Intraoperative cerebral cortex electrode intelligent positioning system based on deep learning

Abstract

The invention relates to an intraoperative cerebral cortex electrode intelligent positioning system based on deep learning, which belongs to the technical field of surgical navigation, and comprises a data acquisition module used for acquiring data of intraoperative ultrasound, intraoperative MRI and preoperative MRI; the preprocessing module is used for carrying out denoising, contrast enhancement, motion artifact correction, bias field correction, skull stripping and standardization processing on the multi-modal data; the dynamic collaborative segmentation network DSCN fuses MRI and CT features based on a mutual information maximization principle to generate a segmentation Mask; the variational adaptive fusion module VAFM is used for learning probability distribution of multi-modal features through variational reasoning and dynamically adjusting a fusion weight; a Riemannian manifold migration classifier RMTC is used; and the electrode positioning and function mapping module is used for outputting an electrode position and a functional area heat map in combination with an optimized objective function and 3D visualization. Through the dynamic collaborative segmentation network, the variational adaptive fusion module and the Riemannian manifold migration classifier, the precision and efficiency of electrode positioning and functional area mapping are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of surgical navigation, and particularly to an intelligent intraoperative cerebral cortex electrode positioning system based on deep learning. Background Art

[0002] In neurosurgery, such as epilepsy focus resection or brain tumor resection, the accurate positioning of intraoperative cerebral cortex electrodes and functional area mapping are key links determining the success of the surgery. Traditional positioning methods mainly rely on the combination of preoperative images (such as MRI) and intraoperative doctor experience, and identify functional areas by manually placing electrodes and recording electroencephalogram (EEG) signals. However, this method has significant limitations: preoperative images cannot fully reflect the real-time changes of intraoperative brain tissue, and single-modal data such as intraoperative ultrasound and MRI are difficult to comprehensively capture complex anatomical and functional information. Moreover, manual operation not only takes a long time, but is also easily affected by subjective judgment, resulting in insufficient positioning accuracy and low surgical efficiency.

[0003] With the development of multi-modal imaging technology and deep learning, it has become possible to fuse multi-source data such as intraoperative ultrasound, MRI, and EEG, and use intelligent algorithms to achieve automatic electrode positioning and functional area mapping.

[0004] Based on this demand, this technical solution proposes an innovative intelligent positioning system, aiming to significantly improve the accuracy, visualization, and operation efficiency of intraoperative electrode positioning through collaborative processing of multi-modal data and deep learning drive, providing more reliable technical support for neurosurgery. Summary of the Invention

[0005] The object of the present invention is to address the problems in the background art and propose an intelligent intraoperative cerebral cortex electrode positioning system based on deep learning, which significantly improves the accuracy and efficiency of electrode positioning and functional area mapping through a dynamic collaborative segmentation network, a variational adaptive fusion module, and a Riemannian manifold transfer classifier.

[0006] The technical solution of the present invention: An intelligent intraoperative cerebral cortex electrode positioning system based on deep learning, comprising A data acquisition module for acquiring intraoperative ultrasound, intraoperative MRI, and preoperative MRI data; A preprocessing module for denoising, contrast enhancement, motion artifact correction, bias field correction, skull stripping, and normalization processing of multi-modal data; A dynamic collaborative segmentation network DSCN that fuses MRI and CT features based on the principle of maximizing mutual information to generate a segmentation Mask; A variational adaptive fusion module VAFM that dynamically adjusts the fusion weight by learning the probability distribution of multi-modal features through variational inference; Riemannian manifold transfer classifier RMTC aligns the feature distributions of the source domain and the target domain on the Riemannian manifold to achieve functional area classification; Electrode positioning and functional mapping module, combining the optimization objective function and 3D visualization, outputs the electrode positions and the heat map of the functional area.

[0007] Preferably, for preoperative ultrasound data acquisition, a high-frequency linear array ultrasound probe is used to acquire intraoperative cerebral cortex images; The preprocessing module preprocesses the data. First, the wavelet transform threshold denoising method is used for only denoising processing, then histogram equalization is applied to map the image gray values to the interval [0, 255] for contrast enhancement, and finally the Horn-Schunck optical flow method is used to estimate the motion vectors for only motion artifact correction.

[0008] Preferably, for intraoperative MRI data acquisition, a 3T intraoperative MRI device is used to acquire T1-weighted images T1weighted; The preprocessing module preprocesses the data. The N4ITK algorithm is used to eliminate magnetic field inhomogeneity through iterative optimization for bias field correction. The BET tool is used to strip the skull of the brain tissue based on morphological operations and active contour segmentation, and the image gray values are linearly mapped to the interval [0, 1] for normalization processing.

[0009] Preferably, for preoperative MRI data acquisition, a 3T MRI device is used preoperatively to acquire high-resolution T1-weighted images for constructing a 3D brain model; The preprocessing module preprocesses the data. A 6th-order Butterworth filter is used to filter out low-frequency drift and high-frequency noise for band-pass filtering, a notch filter is used to remove power frequency interference, and independent component analysis ICA is used to decompose the signal into independent components for artifact removal processing.

[0010] Preferably, the dynamic cooperative segmentation network DSCN includes a dual-path encoder structure that processes MRI and CT images respectively and extracts features step by step; Mutual information optimization module MIOM, calculates the mutual information of features between modalities and generates a cooperative weight matrix; Decoder, generates the segmentation result by combining the residual correction term.

[0011] Preferably, the variational adaptive fusion module VAFM includes, Variational autoencoder VAE, maps multi-modal features to the latent space and learns the mean and variance parameters; Dynamic weight function, calculates the contribution of each modality based on KL divergence and fuses the feature representation; Evidence lower bound ELBO optimization objective, balances the reconstruction error and distribution regularization.

[0012] Preferably, the Riemannian manifold transfer classifier RMTC includes Covariance matrix representation, which converts the feature distribution into a symmetric positive definite matrix; Riemannian manifold distance metric, which calculates the feature difference through matrix logarithm and Frobenius norm; Alignment loss function, which combines the kernel SVM classifier to optimize the geometric consistency between domains.

[0013] Preferably, the classifier uses kernel SVM; the Riemannian manifold method makes full use of the geometric structure of the features and improves the classification accuracy.

[0014] Preferably, the electrode localization and functional mapping module includes Rigid body transformation registration, which aligns the preoperative and intraoperative spaces through an optical tracking system; Optimization objective function, which balances the EEG signal matching and anatomical constraints.

[0015] Preferably, the electrode localization and functional mapping module further includes 3D visualization interface, which superimposes the functional area heat map and electrode positions, and updates the electrode positions through gradient descent . The functional area mapping will fuse the features Input into RMTC, output the classification probability, and generate a heat map superimposed on the 3D brain model.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: In the present invention, the Riemannian manifold method makes full use of the geometric structure of the features and improves the classification accuracy. The classification module outputs the functional area labels of the electrodes and the optimized localization coordinates, and transmits them to the localization and mapping module.

[0017] This method significantly improves the accuracy and efficiency of electrode localization and functional area mapping through a dynamic cooperative segmentation network, a variational adaptive fusion module, and a Riemannian manifold transfer classifier. The electrode contact localization error will be greatly reduced, the EEG signal analysis response time will be shortened, an intuitive relationship between the brain structure and the electrodes will be provided, facilitating the replication of the surgery, and the intraoperative brain mapping time will be shortened.

[0018] The above description is only an overview of the technical solution of the present invention. In order to be able to understand the technical means of the present invention more clearly and implement it in accordance with the content of the description, the following takes the preferred embodiments of the present invention and combines the accompanying drawings to describe in detail as follows. The specific implementation manners of the present invention are given in detail by the following embodiments and their accompanying drawings. Description of the Drawings

[0019] The drawings described herein are used to provide a further understanding of the present invention, form a part of this application, and the schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings: Figure 1 It is a schematic structural diagram of an embodiment in the present invention. Specific embodiments

[0020] The principles and features of the present invention will be described below in conjunction with the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention. In the following paragraphs, the present invention will be described more specifically by way of example with reference to the accompanying drawings. It should be noted that the accompanying drawings are all in very simplified forms and use non-precise scales, only for the purpose of facilitating and clearly assisting in explaining the objectives of the embodiments of the present invention.

[0021] It should be noted that when a component is referred to as "fixed to" another component, it can be directly on the other component or there can also be an intermediate component. When a component is considered to be "connected" to another component, it can be directly connected to the other component or there may be an intermediate component at the same time. When a component is considered to be "disposed on" another component, it can be directly disposed on the other component or there may be an intermediate component at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are only for the purpose of illustration.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs. The terms used in the description of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0023] Embodiment 1 As Figure 1 shown, an intraoperative cerebral cortex electrode intelligent positioning system based on deep learning proposed by the present invention includes a data acquisition module for acquiring intraoperative ultrasound, intraoperative MRI, and preoperative MRI data; a preprocessing module for denoising, contrast enhancement, motion artifact correction, bias field correction, skull stripping, and normalization processing of multi-modal data; a dynamic cooperative segmentation network DSCN for generating a segmentation Mask by fusing MRI and CT features based on the principle of maximizing mutual information; a variational adaptive fusion module VAFM for dynamically adjusting the fusion weight by learning the probability distribution of multi-modal features through variational inference; a Riemannian manifold transfer classifier RMTC for aligning the feature distributions of the source domain and the target domain on the Riemannian manifold to achieve functional area classification; an electrode positioning and functional mapping module for outputting the electrode position and the functional area heat map by combining an optimization objective function and 3D visualization.

[0024] The Riemannian manifold transfer classifier RMTC includes Covariance matrix representation, which converts the feature distribution into a symmetric positive definite matrix; Riemannian manifold distance metric, which calculates the feature differences through matrix logarithm and Frobenius norm; Alignment loss function, which combines the kernel SVM classifier to optimize the geometric consistency between domains.

[0025] The classifier uses kernel SVM; the Riemannian manifold method makes full use of the geometric structure of the features and improves the classification accuracy.

[0026] In this embodiment, first, a segmentation network is pre-trained on a public dataset (such as BraTS) to extract general features

[0027] : The fused feature matrix of the target domain, with dimension , and each column is a feature vector of a sample. The transpose matrix of, with dimension . : Matrix multiplication, which calculates the inner product between features, and the result is A matrix representing the correlation between features. : Normalization factor, divided by the number of samples to make the covariance matrix reflect the average statistical characteristics of the features. : The output covariance matrix, with dimension , which is a symmetric positive definite matrix describing the distribution shape and direction of the target domain features.

[0028]

[0029]

[0030] : Respectively represent two covariance matrices (such as and ), both of which are symmetric positive definite matrices. The inverse of the matrix square root of, and the calculation method is based on eigenvalue decomposition , then , which is used to normalize the scale of. : Matrix similarity transformation, which projects onto the space with as the reference, and the result is still a symmetric matrix, reflecting the relative geometric relationship between the two. Matrix logarithm operation, which calculates , where is a diagonal eigenvalue matrix, takes the natural logarithm of each eigenvalue to capture the non - linear differences between matrices. : Frobenius norm, defined as the square root of the sum of the squares of the matrix elements , which measures the overall distance between matrices. : The output is the geodesic distance between two covariance matrices on the Riemannian manifold, reflecting the geometric differences in the feature distributions.

[0031] To align the source domain and the target domain, an alignment loss is designed:

[0032] The number of sample pairs in a batch. Assume that both the source domain and the target domain have covariance matrices. : The covariance matrix of the th sample in the source domain. : The covariance matrix of the th sample in the target domain. : The projection operator on the Riemannian manifold that projects onto the sub - manifold consistent with , usually achieved by optimizing the geodesic path to ensure that the target - domain features adapt to the source - domain geometric structure. : The aforementioned Riemannian distance, which calculates the difference between the covariance matrices before and after projection. : The averaging operation, which calculates the mean of the distances of all sample pairs to obtain the overall alignment loss. : The alignment loss. The optimization goal is to reduce the geometric differences in the feature distributions between the source domain and the target domain and improve the transfer ability.

[0033] The classifier uses kernel SVM, and the kernel function is .

[0034] : The aforementioned Riemannian distance, which measures the geometric difference between two covariance matrices.

[0035] : The square of the distance, which amplifies the impact of the difference. The bandwidth parameter of the kernel function, which controls the decay rate of similarity and can be adjusted by cross - validation. : The exponential function, which converts the distance into a similarity measure. The closer the value is to 1, the more similar it is. The output of the kernel function, representing the and similarity, which is used for SVM classification.

[0036] The total loss is:

[0037] is the classification loss, calculated based on kernel SVM, is a hyperparameter that balances the importance of classification accuracy and domain alignment. is the aforementioned alignment loss. is the total loss.

[0038] The Riemannian manifold method fully utilizes the geometric structure of features and improves classification accuracy. The classification module outputs the functional area labels of the electrodes and the optimized positioning coordinates, which are transmitted to the positioning and mapping module.

[0039] This method significantly improves the accuracy and efficiency of electrode positioning and functional area mapping through a dynamic cooperative segmentation network, a variational adaptive fusion module, and a Riemannian manifold transfer classifier. The positioning error of electrode contacts will be greatly reduced, the response time of EEG signal analysis will be shortened, providing an intuitive relationship between brain structure and electrodes, facilitating the replication of surgeries, and shortening the intraoperative brain mapping time.

[0040] Example Two As Figure 1 shown, an intraoperative cerebral cortex electrode intelligent positioning system based on deep learning proposed by the present invention. Compared with Example One, in this example, for preoperative ultrasound data acquisition, a high-frequency linear array ultrasound probe is used to acquire intraoperative cerebral cortex images; a high-frequency linear array ultrasound probe with a frequency range of is used to acquire intraoperative cerebral cortex images. The spatial resolution of the images is , and the temporal resolution is 2030 frames per second (FPS). The acquisition area covers the surgically exposed cerebral cortex surface.

[0041] The preprocessing module preprocesses the data. First, only denoising processing is performed using the wavelet transform threshold denoising method, then histogram equalization is applied to map the image gray values to the interval [0, 255] for contrast enhancement, and finally, the Horn-Schunck optical flow method is used to estimate the motion vectors for only motion artifact correction.

[0042] The wavelet transform threshold denoising method is adopted. The wavelet basis is selected as Daubechies4 (db4), the decomposition level is 4, and the threshold calculation formula is:

[0043] where, is the total number of image pixels, is the noise standard deviation, estimated by the median absolute deviation (MAD):

[0044] is the wavelet coefficient, and respectively represent the decomposition scale and position.

[0045] Contrast enhancement: Apply histogram equalization to map the image gray values to the interval. The specific operation is as follows: Calculate the gray histogram , where is the gray value.

[0046] Calculate the cumulative distribution function (CDF):

[0047] Normalize and map the new gray values:

[0048] Motion artifact correction: Use the Horn Schunck optical flow method to estimate the motion vector. Assuming that the image brightness is constant, the optimization objective is:

[0049] where, is the partial derivative of the image in the direction and time, is the motion vector component, is the smoothing weight. The motion field is obtained by iterative solution to correct the image.

[0050] During intraoperative MRI data acquisition, a 3T intraoperative MRI device is used to acquire T1-weighted images T1weighted; the voxel resolution is , the scanning time is controlled within 5 - 10 minutes, and the field of view (FOV) is . The preprocessing module preprocesses the data, uses the N4ITK algorithm to eliminate magnetic field inhomogeneity through iterative optimization, performs bias field correction, uses the BET tool to segment the brain tissue based on morphological operations and active contours, performs skull stripping, and linearly maps the image gray values to the [0, 1] interval for normalization.

[0051] Preprocessing steps: Bias field correction: Use the N4ITK algorithm to eliminate magnetic field inhomogeneity through iterative optimization. The objective function is:

[0052] where, is the original image, is the bias field, is the corrected image, is the regularization parameter. Skull stripping: Use the BET (Brain Extraction Tool) tool to segment the brain tissue based on morphological operations and active contours. The threshold is set to 0.3 times the image mean, and the number of iterations is 50. Normalization: Linearly map the image gray values to the range of [0, 1 Interval:

[0053] where and are the minimum and maximum gray values of the image Preoperative MRI data acquisition. Before surgery, use a 3T MRI device to acquire high-resolution T1-weighted images for constructing a 3D brain model; the resolution is , and the field of view is , for constructing a 3D brain model; Preprocessing steps: Consistent with the intraoperative MRI, perform bias field correction, skull stripping, and normalization in sequence, with the same parameter settings.

[0054] The preprocessing module preprocesses the data. Use a 6th-order Butterworth filter to perform band-pass filtering to remove low-frequency drift and high-frequency noise, use a notch filter to remove power frequency interference, and use independent component analysis ICA to decompose the signal into independent components for artifact removal processing.

[0055] In this embodiment, data acquisition: Use a 128-channel EEG device with a sampling rate of 1000 Hz. The electrodes are placed according to the 10-20 system to record intraoperative brain electrical activities for the entire duration of the surgery. Preprocessing steps: Band-pass filtering: Use a 6th-order Butterworth filter with a passband range of , to remove low-frequency drift and high-frequency noise. Power frequency interference removal: Use a 50 Hz (or 60 Hz, depending on the region) notch filter with a bandwidth of 2 Hz and an attenuation coefficient of 40 dB. Artifact removal: Use independent component analysis (ICA) to decompose the signal into independent components:

[0056] where is the observed signal matrix, is the mixing matrix, is the source signal matrix. Estimate and through the FastICA algorithm, and reconstruct the signal after removing the eye movement and electromyogram artifact components.

[0057] Example 3 As Figure 1As shown in the figure, an intraoperative cerebral cortex electrode intelligent positioning system based on deep learning proposed by the present invention. Compared with Embodiment 1 or Embodiment 2, in this embodiment, the dynamic cooperative segmentation network DSCN includes a dual-path encoder structure, which processes MRI and CT images respectively and extracts features step by step; The mutual information optimization module MIOM calculates the mutual information between features of different modalities and generates a cooperative weight matrix; The decoder generates a segmentation result by combining a residual correction term.

[0058] In this embodiment, the network adopts a dual-path encoder structure to process MRI and CT images respectively. Each path contains 5 convolutional blocks (convolution kernel , stride 1) and a max-pooling layer (pooling window ), and extracts features step by step. To achieve modality cooperation, through the mutual information optimization module Mutual Information Optimization Module, MIOM, the mutual information between the MRI feature map and the CT feature map is calculated at each layer of the encoder:

[0059] where represents the information entropy, and the calculation method is ; the joint entropy is obtained by estimating the joint probability distribution of the features. To reduce the computational complexity, a variational approximation method is used to approximate the mutual information, and an auxiliary distribution is introduced:

[0060] By maximizing the mutual information, the network learns a cooperative weight matrix :

[0061] where is obtained by gradient descent optimization, and the loss function is:

[0062] The fused features are input into the decoder, and after upsampling (bilinear interpolation) and convolution operations, a segmentation Mask is generated. To further improve the accuracy, a residual correction term is introduced at the end of the decoder (where is the input feature, is the preliminary prediction), and the final output is ( is a learnable parameter). The total loss of the network is:

[0063] Among them and are the real and predicted segmentation maps respectively. Based on the collaborative mechanism of mutual trust, the segmentation accuracy of complex brain structures such as gyri and sulci has been significantly improved.

[0064] Example 4 As Figure 1 shown, for an intraoperative cerebral cortex electrode intelligent positioning system based on deep learning proposed by the present invention, compared with Example 1 or Example 2 or Example 3, in this embodiment, the variational adaptive fusion module VAFM includes, a variational autoencoder VAE that maps multi-modal features to the latent space and learns the mean and variance parameters; a dynamic weight function that calculates the contribution of each modality based on the KL divergence and fuses the feature representations; an evidence lower bound ELBO optimization objective that balances the reconstruction error and the distribution regularization.

[0065] In this embodiment, first, the image features (extracted from the segmentation network), EEG frequency domain features (obtained through the Morlet wavelet transform, and the transform formula ) and behavioral features (in the form of vector encoding) are preprocessed to unify the dimensions to . Assume that each modality feature follows a multivariate Gaussian distribution , and the latent representation is inferred through a variational autoencoder (VAE):

[0066] represents the mean of the latent representation calculated through a two-layer fully connected neural network , indicating the central tendency of the features. represents the log variance of the latent representation calculated through another two-layer fully connected network , controlling the dispersion degree of the distribution. These two networks learn key information from and generate the compressed latent representation .

[0067] To fuse the multi-modal latent representations, a dynamic weight function is designed:

[0068] where KL is the Kullback Leibler divergence, and the calculation formula is:

[0069] Denotes the trace operation and measures the similarity between two covariance matrices. Denotes the Mahalanobis distance of the mean difference, representing the offset between means. Denotes the dimension correction term, which is the feature dimension. Denotes the ratio of the determinants of the covariance matrices, reflecting the difference in distribution shapes. The formula quantifies the difference between two distributions and is used to dynamically adjust the importance of modalities.

[0070] The fused feature is:

[0071] The core of the fusion process is obtained by optimizing through minimizing the evidence lower bound (ELBO), and ELBO is defined as:

[0072] : The reconstruction term, representing the likelihood of reconstructing the input feature from the latent representation to ensure that the original information is retained. : The regularization term, which restricts the variational distribution to be close to the prior distribution to avoid overfitting.

[0073] By optimizing this objective, the model can not only effectively compress and fuse features, but also maintain the integrity of information and the stability of the distribution, thus obtaining the optimal one. The variational inference method adaptively adjusts the contributions of each modality through probability distributions to ensure that the fusion result has the optimal representation ability for the task. The final fused feature map is transmitted to the classification module.

[0074] Example Five As Figure 1 shown, for an intraoperative cerebral cortex electrode intelligent positioning system based on deep learning proposed by the present invention, compared with Example One or Example Two or Example Three or Example Four, in this example, the electrode positioning and functional mapping module includes Rigid body transformation registration to align the preoperative and intraoperative spaces through an optical tracking system; Optimizing the objective function to balance EEG signal matching and anatomical constraints.

[0075] A 3D visualization interface that superimposes the functional area heat map and electrode positions and updates the electrode positions through gradient descent . The functional area mapping inputs the fused feature into the RMTC, outputs the classification probability, and generates a heat map superimposed on the 3D brain model.

[0076] In this embodiment, during the electrode positioning and functional area mapping phase, first, through rigid body transformation ( is the rotation matrix, is the translation vector) to register the preoperative image space with the intraoperative space, and the optical tracking system is used to calculate the transformation parameters. Subsequently, the segmentation network outputs the brain structure Mask, and combines the electrode planning to predict the initial position . According to the intraoperative EEG and ECS data, the optimization objective is defined:

[0077] The current electrode position vector, with a dimension of , represents the coordinates to be adjusted during the optimization process . : The EEG signal feature vector simulated or measured at the position , which may include frequency domain power, phase, etc., and the dimension depends on the feature extraction method (such as ). It reflects the neural activity recorded by the electrode at the position. : The target EEG signal feature vector, with the same dimension as , is usually obtained from intraoperative ECS data or predefined functional area features, and represents the desired neural activity pattern. : The square of the L2 norm, which calculates the square of the Euclidean distance between the simulated EEG signal and the target signal, that is . This term measures the matching degree between the electrode position and the target functional area, and the smaller the value, the closer it is to the ideal position.

[0078] : The position deviation vector, with a dimension of , represents the offset of the current position relative to the initial position . The square of the L2 norm, that is , measures the degree of deviation of the current position from the initial plan, and prevents the optimization from deviating too much from the anatomical constraints. Weigh the weights of the two terms. A larger makes closer to , and a smaller emphasizes the EEG signal matching more. : The total objective function, a scalar, and the optimization aim is to find the that minimizes , that is, to achieve the best balance between signal matching and position constraints : The regularization coefficient, a scalar set to 0.1, is used for balancing.

[0079] Update the electrode position through gradient descent The functional area mapping will fuse features Input RMTC, output classification probability, generate a heat map and overlay it on the 3D brain model. This module obtains the optimized electrode position P and the functional area mapping result. Calculate the positioning error Error during the verification phase and classification metrics such as accuracy, and provide an interactive interface for doctors to adjust

[0080] As mentioned above, it is only the preferred embodiment of the present invention, and there is no any form of limitation to the present invention; any ordinary technician in the industry can smoothly implement the present invention according to the illustrations in the specification and the above description; however, any slight changes, modifications and evolutions made by those skilled in the art within the scope of the technical solution of the present invention by using the technical content disclosed above are equivalent embodiments of the present invention; at the same time, any changes, modifications and evolutions made to the above embodiments based on the substantial technology of the present invention are still within the protection scope of the technical solution of the present invention

Claims

1. An intelligent intraoperative cerebral cortex electrode positioning system based on deep learning, characterized in that: including, a data acquisition module for acquiring intraoperative ultrasound, intraoperative MRI, and preoperative MRI data; a preprocessing module for denoising, contrast enhancement, motion artifact correction, bias field correction, skull stripping, and normalization of multimodal data; a dynamic collaborative segmentation network DSCN that fuses MRI and CT features based on the principle of maximizing mutual information to generate a segmentation mask; a variational adaptive fusion module VAFM that learns the probability distribution of multimodal features through variational inference and dynamically adjusts the fusion weights; a Riemannian manifold transfer classifier RMTC that aligns the feature distributions of the source domain and the target domain on the Riemannian manifold to achieve functional area classification; an electrode positioning and functional mapping module that combines an optimization objective function and 3D visualization to output the electrode positions and functional area heat maps.

2. The intraoperative cerebral cortex electrode intelligent positioning system based on deep learning according to claim 1, wherein Preoperative ultrasound data acquisition, using a high-frequency linear array ultrasound probe to acquire intraoperative cerebral cortex images; The preprocessing module preprocesses the data. First, it uses the wavelet transform threshold denoising method for only denoising, then applies histogram equalization to map the image gray values to the interval [0, 255] for contrast enhancement, and finally uses the Horn Schunck optical flow method to estimate the motion vectors for only motion artifact correction.

3. The intraoperative cerebral cortex electrode intelligent positioning system based on deep learning according to claim 2, characterized in that Intraoperative MRI data acquisition, using a 3T intraoperative MRI device to acquire T1-weighted images T1weighted; The preprocessing module preprocesses the data. It uses the N4ITK algorithm to eliminate magnetic field inhomogeneity through iterative optimization for bias field correction, uses the BET tool to segment the brain tissue based on morphological operations and active contours for skull stripping, and linearly maps the image gray values to the interval [0, 1] for normalization.

4. The intraoperative cerebral cortex electrode intelligent positioning system based on deep learning according to claim 3, characterized in that, Preoperative MRI data acquisition, using a 3T MRI device preoperatively to acquire high-resolution T1-weighted images for constructing a 3D brain model; The preprocessing module preprocesses the data. It uses a 6th-order Butterworth filter to filter out low-frequency drift and high-frequency noise for band-pass filtering, uses a notch filter to remove power frequency interference, and uses independent component analysis ICA to decompose the signal into independent components for artifact removal.

5. The intraoperative cerebral cortex electrode intelligent positioning system based on deep learning according to claim 4, characterized in that, The dynamic collaborative segmentation network DSCN includes a dual-path encoder structure that processes MRI and CT images separately and extracts features step by step; a mutual information optimization module MIOM that calculates the mutual information between modal features and generates a collaborative weight matrix; a decoder that generates a segmentation result in combination with a residual correction term.

6. The intraoperative cerebral cortex electrode intelligent positioning system based on deep learning according to claim 5, wherein The variational adaptive fusion module VAFM includes a variational autoencoder VAE that maps multimodal features to the latent space and learns the mean and variance parameters; a dynamic weight function that calculates the contribution of each modality based on the KL divergence and fuses the feature representations; an evidence lower bound ELBO optimization objective that balances the reconstruction error and distribution regularization.

7. An intraoperative cerebral cortex electrode intelligent positioning system based on deep learning according to claim 6, characterized in that, The Riemannian manifold transfer classifier RMTC includes a covariance matrix representation that converts the feature distribution into a symmetric positive definite matrix; a Riemannian manifold distance metric that calculates the feature difference through matrix logarithm and Frobenius norm; an alignment loss function that combines a kernel SVM classifier to optimize the geometric consistency between domains.

8. An intraoperative cerebral cortex electrode intelligent positioning system based on deep learning according to claim 7, characterized in that, The classifier uses kernel SVM; the Riemannian manifold method makes full use of the geometric structure of features and improves the classification accuracy.

9. The intraoperative cerebral cortex electrode intelligent positioning system based on deep learning according to claim 8, characterized in that, The electrode localization and functional mapping module includes rigid body transformation registration to align the pre-operative and intra-operative spaces through an optical tracking system; optimizing the objective function to balance EEG signal matching and anatomical constraints.

10. The intraoperative cerebral cortex electrode intelligent positioning system based on deep learning according to claim 9, characterized in that, The electrode localization and functional mapping module also includes 3D visualization interface, overlaying the functional area heatmap with electrode positions, updating electrode positions through gradient descent Functional area mapping will fuse features Input RMTC, output classification probability, generate a heatmap and overlay it on the 3D brain model.

Images