura, fascia, etc.), for example: blocking targets → yellow; blood vessel structures → red; fascia structures → blue; image processing libraries (such as OpenCV, PyTorch Tensor API) are used to convert the structure mask binary map (or probability map) into a semi-transparent color layer, and the transparency of each structure layer is The α-value can be controlled within the range of 0.3–0.6, making it easier to observe US image details. The pseudo-color structural layer is aligned with the intraoperative US image in pixel space (based on the aforementioned registration and dynamic compensation), and weighted superposition is used to achieve gradient fusion to obtain the pseudo-color US image. The implementation of weighted superposition can be described as follows:

[0095] ;

[0096] in, represents the fused image, represents the intraoperative US image, Represents the pseudo-color structure layer, Indicates the transparency of the layer.

[0097] By using the spatial tracking information of the B-ultrasound probe, the acquisition posture of each pseudo-color US image is recorded to obtain the position and direction of each image in the MRI reference coordinate system. The posture is defined as: ,in is the rotation matrix, is the translation vector. Using the position and posture parameters of each frame image, the pixel intensity in the two-dimensional pseudo-color image is back-projected to the three-dimensional space coordinates of the preoperative structural model. The back-projection method based on voxel space is used to The image contents of each frame are projected and superimposed. The basic principle is as follows: for each pseudo-color ultrasound image frame, its pixels Corresponding spatial point coordinates It can be obtained by back projection through the following formula:

[0098] ;

[0099] in, and represents the pixel pitch, Indicates the center of the image.

[0100] The above three-dimensional back projection results The 3D anatomical model constructed from the preoperative MRI is spatially registered (e.g., using a B-spline deformation field or the Demons algorithm). The resulting fused volume is a 3D reference model that includes key structures, block targets, and preoperative anatomical reference information.

[0101] Example 2

[0102] This embodiment takes brachial plexus block as the target, and uses the multi-source data-based anesthesia block target positioning system proposed by the present invention to accurately locate the block target around the brachial plexus and guide clinical anesthesia operations. Figure 2 As shown, including:

[0103] MRI image structure extraction module, used to obtain MRI images of the target area, perform image segmentation and 3D modeling on the MRI images, extract boundary information of anatomical structures, and construct a preoperative structural model;

[0104] The image registration module is used to obtain the US image of the target area as the intraoperative guidance image. Based on the anatomical structure information between the MRI image and the US image, the mutual information maximization is used to spatially align the MRI image and the US image to generate a registered multimodal image;

[0105] The image compensation module is used to construct a dynamic displacement model of the target structure based on the displacement field of the anatomical structure's periodic motion identified in the intraoperative continuous US image frame sequence, and to perform temporal transformation on the multimodal registration images to generate compensated multimodal images;

[0106] The mask acquisition module is used to input the compensated multimodal image into the trained deep neural network model to identify and segment the anatomical structure and output the structural mask information;

[0107] The blocking target positioning module is used to obtain multiple candidate injection points within the segmentation mask range, design an objective function to score each point, and select one or more points with the largest score as the blocking target;

[0108] The three-dimensional model reference module is used to superimpose the block target and the structural segmentation results on the intraoperative US image, and perform 3D matching of the superimposed results with the preoperative structural model to generate a three-dimensional reference model to assist clinical anesthesia operations in achieving target positioning.

[0109] The specific implementation process is as follows:

[0110] The patient underwent a high-resolution T1-weighted MRI scan of the shoulder region, acquiring a series of transverse images. A UNet-based deep neural network model was used to perform semantic segmentation on the MRI images, identifying key structures such as the brachial plexus, axillary artery, subclavian vein, scalene muscle, and first rib. Three-dimensional surface reconstruction was then performed using the Marching Cubes algorithm to obtain a preoperative 3D model of the anatomical structure.

[0111] During the anesthesia preparation phase, a high-frequency linear array probe is used to acquire real-time B-ultrasound images of the target area, and an electromagnetic tracker records the image pose. Image registration is performed using structural similarities (such as muscle boundaries and vascular distribution). Maximizing mutual information is used as the registration criterion. By optimizing rigid transformation parameters, the MRI image is aligned with the current US image to generate a registered multimodal image.

[0112] A continuous sequence of US image frames spanning a complete respiratory cycle is acquired, and optical flow estimation is performed on the brachial plexus and surrounding muscle tissue to extract the inter-frame displacement field. This displacement field data is then fed into a temporal convolutional neural network to build a dynamic deformation field model that describes the spatial trajectory of the brachial plexus as it moves with respiration. This deformation field is then applied to the registered images, generating images that dynamically compensate for respiratory motion at each time point, thereby improving the temporal consistency of structural registration.

[0113] The dynamic compensation image is fed into the DeepLabv3+ semantic segmentation network, which outputs a structural mask that includes the precise boundaries of the brachial plexus region. Based on the segmentation results, multiple candidate injection points are extracted. An objective function is constructed to calculate the scores of these candidates, and the point with the highest score is ultimately selected as the target for blockade.

[0114] The segmentation results and the location of the block target were superimposed on each intraoperative US image frame and presented as a pseudo-color layer. Using electromagnetic tracking to register the pose information and a back-projection reconstruction algorithm, the pseudo-color US frame sequence was mapped into 3D space. The voxels of the intraoperative pseudo-color image were obtained and registered to the preoperative model to generate a 3D reference model.

[0115] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for locating anesthesia block targets based on multi-source data, characterized in that: include: Obtain MRI images of the target area, perform image segmentation and 3D modeling on the MRI images, extract boundary information of the anatomical structure, and construct a preoperative structural model; Obtain a US image of the target area as an intraoperative guidance image. Based on the anatomical structure information between the MRI image and the US image, the mutual information maximization is used to spatially align the MRI image and the US image to generate a registered multimodal image. Based on the displacement field of the anatomical structure's periodic motion identified in the continuous intraoperative US image sequence, a dynamic displacement model of the target structure is constructed, and the multimodal registration images are temporally transformed to generate compensated multimodal images. Input the compensated multimodal image into the trained deep neural network model to identify and segment the anatomical structure and output the structural mask information; Obtain multiple candidate injection points within the segmentation mask range, design an objective function to score each point, and select one or more points with the largest score as the blocking target; The block target and the structural segmentation results are superimposed and displayed on the intraoperative US image, and the superimposed results are 3D matched to the preoperative structural model to generate a three-dimensional reference model to assist clinical anesthesia operations in achieving target positioning.

2. The method for locating anesthesia block targets based on multi-source data according to claim 1, characterized in that: The steps to construct the preoperative structural model are: applying image enhancement and denoising algorithms to the MRI images to improve soft tissue edge clarity; Perform semantic segmentation on MRI images based on a deep neural network model to extract boundary information of anatomical structures; The segmentation results are input into the 3D reconstruction algorithm, and a 3D model of the preoperative anatomical structure is constructed through interpolation reconstruction and surface mesh generation.

3. The method for locating anesthesia block targets based on multi-source data according to claim 1, characterized in that: The steps to generate registered multimodal images are: The MRI images were used as reference images, and the intraoperative US images were used as moving images; Based on the spatial distribution relationship of key anatomical structures in MRI images and US images, the mutual information value between the two images is calculated; A non-rigid registration algorithm based on mutual information maximization is used to perform elastic deformation on the US image to align the US image with the MRI image on key anatomical structures and generate a registered multimodal image.

4. The method for locating anesthesia block targets based on multi-source data according to claim 1, characterized in that: The steps to generate compensated multimodal images are: Based on the continuously acquired intraoperative US image sequence, the target area is modeled in time series to identify the displacement field of the anatomical structure in multiple frames; The dynamic deformation field of key anatomical structures is established by using a temporal convolutional neural network to transform the displacement field; The dynamic deformation field is applied to the registered multimodal images, and the static structural model is transformed in a temporal spatial manner to generate compensated multimodal images that change dynamically with breathing.

5. The method for locating anesthesia block targets based on multi-source data according to claim 1, characterized in that: The calculation formula of the objective function is: ； in, Indicates candidate injection points The comprehensive score of Indicates candidate points The minimum Euclidean distance to the target neuron, Indicates candidate points The minimum Euclidean distance to the blood vessel, Indicates candidate points The width of the space between the fascia layers, 、 and is the weighting coefficient.

6. The method for locating anesthesia block targets based on multi-source data according to claim 1, characterized in that: The steps to generate a 3D reference model are: The structural mask and the blocking target are converted into a pseudo-color layer, and the pseudo-color layer is superimposed with the intraoperative US image using gradient fusion technology to obtain a pseudo-color US image; The two-dimensional pseudo-color US image is reconstructed into three dimensions using back-projection reconstruction technology, aligned with the three-dimensional preoperative structural model, and a three-dimensional reference model is generated.

7. A multi-source data-based anesthesia block target positioning system, characterized in that: include: MRI image structure extraction module, used to obtain MRI images of the target area, perform image segmentation and 3D modeling on the MRI images, extract boundary information of anatomical structures, and construct a preoperative structural model; The image registration module is used to obtain the US image of the target area as the intraoperative guidance image. Based on the anatomical structure information between the MRI image and the US image, the mutual information maximization is used to spatially align the MRI image and the US image to generate a registered multimodal image; An image compensation module is used to construct a dynamic displacement model of the target structure based on the displacement field of the anatomical structure's periodic motion identified in a continuous intraoperative US image sequence, and to perform temporal transformation on the multimodal registration images to generate compensated multimodal images; The mask acquisition module is used to input the compensated multimodal image into the trained deep neural network model to identify and segment the anatomical structure and output the structural mask information; The blocking target positioning module is used to obtain multiple candidate injection points within the segmentation mask range, design an objective function to score each point, and select one or more points with the largest score as the blocking target; The three-dimensional model reference module is used to superimpose the block target and the structural segmentation results on the intraoperative US image, and perform 3D matching of the superimposed results with the preoperative structural model to generate a three-dimensional reference model to assist clinical anesthesia operations in achieving target positioning.

Images