Operation process management method and system

Abstract

The invention provides a surgical process management method and system, and relates to the technical field of surgical process management.The method comprises the steps that DTI, fMRI and 3D-T1WI image data are fused through a non-rigid registration algorithm, a three-dimensional brain network atlas is generated, then an LSTM deep learning model is constructed, a brain surface displacement field monitored in real time in an operation is used as input, and an LSTM deep learning model is constructed; a deep tissue displacement value is used as a label for training to obtain a deep tissue displacement predicted value, and a dynamic relation between brain neural element electrical activity and oxyhemoglobin saturation is analyzed by combining data monitored by high-frequency EEG and near infrared spectrum and utilizing a wavelet time-frequency coherence algorithm to obtain a dynamic influence coefficient; and finally, combining the deep tissue displacement predicted value with the dynamic influence coefficient, calculating a comprehensive evaluation index, and comparing the comprehensive evaluation index with a preset threshold value, so that a surgeon can quickly adjust an operation strategy and timely deal with possible risks.

Description

Technical Field

[0001] The present invention relates to the technical field of surgical process management, and in particular to a surgical process management method and system. Background Art

[0002] In modern neurosurgery, accurate assessment of brain tissue displacement and deformation is a key factor in ensuring surgical safety and effectiveness. Although traditional imaging techniques can provide anatomical information, they cannot reflect the dynamic changes of brain tissue in real time during surgery. In addition, due to the complexity and individual differences of brain tissue, real-time monitoring of brain surface displacement and accurate prediction of deep tissue displacement during surgery remain challenges. Currently, there is a lack of effective methods to integrate multiple imaging technologies and real-time monitoring data. Providing accurate brain tissue displacement prediction and dynamic feedback is an important guarantee to ensure that surgeons can adjust surgical strategies in a timely manner, especially in high-risk surgeries.

[0003] Existing technologies mainly rely on traditional imaging methods to assess the structure and position of brain tissue. However, these methods can only provide static information and cannot reflect the dynamic changes of brain tissue during surgery in real time. Although some emerging technologies, such as intraoperative ultrasound imaging, have been introduced to improve real-time monitoring capabilities, there are still technical limitations in predicting deep tissue displacement. These traditional technologies fail to effectively integrate multiple data sources, resulting in doctors lacking sufficient real-time information during surgery, making it difficult to make timely and accurate decisions.

[0004] Although existing technologies have improved the management of brain surgery to a certain extent, they generally have the following shortcomings: first, the lack of comprehensive assessment of the dynamic relationship between the brain surface and deep tissues often leads to inaccurate predictions of deep tissue displacement; second, traditional methods have a delayed response to emergencies, which invisibly increases the risk of surgery; finally, the fusion processing and intelligent analysis capabilities of multimodal data are limited, and cannot provide individualized, real-time surgical management strategies.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] The purpose of the present invention is to provide a surgical process management method and system to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A surgical process management method, comprising the following steps:

[0009] Step 1: Preoperatively, the patient's DTI, fMRI, and 3D-T1WI imaging data are collected and fused using a non-rigid registration algorithm to generate a three-dimensional brain network map. During surgery, a laser speckle imaging device is integrated into the surgical microscope to capture brain surface displacement data in real time, and the surface displacement field is constructed using an optical flow algorithm.

[0010] Step 2: Build an LSTM deep learning model, using the brain surface displacement field monitored during surgery as the model input and the deep tissue displacement verified by intraoperative ultrasound imaging as the label training model. Input the real-time brain surface displacement data into the trained model, and output the predicted deep tissue displacement value.

[0011] Step 3: Deploy high-frequency EEG and near-infrared spectroscopy equipment in the surgical area. High-frequency EEG is used to detect the electrical activity of brain neurons, and near-infrared spectroscopy is used to monitor changes in local blood oxygen saturation in the brain. Dynamic influence coefficients of changes in neuronal electrical activity and blood oxygen saturation are constructed using a wavelet time-frequency coherence algorithm.

[0012] Step 4: Calculate a comprehensive evaluation index based on the predicted deep tissue displacement of the brain tissue during surgery and the dynamic influence coefficient, and compare the comprehensive evaluation index with the preset threshold to determine whether to adjust the surgery.

[0013] Furthermore, the patient's DTI, fMRI, and 3D-T1WI imaging data were collected before surgery, and the logic for generating a three-dimensional brain network map by fusing them through a non-rigid registration algorithm is as follows:

[0014] DTI is used to obtain white matter fiber bundle information, fMRI is used to identify important functional areas, and 3D-T1WI is used to obtain high-resolution brain anatomical structure and define gray matter, white matter, and cerebrospinal fluid segmentation masks;

[0015] Setting I DTI ,I T1 Represent DTI and T1 image data respectively, and optimize the deformation field through the non-rigid registration algorithm of the B-spline free deformation model Minimize the difference between multimodal images. The objective function is defined as:

[0016]

[0017] Among them, NMI is the normalized mutual information, is the Dice coefficient, M fMRI is the fMRI functional area mask, M T1 is the anatomical segmentation mask of the T1 image; α, β are weight coefficients, and α+β=1;

[0018] By achieving spatial alignment of multimodal data through non-rigid registration and combining fiber tracking, functional area probability, and anatomical segmentation, we ultimately generate a three-dimensional brain network map G = (V, E, W) with brain regions as nodes, white matter connections as edges, and functional area weights as attributes. V is a set of brain regions, with each node corresponding to a brain region; E is the set of edges of white matter fiber connections, representing the anatomical connections between brain regions; and W is the set of edge weights, representing the functional importance of the connections.

[0019] Edge weights of white matter fiber connections The strength of the fiber bundle is expressed based on the formula:

[0020]

[0021] Among them, S ij represents the number of fiber bundles from brain area i to j, FA ij is the anisotropy fraction of the fiber bundle, ranging from [0,1], mean(FA ij ) the mean FA of all fibers connecting brain areas i and j;

[0022] The node weights based on fMRI are based on the formula:

[0023]

[0024] Among them, V i is the i-th brain region, |V i | is the number of voxels in brain region i, i.e., its volume, P func (x) is the fMRI activation probability at position x;

[0025] Weight set: Where a, b are weight coefficients, and a+b=1, is the edge weight of the white matter fiber connection, is the weight of the brain region node.

[0026] Furthermore, during surgery, a laser speckle imaging device was integrated into the surgical microscope to capture brain surface displacement data in real time. The logic for constructing the surface displacement field using the optical flow algorithm was as follows:

[0027] Setting I t (x, y) is the speckle image collected at time t, I t+1 (x, y) is the speckle image acquired at time t+1, where t represents the time frame number and (x, y) represents the coordinates on the image plane. Assuming that the brain surface displacement is a small-scale smooth deformation, the Lucas-Kanade optical flow algorithm is used to calculate the displacement field. The formula is:

[0028]

[0029] in, are the spatial gradients of the speckle intensity I in the x and y directions, is the time gradient of speckle intensity I, v x ,v y are the velocity components of the surface displacement in the x and y directions, respectively;

[0030] The above equations are solved by weighted least squares method to obtain a more accurate displacement estimate, with the goal of minimizing the following error function, based on the formula:

[0031]

[0032] Among them, (x, y) represents the pixel coordinates on the image plane, are the spatial gradient and temporal gradient of the speckle intensity in the x and y directions, respectively. w(x,y) represents the weight function, and w(x,y)≥0;

[0033] Finally, the two-dimensional displacement field V(x,y,t)=(v x ,v y ,t), the displacement field describes the displacement of the brain surface on a two-dimensional plane. The brain surface is assumed to be a plane, i.e. z = 0, and the surface data is approximately two-dimensional. The displacement field only considers the displacement components v in the x and y directions. x ,v y , time t is the dynamic change dimension of the displacement field.

[0034] Furthermore, an LSTM deep learning model was constructed, using the brain surface displacement field monitored during surgery as the model input and the deep tissue displacement verified by intraoperative ultrasound imaging as the label. The logic of the training model is as follows:

[0035] The input features include the brain surface displacement field data set as V s (x, y, t), captured by the intraoperative laser speckle imaging device, each sample is a displacement sequence within the time window tk:t, the three-dimensional brain network map G = (V, E, W) Reflects the strength of white matter fiber connections between brain regions. For each brain region V i The mean fMRI activation intensity;

[0036] The label data is the deep tissue displacement value, set to V d (x, y, z, t), verified by intraoperative ultrasound imaging, and the surface displacement field V s Align in time and space by interpolation; the surface displacement field V s Normalized to [-1,1], the deep displacement field V d Normalize to [0,1];

[0037] Use GAE to compress the brain network map G into a low-dimensional vector, define the node feature matrix X, and each brain region V i The characteristics of Where (x i ,y i ,z i ) is the coordinate of the center of the brain area, For brain area V i The mean fMRI activation intensity is defined as the adjacency matrix A, which is Binarization is obtained;

[0038] The encoder of GAE uses a graph convolutional network, and the decoder reconstructs the graph structure through the adjacency matrix and outputs a low-dimensional embedding vector g with a dimension of 64:

[0039] g=GAE(A,X)

[0040] The embedding vector g is fused with the input features of LSTM through the attention mechanism, guiding LSTM to focus on the displacement propagation law under anatomical constraints;

[0041] The 3D convolution layer is used to extract the spatiotemporal features of the surface displacement field. The convolution kernel size is 3×3×3, the step size is 1×1×1, the number of channels is 32, and the output feature is expressed as:

[0042] h s =Conv3D[V s (tk:t)]

[0043] The graph is embedded into the vector g and mapped to the same dimension through the fully connected layer with the spatiotemporal feature h s Concatenate on the feature dimension:

[0044] h concat =[h s ,FC(g)]

[0045] Among them, FC(g) represents the fully connected transformation of g;

[0046] The bidirectional LSTM layer processes the concatenated feature sequence with a time step of k and 128 hidden units. The output is represented as:

[0047]

[0048] The output of LSTM is further mapped to high-dimensional space through a fully connected layer;

[0049] The fully connected layer maps the LSTM output to a high-dimensional space:

[0050]

[0051] Among them, h FC is the output of the bidirectional LSTM layer, which is mapped to a high-dimensional space through a fully connected layer, and the dimension matches the number of input channels of the deconvolution layer;

[0052] The 3D deconvolution layer reconstructs the deep displacement field, the convolution kernel size is 3×3×3, the step size is 2×2×2, and the number of channels is gradually upsampled to the same value as V d At the same resolution, the output is represented as:

[0053]

[0054] The loss function uses weighted mean square error, and the risk weight matrix W risk Definition of functional importance based on deep tissue:

[0055]

[0056] Among them, W risk (x,y,z) is the weight of the deep tissue region (x,y,z), and its value range is [0,1]. is the true deep tissue displacement value, To predict the deep tissue displacement value.

[0057] Furthermore, the real-time brain surface displacement data is input into the trained model, and the logic for outputting the deep tissue displacement prediction value is as follows:

[0058] The intraoperative laser speckle imaging device captures the brain surface displacement field in real time at a fixed sampling rate. After normalizing the real-time surface displacement field using the mean and standard deviation calculated during the training phase, prediction is performed using the trained model. The core components of the model include a bidirectional LSTM layer and a deconvolution layer. Finally, the deconvolution layer Deconv3D generates the predicted deep tissue displacement value:

[0059]

[0060] Among them, h FC is the output of the bidirectional LSTM layer;

[0061] The predicted deep displacement value is converted to Mapping to the preoperative 3D brain network map coordinate system to generate the deep displacement value in the preoperative 3D brain network map coordinate system

[0062]

[0063] Among them, B-spline basis function is used to quickly calculate φ -1 .

[0064] Furthermore, sub-nodes are created to deploy high-frequency EEG and near-infrared spectroscopy equipment in the surgical area. The logic for detecting the electrical activity of brain neurons through high-frequency EEG and monitoring changes in local blood oxygen saturation in the brain through near-infrared spectroscopy is as follows:

[0065] The high-frequency EEG system uses a 128-channel high-density array, positioned according to the international 10-10 system, covering key functional areas identified by preoperative fMRI. Electrode positioning is achieved through automatic registration of preoperative fMRI data with the intraoperative electrode cap, ensuring that the electrode spacing is ≤20 mm. Electrode impedance is controlled to <5 kΩ by preoperative impedance testing. The sampling frequency is set at 2000 Hz, the EEG system resolution is 24 bits, and the bandwidth is 0.1-500 Hz.

[0066] The near-infrared spectroscopy device deploys a dual-wavelength probe array of 690 nm and 830 nm on the scalp adjacent to the surgical area. It uses a 16-channel configuration with 8 transmitters and 8 receivers, a probe spacing of 30 mm, and a detection depth of 2-3 cm. Probe placement is based on automatic registration of preoperative MRI data with intraoperative probes. The device is calibrated using Monte Carlo simulations, with a sampling rate of 10 Hz and a temporal resolution of 100 ms for NIRS data acquisition.

[0067] The logic of constructing the dynamic influence coefficient of neuronal electrical activity and blood oxygen saturation changes through the wavelet time-frequency coherence algorithm is as follows:

[0068] EEG signal processing uses a 1-100 Hz bandpass filter, a fourth-order Butterworth filter to retain the neural oscillation signal, and the FastICA algorithm to decompose the signal into 128 independent components. Artifact components with correlation coefficients > 0.8 are removed. After artifact removal, the EEG signal is reconstructed by linear combination of the independent components.

[0069] NIRS signal processing calculates the concentrations of oxyhemoglobin and deoxyhemoglobin based on the modified Beer-Lambert law, as follows:

[0070]

[0071] where d is the photon path length, estimated by preoperative MRI and Monte Carlo simulation, ∈ HbR =0.18cm -1 mmol -1 L are the extinction coefficients of HbO2 and HbR, and is the light intensity at wavelengths of 690nm and 830nm, and is the reference light intensity, defined as the light intensity when there is no absorption;

[0072] Continuous wavelet transform is used to analyze the time-frequency characteristics. The wavelet basis function is complex Morlet wavelet, which is defined as:

[0073]

[0074] Among them, f0=1Hz is the center frequency, σ ψ =7 is the bandwidth parameter, covering the 1-100 Hz frequency band; EEG and NIRS data with different sampling rates are unified to the same time resolution through interpolation method; i (t) and NIRS signal Y j (t) Perform continuous wavelet transform respectively to obtain wavelet coefficients and

[0075] Calculate the complex wavelet coherence of the EEG of brain area i and the NIRS of area j in frequency band f and time window t:

[0076]

[0077] Here, <·> represents a sliding average with a window length of 1 second and a step length of 0.1 seconds. is the complex conjugate of the NIRS wavelet coefficient, is the wavelet coefficient of the EEG signal, is the NIRS signal of region j; C ij (f, t) is the time-frequency coherence between EEG and NIRS in frequency band f and time t, which quantifies the strength of neurovascular coupling. The value range is [0, 1], where 1 indicates complete synchronization and 0 indicates no correlation. f is the frequency domain division;

[0078] C ij (f, t) is normalized and the average coherence of each frequency band is calculated:

[0079]

[0080] Among them, N f is the number of wavelet scales in the frequency band; For brain regions to ij in a specific frequency band;

[0081] Dynamic influence coefficient matrix:

[0082]

[0083] in, is the dynamic influence coefficient matrix, with a range of [0,1], w band is the frequency band weight, and its value range is [0,1].

[0084] Furthermore, the comprehensive evaluation index is calculated by combining the predicted value of deep tissue displacement of the brain tissue during surgery and the dynamic influence coefficient, and the logic of comparing the comprehensive evaluation index with the preset threshold is as follows:

[0085] The spatial distribution of the predicted value of deep tissue displacement is combined with the functional coupling strength of the dynamic influence coefficient point by point to form a comprehensive evaluation index, which is:

[0086]

[0087] in, is the predicted value of deep tissue displacement of the i-th brain region, is the average dynamic influence coefficient between the i-th brain region and the j-th brain region; Combine the displacement value with the dynamic influence coefficient point by point; is a nonlinear function; S is a comprehensive evaluation index with a value range of [0,1], and N is the number of brain regions;

[0088] Through comprehensive evaluation index S and preset threshold S threshold For comparison:

[0089] When S≥S threshold , indicating that the surgical risk is high and needs to be adjusted; when S threshold , indicating that the surgical risk is low and the current operation can be continued.

[0090] The present invention further provides a surgical process management system, which is used to implement the above-mentioned surgical process management method, specifically comprising:

[0091] The data acquisition module is used to collect the patient's DTI, fMRI, and 3D-T1WI imaging data before surgery, fuse them to generate a three-dimensional brain network map through a non-rigid registration algorithm, and integrate the laser speckle imaging device into the surgical microscope during surgery to capture brain surface displacement data in real time and construct the surface displacement field through the optical flow field algorithm;

[0092] The model building module is used to build a LSTM-based deep learning model. The brain surface displacement field monitored during surgery is used as the model input, and the deep tissue displacement verified by intraoperative ultrasound imaging is used as the label training model. The real-time brain surface displacement data is input into the trained model, and the deep tissue displacement prediction value is output.

[0093] The data calculation module is used to deploy high-frequency EEG and near-infrared spectroscopy equipment in the surgical area. High-frequency EEG is used to detect the electrical activity of brain neurons, and near-infrared spectroscopy is used to monitor changes in local blood oxygen saturation in the brain. The dynamic influence coefficient of neuronal electrical activity and blood oxygen saturation changes is constructed using a wavelet time-frequency coherence algorithm.

[0094] ​The result judgment module is used to calculate the comprehensive evaluation index by combining the deep tissue displacement prediction value of the brain tissue during surgery and the dynamic influence coefficient, and compare the comprehensive evaluation index with the preset threshold to determine whether the surgery should be adjusted.

[0095] Compared with the prior art, the present invention has the following beneficial effects:

[0096] The present invention collects the patient's DTI, fMRI and 3D-T1WI imaging data and uses a non-rigid registration algorithm to generate a three-dimensional brain network map. This technical feature solves the problem that traditional imaging technology is difficult to provide real-time and continuous information in dynamic monitoring. By fusing multiple imaging data, it can more comprehensively reflect the structure and function of brain tissue, enabling surgeons to have a more accurate understanding of the anatomy and functional positioning of brain regions. By constructing an LSTM deep learning model, the brain surface displacement field monitored in real time during surgery is used as input, and the deep tissue displacement verified by intraoperative ultrasound imaging is used as a label for model training. This technical feature effectively solves the problem of insufficient accuracy and timeliness of traditional methods in real-time prediction of deep tissue displacement.

[0097] The present invention introduces high-frequency EEG and near-infrared spectroscopy monitoring technology to detect the electrical activity of brain neurons and changes in blood oxygen saturation in real time, and constructs the dynamic influence coefficient of neuronal electrical activity and blood oxygen saturation changes through the wavelet time-frequency coherence algorithm. This technical feature further enhances the real-time assessment capability of brain functional status, and solves the deficiency of existing technologies in the lack of comprehensive analysis when processing multimodal monitoring data. Real-time monitoring of EEG activity and blood oxygen changes can help surgeons identify danger signals in a timely manner, thereby adjusting surgical strategies more accurately.

[0098] The present invention combines the predicted value of deep tissue displacement with the dynamic influence coefficient to calculate a comprehensive evaluation index, and compares it with a preset threshold. This process not only realizes the dynamic assessment of surgical risks, but also enables surgeons to make rapid and effective adjustments during the operation. This technical feature, which integrates deep learning and multimodal data analysis, significantly improves the response speed and decision-making quality during the operation, and reduces the probability of postoperative complications. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 Schematic diagram of the overall method flow of the present invention;

[0100] Figure 2 This is a flow chart of the overall system module of the present invention. DETAILED DESCRIPTION

[0101] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0102] It should be noted that, unless otherwise defined, the technical or scientific terms used in the present invention should have the usual meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0103] Example:

[0104] See also Figure 1 , the present invention provides a technical solution:

[0105] A surgical process management method, comprising the following steps:

[0106] Step 1: Preoperatively, the patient's DTI, fMRI, and 3D-T1WI imaging data are collected and fused using a non-rigid registration algorithm to generate a three-dimensional brain network map. During surgery, a laser speckle imaging device is integrated into the surgical microscope to capture brain surface displacement data in real time, and the surface displacement field is constructed using an optical flow algorithm.

[0107] Before surgery, the patient's DTI, fMRI, and 3D-T1WI imaging data were collected and fused using a non-rigid registration algorithm to generate a three-dimensional brain network map. The logic is as follows:

[0108] Diffusion tensor imaging (DTI) is used to obtain information about white matter fiber bundles, estimate the diffusion characteristics of water molecules along axons, and reconstruct the orientation of white matter fibers. Functional magnetic resonance imaging (fMRI) reflects the functional activity of different brain regions and is used to identify important functional areas. 3D-T1WI (three-dimensional T1-weighted imaging) obtains high-resolution brain anatomical structure and defines gray matter, white matter, and cerebrospinal fluid segmentation masks.

[0109] Anisotropic diffusion filtering is applied to DTI and fMRI data. DTI uses Log-Euclidean anisotropic diffusion filtering, and fMRI uses the Perona-Malik equation to reduce noise. Multimodal data are aligned to a standard space, such as MNI152, through affine transformation. Non-brain tissue is removed using morphological opening operations based on 3D-T1WI masks. The goal of non-rigid registration is to align multimodal data DTI, fMRI, and 3D-T1WI so that the data of different modalities are spatially consistent, providing a unified spatial framework for subsequent brain network construction.

[0110] Setting I DTI ,I T1 Represent DTI and T1 image data respectively, and optimize the deformation field through the non-rigid registration algorithm of the B-spline free deformation model Minimize the difference between multimodal images. The objective function is defined as:

[0111]

[0112] Among them, NMI is the normalized mutual information, which is used to measure the similarity of the grayscale distribution between the registered T1 image and the deformed DTI image. is the Dice coefficient, which is used to measure the spatial overlap between the registered T1 image and the fMRI functional area mask. fMRI is the fMRI functional area mask, indicating the spatial distribution of the functional area, M T1 is the anatomical segmentation mask of the T1 image, representing the spatial distribution of the anatomical region; α and β are weight coefficients, and α + β = 1, which are used to balance the proportion of grayscale distribution similarity and spatial overlap in the objective function; the objective function is solved by gradient descent or other optimization algorithms to obtain the optimal deformation field φ; the non-rigid deformation field φ is to align the DTI and fMRI images to the 3D-T1WI mapping to minimize the difference between multimodal images; the non-rigid deformation field φ represents the displacement of each 3D coordinate point:

[0113] NMI normalized mutual information, which measures the anatomical structure of two images. T1 and after registration Statistical dependence of:

[0114]

[0115] Where H(A) is the entropy of image A, H(B) is the entropy of image B, and H(A,B) is the joint entropy. The anatomical structures of DTI and T1WI are quite different, i.e., white matter fibers vs. gray matter boundaries. Mutual information is suitable for cross-modality registration.

[0116] Dice is the Dice coefficient, and the formula is:

[0117]

[0118] Measuring fMRI activation area mask M fMRI Anatomical mask with T1WI Dice ensures that the fMRI activation area accurately fits the anatomical functional area, such as the motor area in the precentral gyrus;

[0119] Through non-rigid registration, multimodal data spatial alignment is achieved. Combined with fiber tracking, functional area probability and anatomical segmentation, a three-dimensional brain network map G = (V, E, W) is finally generated with brain regions as nodes, white matter connections as edges, and functional area weights as attributes. Among them, V is a set of brain regions, each node corresponds to a brain region, and the definition of brain regions is based on the functional activation map of fMRI. Brain regions are usually divided into several functional areas through clustering algorithms such as K-means, or based on anatomical templates such as AAL90 definition. E is the edge set of white matter fiber connections, which represents the anatomical connection between brain regions. The definition of edges is based on the fiber tracking results of DTI. Common algorithms include deterministic tracking, such as FACT algorithm and probabilistic tracking. W is the weight set of edges, which represents the functional importance of the connection.

[0120] The definition of brain region nodes V is based on functional activation maps of fMRI, and the brain region is usually divided into several functional areas through clustering algorithms such as K-means. The definition of white matter fiber connections E is based on the tracking results of DTI. Common algorithms include deterministic tracking, such as the FACT algorithm, and probabilistic tracking.

[0121] Edge weights of white matter fiber connections The strength of the fiber bundle is expressed based on the formula:

[0122]

[0123] Among them, S ij represents the number of fiber bundles from brain region i to j, obtained by DTI fiber tracking statistics, FA ij is the anisotropy fraction of the fiber bundle, reflecting the integrity and directional consistency of the white matter fibers, with a value range of [0,1]; mean(FA ij ) the mean FA of all fibers connecting brain areas i and j;

[0124] The node weights based on fMRI are based on the formula:

[0125]

[0126] Among them, V i is the spatial extent of brain region i based on the anatomical template, such as the AAL90 definition, |Vi | is the number of voxels in brain region i, i.e., its volume, P func (x) is the fMRI activation probability at position x, obtained by thresholding the BOLD signal and defined as:

[0127]

[0128] Among them, S BOLD is the BOLD signal intensity at position x, after normalization P func (x)∈[0,1];

[0129] Weight set: Where a and b are weight coefficients, and a+b=1, which are used to balance the proportion of anatomical connection strength and functional activity importance. is the edge weight of the white matter fiber connection, is the weight of the brain region node;

[0130] During surgery, the laser speckle imaging device was integrated into the surgical microscope to capture brain surface displacement data in real time. The logic for constructing the surface displacement field using the optical flow algorithm was as follows:

[0131] A miniature laser speckle imaging device (LSCI) is integrated into the surgical microscope. Laser speckle imaging technology is based on the principle of light scattering. When laser light irradiates the brain surface, the scattered light produces a speckle pattern due to the microstructure and dynamic changes of brain tissue. The temporal changes in the speckle pattern reflect the microscopic movement of the brain surface. By capturing and analyzing these temporal changes in the speckle pattern, real-time displacement information of the brain surface can be obtained. The spatial and temporal resolution of speckle imaging has a significant impact on the accuracy of displacement estimation, and appropriate parameters must be selected based on surgical requirements.

[0132] Setting I t (x, y) is the speckle image collected at time t, I t+1 (x, y) is the speckle image collected at time t+1, where t represents the time frame number and (x, y) represents the coordinates on the image plane. Assuming that the brain surface displacement is a small-scale smooth deformation that satisfies the Lucas-Kanade optical flow algorithm's brightness constant assumption, that is, the change in speckle intensity in a short period of time can be ignored, the Lucas-Kanade optical flow algorithm is used to calculate the displacement field, based on the formula:

[0133]

[0134] in, are the spatial gradients of the speckle intensity I in the x and y directions, is the time gradient of speckle intensity I, v x ,v yare the velocity components of the surface displacement in the x and y directions, respectively;

[0135] According to the constant brightness assumption, the change of speckle intensity can be expressed as:

[0136] I(x,y,t)≈I(x+v x ,y+v y ,t+1)

[0137] The above formula is obtained by Taylor expansion and ignoring high-order terms. The Lucas-Kanade algorithm calculates the velocity component of each pixel by minimizing this formula.

[0138] The above equations are solved by weighted least squares method to obtain a more accurate displacement estimate, with the goal of minimizing the following error function, based on the formula:

[0139]

[0140] Among them, (x, y) represents the pixel coordinates on the image plane, are the spatial gradient and temporal gradient of the speckle intensity in the x and y directions, respectively. w(x,y) represents the weight function, which is used to consider the contribution of different pixels to the displacement estimation. w(x,y)≥0. The weight is assigned based on the amplitude of the image gradient and is defined as:

[0141]

[0142] Among them, ∈ is a small positive number used to avoid the denominator being zero; by minimizing the error function, the velocity component v of each pixel can be obtained x ,v y When solving the estimated value of , the error function is usually linearized and solved iteratively through matrix operations, such as the Gauss-Newton method;

[0143] Finally, the two-dimensional displacement field V(x,y,t)=(v x ,v y ,t), the displacement field describes the displacement of the brain surface on a two-dimensional plane. The brain surface is assumed to be a plane, i.e. z = 0, and the surface data is approximately two-dimensional. The displacement field only considers the displacement components v in the x and y directions. x ,v y , time t is the dynamic change dimension of the displacement field.

[0144] Step 2: Build an LSTM-based deep learning model. Use the brain surface displacement field monitored during surgery as the model input and the deep tissue displacement verified by intraoperative ultrasound imaging as the label training model. Input the real-time brain surface displacement data into the trained model and output the predicted deep tissue displacement value.

[0145] Build an LSTM deep learning model, use the brain surface displacement field monitored during surgery as the model input, and the deep tissue displacement verified by intraoperative ultrasound imaging as the label. The logic of training the model is as follows:

[0146] The input features include the brain surface displacement field data set as V s (x, y, t), captured by the intraoperative laser speckle imaging device. Each sample is a displacement sequence within the time window tk:t. The time window length k is set according to the data sampling rate and task requirements. For example, k = 5 means using the data of the past 5 frames. The three-dimensional brain network map G = (V, E, W) Reflects the strength of white matter fiber connections between brain regions. For each brain region V i The mean fMRI activation intensity of the label data is the deep tissue displacement value, set as V d (x, y, z, t), verified by intraoperative ultrasound imaging, and the surface displacement field V s Align in time and space by interpolation to ensure consistent temporal and spatial resolution; the surface displacement field V s Normalized to [-1,1], the deep displacement field V d Normalize to [0,1];

[0147] The normalization formula is:

[0148]

[0149] Among them, min(V s ),max(V s ) are V s The minimum and maximum values, min(V d ),max(V d ) are V d The minimum and maximum values ​​of,the normalization parameters are calculated based on the training set and,applied to the test set to ensure consistency;

[0150] GAE is used to compress the brain network map G into a low-dimensional vector, preserving topological characteristics such as node degree distribution and modularity; defining the node feature matrix X, each brain region V i The characteristics of Where (x i ,y i ,z i ) is the coordinate of the center of the brain area, For brain area V i The mean fMRI activation intensity is defined as the adjacency matrix A, which is The binarization threshold is set to to ensure that the primary connections are preserved:

[0151]

[0152] The encoder of GAE uses a graph convolutional network, and the decoder reconstructs the graph structure through the adjacency matrix and outputs a low-dimensional embedding vector g with a dimension of 64:

[0153] g=GAE(A,X)

[0154] Among them, g encodes the structural connectivity pattern and functional importance distribution between brain regions. The embedding vector g is fused with the input features of LSTM through the attention mechanism, guiding LSTM to focus on the displacement propagation law under anatomical constraints.

[0155] The 3D convolution layer is used to extract the spatiotemporal features of the surface displacement field. The convolution kernel size is 3×3×3, the step size is 1×1×1, the number of channels is 32, and the output feature is expressed as:

[0156] h s =Conv3D[V s (tk:t)]

[0157] The graph is embedded into the vector g and mapped to the same dimension through the fully connected layer with the spatiotemporal feature h s Concatenate on the feature dimension:

[0158] h concat =[h s ,FC(g)]

[0159] Among them, FC(g) represents the fully connected transformation of g;

[0160] The bidirectional LSTM layer processes the concatenated feature sequence with a time step of k, which is the same as the input time window length, and 128 hidden units. The output is represented as:

[0161]

[0162] The output of LSTM is further mapped to high-dimensional space through a fully connected layer;

[0163] The fully connected layer maps the LSTM output to a high-dimensional space:

[0164]

[0165] Among them, h FC is the output of the bidirectional LSTM layer, which is mapped to a high-dimensional space through a fully connected layer, and the dimension matches the number of input channels of the deconvolution layer;

[0166] The 3D deconvolution layer reconstructs the deep displacement field, the convolution kernel size is 3×3×3, the step size is 2×2×2, and the number of channels is gradually upsampled to the same value as V dAt the same resolution, the output is represented as:

[0167]

[0168] The loss function uses weighted mean square error, and the risk weight matrix W risk Definition of functional importance based on deep tissue:

[0169]

[0170] Among them, W risk (x,y,z) is the weight of the deep tissue region (x,y,z), and its value range is [0,1]. The higher the weight, the more important the region. is the true deep tissue displacement value, To predict the displacement value of deep tissue;

[0171] The logic of inputting the real-time brain surface displacement data into the trained model and outputting the predicted value of deep tissue displacement is as follows:

[0172] The intraoperative laser speckle imaging device captures the brain surface displacement field in real time at a fixed sampling rate, such as 30 Hz. After normalizing the real-time surface displacement field using the mean and standard deviation calculated during the training phase, prediction is performed using the trained model. The core components of the model include a bidirectional LSTM layer and a deconvolution layer. Finally, the deconvolution layer Deconv3D generates the predicted deep tissue displacement value:

[0173]

[0174] Among them, h FC is the output of the bidirectional LSTM layer;

[0175] The predicted deep displacement value is converted to Mapping to the preoperative 3D brain network map coordinate system to generate the deep displacement value in the preoperative 3D brain network map coordinate system

[0176]

[0177] Among them, B-spline basis function is used to quickly calculate φ -1 .

[0178] set up is the preoperative non-rigid registration deformation field, The inverse transformation is used to map the intraoperative coordinates (x, y, z) to the preoperative three-dimensional brain network map coordinates (x', y', z'): (x', y', z') = φ -1 (x,y,z);

[0179] Fast calculation of φ using B-spline basis functions -1 , the formula is:

[0180]

[0181] Among them, B i (x),B y (y),B k (z) is the B-spline basis function, c ijk is the corresponding control point coefficient;

[0182] For the predicted deep displacement value The mapped value is calculated using the trilinear interpolation method:

[0183]

[0184] Where W(x',x) is the trilinear interpolation weight, defined as:

[0185]

[0186] Define W(y',y) and W(z',z) similarly.

[0187] Step 3: Deploy high-frequency EEG and near-infrared spectroscopy equipment in the surgical area. High-frequency EEG is used to detect the electrical activity of brain neurons, and near-infrared spectroscopy is used to monitor changes in local blood oxygen saturation in the brain. Dynamic influence coefficients of changes in neuronal electrical activity and blood oxygen saturation are constructed using a wavelet time-frequency coherence algorithm.

[0188] The logic behind deploying high-frequency EEG and near-infrared spectroscopy equipment in the surgical area, using high-frequency EEG to detect the electrical activity of brain neurons and near-infrared spectroscopy to monitor changes in local brain blood oxygen saturation, is as follows:

[0189] The high-frequency EEG system uses a 128-channel high-density array, positioned according to the international 10-10 system, covering key functional areas identified by preoperative fMRI, such as motor and language areas. Electrode positioning is achieved through automatic registration of preoperative fMRI data with the intraoperative electrode cap, ensuring that the electrode spacing is ≤20mm and spatial resolution is guaranteed. Electrode impedance is controlled to <5kΩ through preoperative impedance testing. The sampling frequency is set to 2000Hz to meet the Nyquist sampling theorem and avoid high-frequency signal aliasing. The EEG system has a resolution of 24 bits and a bandwidth of 0.1-500Hz, covering the main frequency bands of EEG signals. Active shielding electrodes are used to reduce intraoperative electrocautery interference and ensure signal quality.

[0190] The near-infrared spectroscopy device deploys a dual-wavelength probe array, 690nm and 830nm, on the scalp adjacent to the surgical area. It uses a 16-channel configuration with 8 transmitters and 8 receivers, a probe spacing of 30mm, and a detection depth of 2-3cm. Probe placement is based on automatic registration of preoperative MRI data with intraoperative probes to ensure coverage of critical functional areas. The device is calibrated using Monte Carlo simulations, with a sampling rate of 10Hz and a temporal resolution of 100ms for NIRS data acquisition.

[0191] The EEG, NIRS, and surgical navigation system master clocks are connected via BNC interfaces, achieving time alignment accuracy of ≤1ms. Time synchronization is verified through preoperative calibration and intraoperative real-time monitoring to ensure spatiotemporal consistency of multimodal data.

[0192] The logic of constructing the dynamic influence coefficient of neuronal electrical activity and blood oxygen saturation changes through the wavelet time-frequency coherence algorithm is as follows:

[0193] EEG signal processing uses a 1-100 Hz bandpass filter. A fourth-order Butterworth filter preserves neural oscillation signals and suppresses low-frequency drift and high-frequency noise. A 50 / 60 Hz notch filter eliminates power frequency interference. A Q factor of 30 represents the ratio of the filter bandwidth to the center frequency. The FastICA algorithm is used to decompose the signal into 128 independent components. The correlation coefficient of each component with the electrooculogram (EOG) and electrocardiogram (ECG) templates is calculated. Artifact components with correlation coefficients > 0.8 are removed. After artifact removal, the EEG signal is reconstructed by linear combination of the independent components.

[0194] NIRS signal processing calculates the concentrations of oxyhemoglobin (HbO2) and deoxyhemoglobin (HbR) based on the modified Beer-Lambert law, as follows:

[0195]

[0196] where d is the photon path length, estimated by preoperative MRI and Monte Carlo simulation, ∈ HbR =0.18cm -1 mmol -1 L are the extinction coefficients of HbO2 and HbR, and is the light intensity at wavelengths of 690nm and 830nm, and is the reference light intensity, defined as the light intensity when there is no absorption;

[0197] Based on the surgical event markers, a 5-second time window was captured with a 50% overlap rate to generate synchronized EEG-NIRS data segments. Each sample contained: EEG, 128 channels × 2000 Hz × 5 seconds = 10,000 points / channel; NIRS, 16 channels × 10 Hz × 5 seconds = 50 points / channel;

[0198] Continuous wavelet transform is used to analyze the time-frequency characteristics. The wavelet basis function is complex Morlet wavelet, which is defined as:

[0199]

[0200] Among them, f0=1Hz is the center frequency, σ ψ =7 is the bandwidth parameter, covering the 1-100 Hz frequency band; EEG and NIRS data with different sampling rates are unified to the same temporal resolution through interpolation method;

[0201] For EEG signal X i (t) and NIRS signal Y j (t) Perform continuous wavelet transform respectively to obtain wavelet coefficients and

[0202] Calculate the complex wavelet coherence of the EEG of brain area i and the NIRS of area j in frequency band f and time window t:

[0203]

[0204] Here, <·> represents a sliding average with a window length of 1 second and a step length of 0.1 seconds. is the complex conjugate of the NIRS wavelet coefficient, is the wavelet coefficient of the EEG signal, is the NIRS signal of region j, i.e., the wavelet coefficient of HbO2 or HbR concentration; C ij (f, t) is the time-frequency coherence between EEG and NIRS in frequency band f and time t, which quantifies the strength of neurovascular coupling. The value range is [0, 1], where 1 indicates complete synchronization and 0 indicates no correlation. f is the frequency domain division;

[0205] C ij (f, t) is normalized to avoid numerical instability and the average coherence of each frequency band is calculated:

[0206]

[0207] Among them, N f is the number of wavelet scales in the frequency band; It is the average coherence of brain regions to ij in a specific frequency band, such as Gamma, reflecting the coupling strength of a specific neural oscillation frequency band;

[0208] Dynamic influence coefficient matrix:

[0209]

[0210] in, is the dynamic influence coefficient matrix, which integrates the multi-band coherence and quantifies the real-time functional coupling state during the operation. Its value range is [0,1]. band is the frequency band weight, with a value range of [0,1]; Gamma (30-100Hz): reflects cognition and motor control, Beta (13-30Hz): reflects motor inhibition and concentration, Alpha (8-12Hz): reflects resting state relaxation, Theta (4-7Hz): reflects memory and emotion, Delta (1-3Hz): reflects deep sleep or pathological state.

[0211] Step 4: Calculate a comprehensive evaluation index based on the predicted deep tissue displacement of the brain tissue during surgery and the dynamic influence coefficient. Compare the comprehensive evaluation index with the preset threshold to determine whether the surgery should be adjusted.

[0212] The logic for calculating the comprehensive evaluation index by combining the predicted value of deep tissue displacement of the brain tissue during surgery and the dynamic influence coefficient and comparing the comprehensive evaluation index with the preset threshold is as follows:

[0213] Prediction of deep tissue displacement Indicates the actual displacement of brain tissue during surgery. The greater the displacement, the higher the risk to the surgery. represents the functional coupling strength between the i-th brain region and the j-th brain region. The stronger the coupling, the greater the impact of displacement on function.

[0214] The spatial distribution of the predicted value of deep tissue displacement is combined with the functional coupling strength of the dynamic influence coefficient point by point to form a comprehensive evaluation index, which is:

[0215]

[0216] in, is the predicted value of deep tissue displacement of the i-th brain region, is the average dynamic influence coefficient between the i-th brain region and the j-th brain region, after comprehensive processing in the time and frequency domains; The displacement value is combined with the dynamic influence coefficient point by point to reflect the influence of the displacement on the function; It is a nonlinear function used to nonlinearly amplify or reduce the displacement value and enhance the sensitivity to high displacement areas; S is a comprehensive evaluation index with a value range of [0,1], and N is the number of brain regions;

[0217] Squaring the binding value amplifies the effects of high displacement and high coupling regions; right Perform normalization processing to ensure that the values ​​of comprehensive evaluation indicators are within a reasonable range;

[0218] In order to enhance the sensitivity to high displacement areas, a nonlinear function is designed Its form is:

[0219]

[0220] in, To square the displacement value and enhance the impact of high displacement areas; Is a Gaussian function used to adjust the displacement value nonlinearly: V peak is the peak area of ​​displacement, and the influence near this area is further amplified, σ f is the standard deviation of the Gaussian function, which controls the range of nonlinear regulation. λ is the regulation coefficient, which is used to control the influence intensity of the Gaussian function. The value of is always positive, ensuring that the denominator is not zero;

[0221] In order to quantify the dynamic influence coefficient more comprehensively, the time and frequency domains are processed to obtain

[0222]

[0223] in, is the value of the dynamic influence coefficient in the frequency domain f and time t, f∈[f1,f2] is the frequency domain range, T is the time range, which represents the upper limit time of the integration of the dynamic influence coefficient; g(f) is the frequency domain weight function, which is used to enhance the influence of a specific frequency band, such as the Gamma band:

[0224]

[0225] h(t) is a time weight function used to enhance the impact of a specific time period:

[0226]

[0227] Through comprehensive evaluation index S and preset threshold S threshold For comparison:

[0228] When S≥S threshold , indicating that the surgical risk is high and needs to be adjusted; when S threshold , indicating that the surgical risk is low and the current operation can be continued.

[0229] See also Figure 2 ​The present invention also provides a surgical process management system, which is used to implement the above-mentioned surgical process management method, specifically comprising:

[0230] The data acquisition module is used to collect the patient's DTI, fMRI, and 3D-T1WI imaging data before surgery, fuse them to generate a three-dimensional brain network map through a non-rigid registration algorithm, and integrate the laser speckle imaging device into the surgical microscope during surgery to capture brain surface displacement data in real time and construct the surface displacement field through the optical flow field algorithm;

[0231] The model building module is used to build a LSTM-based deep learning model. The brain surface displacement field monitored during surgery is used as the model input, and the deep tissue displacement verified by intraoperative ultrasound imaging is used as the label training model. The real-time brain surface displacement data is input into the trained model, and the deep tissue displacement prediction value is output.

[0232] The data calculation module is used to deploy high-frequency EEG and near-infrared spectroscopy equipment in the surgical area. High-frequency EEG is used to detect the electrical activity of brain neurons, and near-infrared spectroscopy is used to monitor changes in local blood oxygen saturation in the brain. The dynamic influence coefficient of neuronal electrical activity and blood oxygen saturation changes is constructed using a wavelet time-frequency coherence algorithm.

[0233] The result judgment module is used to calculate the comprehensive evaluation index by combining the deep tissue displacement prediction value of the brain tissue during surgery and the dynamic influence coefficient, and compare the comprehensive evaluation index with the preset threshold to determine whether the surgery should be adjusted.

[0234] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0235] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraints of the technical solution.

[0236] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment. The above description is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of this application.

Claims

1. A surgical process management method, characterized in that: The specific steps include: Step 1: Preoperatively, the patient's DTI, fMRI, and 3D-T1WI imaging data are collected and fused using a non-rigid registration algorithm to generate a three-dimensional brain network map. During surgery, a laser speckle imaging device is integrated into the surgical microscope to capture brain surface displacement data in real time, and the surface displacement field is constructed using an optical flow algorithm. Step 2: Build an LSTM deep learning model, using the brain surface displacement field monitored during surgery as the model input and the deep tissue displacement verified by intraoperative ultrasound imaging as the label training model. Input the real-time brain surface displacement data into the trained model, and output the predicted deep tissue displacement value. Step 3: Deploy high-frequency EEG and near-infrared spectroscopy equipment in the surgical area. High-frequency EEG is used to detect the electrical activity of brain neurons, and near-infrared spectroscopy is used to monitor changes in local blood oxygen saturation in the brain. Dynamic influence coefficients of changes in neuronal electrical activity and blood oxygen saturation are constructed using a wavelet time-frequency coherence algorithm. Step 4: Calculate a comprehensive evaluation index based on the predicted deep tissue displacement of the brain tissue during surgery and the dynamic influence coefficient, and compare the comprehensive evaluation index with the preset threshold to determine whether to adjust the surgery.

2. A surgical process management method according to claim 1, characterized in that: Before surgery, the patient's DTI, fMRI, and 3D-T1WI imaging data were collected and fused using a non-rigid registration algorithm to generate a three-dimensional brain network map. The logic is as follows: DTI is used to obtain white matter fiber bundle information, fMRI is used to identify important functional areas, and 3D-T1WI is used to obtain high-resolution brain anatomical structure and define gray matter, white matter, and cerebrospinal fluid segmentation masks; Setting I DTI ,I T1 Represent DTI and T1 image data respectively, and optimize the deformation field through the non-rigid registration algorithm of the B-spline free deformation model Minimize the difference between multimodal images. The objective function is defined as: Among them, NMI is the normalized mutual information, is the Dice coefficient, M fMRI is the fMRI functional area mask, M T1 is the anatomical segmentation mask of the T1 image; α, β are weight coefficients, and α+β=1; By achieving spatial alignment of multimodal data through non-rigid registration and combining fiber tracking, functional area probability, and anatomical segmentation, we ultimately generate a three-dimensional brain network map G = (V, E, W) with brain regions as nodes, white matter connections as edges, and functional area weights as attributes. V is a set of brain regions, with each node corresponding to a brain region; E is the set of edges of white matter fiber connections, representing the anatomical connections between brain regions; and W is the set of edge weights, representing the functional importance of the connections. Edge weights of white matter fiber connections The strength of the fiber bundle is expressed based on the formula: Among them, S ij represents the number of fiber bundles from brain area i to j, FA ij is the anisotropy fraction of the fiber bundle, ranging from [0,1], mean(FA ij ) the mean FA of all fibers connecting brain areas i and j; The node weights based on fMRI are based on the formula: Among them, V i is the i-th brain region, |V i | is the number of voxels in brain region i, i.e., its volume, P func (x) is the fMRI activation probability at position x; Weight set: Where a, b are weight coefficients, and a+b=1, is the edge weight of the white matter fiber connection, is the weight of the brain region node.

3. A surgical process management method according to claim 2, characterized in that: During surgery, the laser speckle imaging device was integrated into the surgical microscope to capture brain surface displacement data in real time. The logic for constructing the surface displacement field using the optical flow algorithm was as follows: Setting I t (x, y) is the speckle image collected at time t, I t+1 (x, y) is the speckle image acquired at time t+1, where t represents the time frame number and (x, y) represents the coordinates on the image plane. Assuming that the brain surface displacement is a small-scale smooth deformation, the Lucas-Kanade optical flow algorithm is used to calculate the displacement field. The formula is: in, are the spatial gradients of the speckle intensity I in the x and y directions, is the time gradient of speckle intensity I, v x ,v y are the velocity components of the surface displacement in the x and y directions, respectively; The above equations are solved by weighted least squares method to obtain a more accurate displacement estimate, with the goal of minimizing the following error function, based on the formula: Among them, (x, y) represents the pixel coordinates on the image plane, are the spatial gradient and temporal gradient of the speckle intensity in the x and y directions, respectively. w(x,y) represents the weight function, and w(x,y)≥0; Finally, the two-dimensional displacement field V(x,y,t)=(v x ,v y ,t), the displacement field describes the displacement of the brain surface on a two-dimensional plane. The brain surface is assumed to be a plane, i.e. z = 0, and the surface data is approximately two-dimensional. The displacement field only considers the displacement components v in the x and y directions. x ,v y , time t is the dynamic change dimension of the displacement field.

4. A surgical process management method according to claim 3, characterized in that: Build an LSTM deep learning model, use the brain surface displacement field monitored during surgery as the model input, and the deep tissue displacement verified by intraoperative ultrasound imaging as the label. The logic of training the model is as follows: The input features include the brain surface displacement field data set as V s (x, y, t), captured by the intraoperative laser speckle imaging device, each sample is a displacement sequence within the time window tk:t, the three-dimensional brain network map G = (V, E, W) Reflects the strength of white matter fiber connections between brain regions. For each brain region V i The mean fMRI activation intensity; The label data is the deep tissue displacement value, set to V d (x, y, z, t), verified by intraoperative ultrasound imaging, and the surface displacement field V s Align in time and space by interpolation; the surface displacement field V s Normalized to [-1,1], the deep displacement field V d Normalize to [0,1]; Use GAE to compress the brain network map G into a low-dimensional vector, define the node feature matrix X, and each brain region V i The characteristics of Where (x i ,y i ,z i ) is the coordinate of the center of the brain area, For brain area V i The mean fMRI activation intensity is defined as the adjacency matrix A, which is Binarization is obtained; The encoder of GAE uses a graph convolutional network, and the decoder reconstructs the graph structure through the adjacency matrix and outputs a low-dimensional embedding vector g with a dimension of 64: g=GAE(A,X) The embedding vector g is fused with the input features of LSTM through the attention mechanism, guiding LSTM to focus on the displacement propagation law under anatomical constraints; The 3D convolution layer is used to extract the spatiotemporal features of the surface displacement field. The convolution kernel size is 3×3×3, the step size is 1×1×1, the number of channels is 32, and the output feature is expressed as: h s =Conv3D[V s (t-k:t)] The graph is embedded into the vector g and mapped to the same dimension through the fully connected layer with the spatiotemporal feature h s Concatenate on the feature dimension: h concat =[h s ,FC(g)] Among them, FC(g) represents the fully connected transformation of g; The bidirectional LSTM layer processes the concatenated feature sequence with a time step of k and 128 hidden units. The output is represented as: The output of LSTM is further mapped to high-dimensional space through a fully connected layer; The fully connected layer maps the LSTM output to a high-dimensional space: Among them, h FC The output of the bidirectional LSTM layer is mapped to a high-dimensional space through a fully connected layer, and the dimension matches the number of input channels of the deconvolution layer; The 3D deconvolution layer reconstructs the deep displacement field, the convolution kernel size is 3×3×3, the step size is 2×2×2, and the number of channels is gradually upsampled to the same value as V d At the same resolution, the output is represented as: The loss function uses weighted mean square error, and the risk weight matrix W risk Definition of functional importance based on deep tissue: Among them, W risk (x,y,z) is the weight of the deep tissue region (x,y,z), and its value range is [0,1]. is the true deep tissue displacement value, To predict the deep tissue displacement value.

5. A surgical process management method according to claim 4, characterized in that: The logic of inputting the real-time brain surface displacement data into the trained model and outputting the predicted value of deep tissue displacement is as follows: The intraoperative laser speckle imaging device captures the brain surface displacement field in real time at a fixed sampling rate. After normalizing the real-time surface displacement field using the mean and standard deviation calculated during the training phase, prediction is performed using the trained model. The core components of the model include a bidirectional LSTM layer and a deconvolution layer. Finally, the deconvolution layer Deconv3D generates the predicted deep tissue displacement value: Among them, h FC is the output of the bidirectional LSTM layer; The predicted deep displacement value is converted to Mapping to the preoperative 3D brain network map coordinate system to generate the deep displacement value under the preoperative 3D brain network map coordinate system Among them, B-spline basis function is used to quickly calculate φ -1 .

6. A surgical process management method according to claim 5, characterized in that: Create a child node and deploy high-frequency EEG and near-infrared spectroscopy equipment in the surgical area. The logic for detecting the electrical activity of brain neurons using high-frequency EEG and monitoring changes in local brain blood oxygen saturation using near-infrared spectroscopy is as follows: The high-frequency EEG system uses a 128-channel high-density array, positioned according to the international 10-10 system, covering key functional areas identified by preoperative fMRI. Electrode positioning is achieved through automatic registration of preoperative fMRI data with the intraoperative electrode cap, ensuring that the electrode spacing is ≤20 mm. Electrode impedance is controlled to <5 kΩ by preoperative impedance testing. The sampling frequency is set at 2000 Hz, the EEG system resolution is 24 bits, and the bandwidth is 0.1-500 Hz. The near-infrared spectroscopy device deploys a dual-wavelength probe array of 690 nm and 830 nm on the scalp adjacent to the surgical area. It uses a 16-channel configuration with 8 transmitters and 8 receivers, a probe spacing of 30 mm, and a detection depth of 2-3 cm. Probe placement is based on automatic registration of preoperative MRI data with intraoperative probes. The device is calibrated using Monte Carlo simulations, with a sampling rate of 10 Hz and a temporal resolution of 100 ms for NIRS data acquisition. The logic of constructing the dynamic influence coefficient of neuronal electrical activity and blood oxygen saturation changes through the wavelet time-frequency coherence algorithm is as follows: EEG signal processing uses a 1-100 Hz bandpass filter, a fourth-order Butterworth filter to retain the neural oscillation signal, and the FastICA algorithm to decompose the signal into 128 independent components. Artifact components with correlation coefficients > 0.8 are removed. After artifact removal, the EEG signal is reconstructed by linear combination of the independent components. NIRS signal processing calculates the concentrations of oxyhemoglobin and deoxyhemoglobin based on the modified Beer-Lambert law, as follows: where d is the photon path length, estimated by preoperative MRI and Monte Carlo simulation, ∈ HbR =0.18cm -1 mmol -1 L are the extinction coefficients of HbO2 and HbR, and is the light intensity at wavelengths of 690nm and 830nm, and is the reference light intensity, defined as the light intensity when there is no absorption; Continuous wavelet transform is used to analyze the time-frequency characteristics. The wavelet basis function is complex Morlet wavelet, which is defined as: Among them, f0=1Hz is the center frequency, σ ψ =7 is the bandwidth parameter, covering the 1-100 Hz frequency band; EEG and NIRS data with different sampling rates are unified to the same time resolution through interpolation method; i (t) and NIRS signal Y j (t) Perform continuous wavelet transform respectively to obtain wavelet coefficients and Calculate the complex wavelet coherence of the EEG of brain area i and the NIRS of area j in frequency band f and time window t: Here, <·> represents a sliding average with a window length of 1 second and a step length of 0.1 seconds. is the complex conjugate of the NIRS wavelet coefficient, is the wavelet coefficient of the EEG signal, is the NIRS signal of region j; C ij (f, t) is the time-frequency coherence between EEG and NIRS in frequency band f and time t, which quantifies the strength of neurovascular coupling. The value range is [0, 1], where 1 indicates complete synchronization and 0 indicates no correlation. f is the frequency domain division; C ij (f, t) is normalized and the average coherence of each frequency band is calculated: Among them, N f is the number of wavelet scales in the frequency band; For brain regions to ij in a specific frequency band; Dynamic influence coefficient matrix: in, is the dynamic influence coefficient matrix, with a range of [0,1], w band is the frequency band weight, and its value range is [0,1].

7. A surgical process management method according to claim 6, characterized in that: The logic for calculating the comprehensive evaluation index by combining the predicted value of deep tissue displacement of the brain tissue during surgery and the dynamic influence coefficient and comparing the comprehensive evaluation index with the preset threshold is as follows: The spatial distribution of the predicted value of deep tissue displacement is combined with the functional coupling strength of the dynamic influence coefficient point by point to form a comprehensive evaluation index, which is: in, is the predicted value of deep tissue displacement of the i-th brain region, is the average dynamic influence coefficient between the i-th brain region and the j-th brain region; Combine the displacement value with the dynamic influence coefficient point by point; is a nonlinear function; S is a comprehensive evaluation index with a value range of [0,1], and N is the number of brain regions; Through comprehensive evaluation index S and preset threshold S threshold For comparison: When S≥S threshold , indicating that the surgical risk is high and needs to be adjusted; when S threshold , indicating that the surgical risk is low and the current operation can be continued.​ 8. A surgical process management system, characterized by: The system is used to execute a surgical process management method according to any one of claims 1 to 7, comprising: The data acquisition module is used to collect the patient's DTI, fMRI, and 3D-T1WI imaging data before surgery, fuse them to generate a three-dimensional brain network map through a non-rigid registration algorithm, and integrate the laser speckle imaging device into the surgical microscope during surgery to capture brain surface displacement data in real time and construct the surface displacement field through the optical flow field algorithm; The model building module is used to build a LSTM-based deep learning model. The brain surface displacement field monitored during surgery is used as the model input, and the deep tissue displacement verified by intraoperative ultrasound imaging is used as the label training model. The real-time brain surface displacement data is input into the trained model, and the deep tissue displacement prediction value is output. The data calculation module is used to deploy high-frequency EEG and near-infrared spectroscopy equipment in the surgical area. High-frequency EEG is used to detect the electrical activity of brain neurons, and near-infrared spectroscopy is used to monitor changes in local blood oxygen saturation in the brain. The dynamic influence coefficient of neuronal electrical activity and blood oxygen saturation changes is constructed using a wavelet time-frequency coherence algorithm. The result judgment module is used to calculate the comprehensive evaluation index by combining the deep tissue displacement prediction value of the brain tissue during surgery and the dynamic influence coefficient, and compare the comprehensive evaluation index with the preset threshold to determine whether the surgery should be adjusted.

Images