An individual structure brain network-based epilepsy postoperative curative effect prediction system

Abstract

The application discloses an individual structure brain network-based epilepsy postoperative curative effect prediction system and belongs to the technical field of biomedical image pattern recognition. The system is based on preoperative structural magnetic resonance imaging data of healthy subjects and epilepsy patients. Firstly, a standard structural covariant network model is constructed in the healthy subject group. Then, the preoperative structural data of the patients is added to the standard model one by one to construct an epilepsy individual disturbance offset network, and the network is standardized. Meanwhile, by adding virtual surgical resection information, the structural network characteristics of the reserved brain area after surgery are calculated and input into a support vector machine classification framework to establish a classification model related to the surgical curative effect. The system quantifies the individual structural image characteristics of the epilepsy patients by constructing the structural covariant network, and realizes intelligent prediction of the curative effect of epilepsy surgery by using the characteristics of the network, so that more comprehensive brain image information reference can be provided for preoperative accurate evaluation.

Description

Technical Field

[0001] This method belongs to the field of biomedical image pattern recognition technology, specifically involving structural magnetic resonance brain imaging data processing, individual structural covariance network construction, and methods for predicting the efficacy of surgical treatment for epilepsy. Background Technology

[0002] Epilepsy is one of the most common neurological disorders worldwide, affecting approximately 70 million people globally, with about 2 million new cases diagnosed annually. Antiepileptic drugs are the first-line treatment for epilepsy, but about 30% of cases are drug-resistant. Surgical resection is a crucial component of epilepsy treatment; at least 50% of patients with drug-resistant epilepsy are suitable candidates for surgery. However, a significant proportion of patients experience persistent epilepsy after surgery. Therefore, individual preoperative assessment of surgical efficacy is essential for precision treatment. Despite advancements in multidisciplinary preoperative assessment, traditional epilepsy neurosurgery urgently needs innovation in new technologies and methods to drive the development of precision diagnosis and treatment for epilepsy. Summary of the Invention

[0003] This invention addresses the clinical needs highlighted in the technical background by proposing an intelligent postoperative efficacy assessment system based on the preoperative individual brain network of epilepsy patients. First, a structural covariant network of epilepsy patients at the individual level is constructed using structural magnetic resonance imaging (MRI). This network is then used as an indicator to explore the relationship between preoperative structural brain network features and surgical efficacy. Furthermore, a postoperative efficacy classification model for epilepsy patients is constructed based on machine learning methods, achieving intelligent preoperative assessment.

[0004] The technical solution of this invention is: a system for predicting the efficacy of epilepsy surgery based on individual structural brain networks. The system includes five modules: a magnetic resonance structural image acquisition and surgical prognosis assessment module, a structural image data preprocessing module, a standard surgical resection template construction module, an individual structural covariant network construction module, and a feature screening and surgical efficacy classification prediction module.

[0005] The magnetic resonance imaging acquisition and surgical prognosis module: This module is used to save and record the preoperative magnetic resonance imaging and postoperative recovery of epilepsy patients who have undergone the same type of surgical treatment.

[0006] The processing method of the structural image data preprocessing module is as follows:

[0007] Step A1: Manually adjust the front and back joins so that they are on the same horizontal line;

[0008] Step A2: Use the Cat12 software package to remove skull information from the image, and then use prior tissue probability maps and Naive Bayes algorithm to identify different tissue types;

[0009] Step A3: Register the gray and white matter images into the standard MNI (Montreal Neurological Institute) space using affine and nonlinear transformations;

[0010] Step A4: Use the Jacobian determinant to preserve information in the original spatial location; use Cat12 software to segment the image, score its quality, and select images with a quality index greater than a set threshold for subsequent analysis;

[0011] Step A5: Perform isotropic Gaussian smoothing on the grayscale image to eliminate noise and artifacts;

[0012] The module for constructing a standard surgical resection template:

[0013] Step B1: Subtract the postoperative image from the preprocessed preoperative image, and then manually adjust the threshold to obtain the surgical area of ​​each patient in the MNI space, represented by a three-dimensional matrix as MS. m m represents the patient's ID number;

[0014] Step B2: Overlap the surgical areas of each patient to obtain the resectable region MS at the group level, specifically in the following manner:

[0015] MS = MS1 ∪ MS2 ∪ MS3 … MS m

[0016] Step B3: Based on previous whole-brain templates, the resected region is mapped onto the template using the MS method. Brain regions with an overlap rate of more than 50% between the resected region and the template region are selected as standard resection templates.

[0017] The processing method of the individual structural covariant network construction module includes: extracting morphological features of brain regions, constructing a structural covariant network, and standardizing the structural covariant network;

[0018] Morphological information extraction of the brain region:

[0019] The brain is divided into several local brain regions using a whole-brain template, and the average gray matter volume within each region is then extracted as a local morphological representation.

[0020] The specific method for constructing the structural covariant network is as follows:

[0021] Step C1: Extract morphological information of normal subjects' brain regions to construct a structural covariance network M. Each element in M ​​is the Pearson correlation coefficient between brain regions. M is used as the reference network, and the calculation formula is as follows:

[0022]

[0023] Among them, M ijLet X be the value in the i-th row and j-th column of M, (X i, Y j ), i,j=1,2…,n are the sampled data of the i-th random variable X and the j-th random variable Y. and It is the average of random variables X and Y, (X , Y) specifically represents the gray matter volume (GMV) of a brain region;

[0024] Step C2: Resample all samples, removing one normal subject in each sample and repeatedly constructing the structural covariant network M from the remaining normal subjects, estimating the distribution of the standard structural covariant network variability:

[0025] [|M1-M|,|M2-M|…,|M Q -M|]

[0026] Where Q represents the total number of normal subjects;

[0027] The method for standardizing the structural covariant network is as follows:

[0028] Step D1: For each epilepsy patient, their preoperative structural magnetic resonance imaging (MRI) images are added to the normal subject population as a perturbation factor for the reference covariant network; the data of the i-th patient are added to the normal subject population to obtain a new structural covariant network M. * Furthermore, the individual offset matrix D = M of the patient can be obtained. * -M;

[0029] Step D2: Calculate the score Z for the individual patient's offset matrix. i The standardization calculation formula is as follows:

[0030]

[0031] median(x) is the median of all sample observations, and MAD is the absolute deviation of the median, defined as follows:

[0032] MAD = median|x i -median(x i )|

[0033] Compared to standard deviation, which is more sensitive to outliers, MAD has better robustness. To make MAD a consistent estimator for estimating standard deviation, a transformation of MAD is performed:

[0034] MAD = δ / k

[0035] Where k is a constant factor and δ is the sample standard deviation; Z0 for each patient i The standardized structural covariant network Z is obtained by combining the components.D ;

[0036] The calculation method for the feature screening and surgical efficacy classification prediction module is as follows:

[0037] Step D1: Use 10-fold cross-validation to divide the dataset into 10 folds, where 9 folds are the training set and 1 fold is the training set.

[0038] Step D2: Calculate the network matrix Z obtained above. D Calculate the features of each brain region and use the score F(i) to perform preliminary screening of the features. The calculation method of F(i) is as follows:

[0039]

[0040] Where i represents the i-th feature, and each feature has an F(i). F(i) is the average of all features, where (+) and (-) represent positive and negative samples, respectively, that is, the patient's epilepsy has not disappeared and the patient's epileptic seizures have completely disappeared. The larger F(i) is, the more discriminative this feature is for the classification task.

[0041] Step D3: Input the test set into the trained model to obtain the predicted postoperative efficacy.

[0042] This invention addresses the issue of surgical efficacy in epilepsy patients by designing a system framework that can quantify individual brain specificity. Furthermore, this system takes into account the patient's surgical information, enabling more accurate feature extraction and prediction of surgical efficacy. Attached Figure Description

[0043] Figure 1 This is an overall flowchart of the present invention;

[0044] Figure 2 This is the surgical resection template of the present invention;

[0045] Figure 3 The prediction results are shown in the diagram for 10-fold cross-validation.

[0046] Figure 4 The distribution of important features returned by the model. Implementation

[0047] First, preoperative 3D T1 magnetic resonance imaging (MRI) images of epilepsy patients undergoing the same type of surgical treatment (e.g., anterior temporal lobectomy) are collected clinically. Patients are followed up for one year post-surgery, and the surgical outcome is assessed using the Engel rating system based on their post-operative seizure history. An Engel rating of Ia indicates complete seizure-free (SF); higher Engel ratings (Ib and above) indicate poor surgical outcome (non-seizure-free (NSF), meaning seizures persist.

[0048] Then proceed with the following processing:

[0049] A. Data preprocessing:

[0050] Step A1: Manually adjust the anterior and posterior commissures to align them on the same horizontal line. Simultaneously, set the spatial origin at the anterior commissure and adjust the coronal and transverse views to achieve left-right symmetry about the vertical line.

[0051] Step A2: Use cat12 to remove skull information from the image, and then use prior tissue probability maps and Naive Bayes algorithm to identify different tissue types, segmenting the image into gray matter and white matter images;

[0052] Step A3: Register the gray and white matter images into the standard MNI space using affine and nonlinear transformations. The voxel size is 1.5mm × 1.5mm × 1.5mm, and the spatial dimensions are [121, 145, 121].

[0053] Step A4: In the image after image registration, the Jacobian determinant is used to preserve information at the original spatial location. This process preserves the volume information of voxels. The images are then segmented using Cat12 software, and quality scores are assigned. Samples with an IQR > 80 are selected for further analysis based on these scores.

[0054] Step A5: Perform isotropic Gaussian smoothing on the grayscale image to eliminate noise and artifacts. The full width at half maximum of the Gaussian smoothing kernel is [8mm 8mm 8mm].

[0055] B. Constructing a standard resection template

[0056] Step B1: Subtract the postoperative image from the preprocessed preoperative image, and then manually adjust the threshold to obtain the surgical area of ​​each patient in the MNI space. A three-dimensional matrix is ​​used to represent the non-MS area. m 'm' represents the patient's ID number, MS m The spatial dimensions are [121, 145, 121];

[0057] Step B2: MS = MS1∪MS2∪MS3…MSm This method is used to overlap the surgical areas of each patient to obtain the resection area MS at the group level;

[0058] Step B3: Based on a whole-brain template of 246 regions of interest, the resected regions are mapped onto the template using the MS method. Brain regions with an overlap of more than 50% between the resected region and the template region are selected as standard resection templates for subsequent analysis. This invention uses subjects who underwent temporal lobectomy as an example, identifying 21 brain regions on the 246 templates that were the surgically resected areas.

[0059] C. Construction of Individual Structural Covariant Networks

[0060] Step C1: Use the 246 template to extract the preprocessed gray matter file. The value for each brain region is the average gray matter volume of all voxels in that brain region. Each person will get a 246*1 vector in this step.

[0061] Step C2: Combine the vectors of the healthy control group into a 246*n1 dimensional matrix RM, where n1 is the number of healthy subjects. RM is used to calculate the reference network M.

[0062] Step C3: Resample the healthy control group, removing data from one subject at a time, to obtain the RM. n1-1 The dimension is 246*(n1-1), RM n1-1 The distribution of structural covariance network variability in healthy subjects was obtained by calculation:

[0063] [|M1-M|,|M2-M|…,|M n1 -M|]. The median M of the structural covariant network of healthy subjects was calculated accordingly. mid MAD (Absolute Deviation from Median):

[0064] MAD = median||M i -M|-M mid |

[0065] Step C4: Add the patient's morphological vector to the RM as a perturbation factor for the reference covariant network, thus obtaining the RM. n1+1 The matrix has a dimension of 246*(n1+1), and the correlation is used to obtain the structural covariant network M. * Furthermore, the individual structural covariance network D of the patient can be obtained: D=|M * -M|;

[0066] Step C5: Perform functional standardization on the patient's individual structural covariant network and calculate the corrected z-score of D: Where k is a constant factor, k = 1.4826.

[0067] D. Feature Filtering and Classification

[0068] Step D1: In order to conduct testing, this invention uses 10-fold cross-validation, dividing the dataset into 10 equal folds, where 9 folds are the training set and 1 fold is the validation set;

[0069] Step D2: Calculate the network matrix Z obtained above. D Perform a z-test, convert the z-score to a p-value, and select edges with p < 0.05 under FDR correction, setting the remaining edges to 0;

[0070] Step D3: Based on the resection template obtained above, calculate the sum of the edge strengths between each unresected brain region and the corresponding brain region to obtain the corresponding feature vector F;

[0071] Step D4: Calculate the f-score for each feature in the training set and take the f-score. Then, sort the features according to their f-scores, typically selecting the top 30 features with the highest f-scores. Select these features as the final feature set and use them as input to the SVM for training.

[0072] Step D5: Apply the feature set selected in D4 to the test set and use the trained model for classification.

[0073] Ten-fold cross-validation was performed on the test data, with an average accuracy of 81% and an AUC area of ​​approximately 0.80. The specificity was 88% and the sensitivity was 71%.

Claims

1. A system for predicting the efficacy of epilepsy surgery based on individual structural brain networks, comprising five modules: magnetic resonance structural image acquisition and surgical prognosis assessment module, structural image data preprocessing module, standard surgical resection template construction module, individual structural covariant network construction module, feature screening and surgical efficacy classification prediction module; The magnetic resonance imaging acquisition and surgical prognosis module: This module is used to save and record the preoperative magnetic resonance imaging and postoperative recovery of epilepsy patients who have undergone the same type of surgical treatment. The processing method of the structural image data preprocessing module is as follows: Step A1: Manually adjust the front and back joints so that they are on the same horizontal line; Step A2: Use the Cat12 software package to remove skull information from the image, and then use prior tissue probability maps and Naive Bayes algorithm to identify different tissue types; Step A3: Register the gray and white matter images into the standard MNI (Montreal Neurological Institute) space using affine and nonlinear transformations; Step A4: Use the Jacobian determinant to preserve information in the original spatial location; use Cat12 software to segment the image, score its quality, and select images with a quality index greater than a set threshold for subsequent analysis; Step A5: Perform isotropic Gaussian smoothing on the grayscale image to eliminate noise and artifacts; Construct a standard surgical resection template module: Step B1: Subtract the postoperative image from the preprocessed preoperative image, and then manually adjust the threshold to obtain the surgical area for each patient in the MNI space, represented by a three-dimensional matrix. m represents the patient's ID number; Step B2: Overlap the surgical areas of each patient to obtain the resection area at the group level. The specific method: ； Step B3: Based on previous whole-brain templates, the area to be removed... Mapping onto the template, brain regions with an overlap rate greater than 50% between the resection area and the template area are selected as standard resection templates; The processing method of the individual structural covariant network construction module includes: extracting morphological features of brain regions, constructing a structural covariant network, and standardizing the structural covariant network; Morphological information extraction of the brain region: The brain was divided into several local brain regions using a whole-brain template, and the average gray matter volume within each region was extracted as a local morphological representation. The specific method for constructing the structural covariant network is as follows: Step C1: Extract morphological information of normal subjects' brain regions to construct a structural covariance network M. Each element in M ​​is the Pearson correlation coefficient between brain regions. M is used as the reference network, and the calculation formula is as follows: ； in, For the first in M The value in row j, For the first random variables and the random variables The sampling data, and It is the average of random variables X and Y. Specifically, it represents the gray matter volume (GMV) of a brain region; Step C2: Resample all samples, removing one normal subject in each sample and repeatedly constructing the structural covariant network from the remaining normal subjects. Estimate the distribution of variability in standard structure covariant networks: ； Where Q represents the total number of normal subjects; The method for standardizing the structural covariant network is as follows: Step D1: For each epilepsy patient, their preoperative structural MRI images were added to the normal subject population as a perturbation factor for the reference covariant network; the first... Data from these patients were added to normal subjects to obtain new structural covariant networks. Further, the individual offset matrix of the patient was obtained. ; Step D2: Calculate the score for the offset matrix of the individual patient. The standardization calculation formula is as follows: ； The median is the median of all sample observations. The MAD is the absolute deviation of the median, defined as follows: ； Compared to standard deviation, which is more sensitive to outliers, To have better robustness, in order to As a consistent estimator for estimating standard deviation, Perform the transformation: ； Where k is a constant factor, It is the sample standard deviation; it represents the standard deviation of each patient. The standardized structural covariant network is obtained by combining the components. ; The calculation method for the feature screening and surgical efficacy classification prediction module is as follows: Step E1: Use 10-fold cross-validation to divide the dataset into 10 folds, where 9 folds are the training set and 1 fold is the training set. Step E2: The network matrix obtained above... Calculate the characteristics of each brain region and use the scores. Initial screening of features. Calculation method: ； Where i represents the i-th feature, and each feature has a set of features. , It is the average of all characteristics, where (+) and (-) represent positive and negative samples, respectively, that is, the patient's epilepsy has not disappeared and the patient's epileptic seizures have completely disappeared. The larger the value, the more discriminative the feature is for classification tasks; Step E3: Input the test set into the trained model to obtain the predicted postoperative efficacy.

Images