Method and System for Selecting Light Stimulation Parameters for Patients Based on Brain Functional Networks

Abstract

The present invention provides a method and system for selecting patient light stimulation parameters based on a brain functional network. The method involves acquiring and preprocessing brain diffusion tensor modal data and resting-state functional modal data from patients with cerebral small vessel disease (SVD). Based on the preprocessed resting-state functional modal data, statistical analysis is performed to obtain multiple different global network attributes and multiple edges with different connection weights. A convolutional neural network is then used to segment the hippocampus region of the preprocessed brain diffusion tensor modal data, quantifying the information contained in the image to obtain multiple radiomic features. The acquired global network attributes and edges with different connection weights are then fused with the radiomic features to form a feature matrix. Multiple basic classifiers are constructed and trained using ensemble learning, and finally a vector is output. The present invention automates the selection of CSVD light stimulation parameters, overcoming the many limitations of complex scoring methods used clinically.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and more particularly to a method and system for selecting light stimulation parameters for patients based on a brain functional network. In particular, the present invention preferably relates to a method for selecting light stimulation parameters for patients with CSVD based on a brain functional network. Background Art

[0002] Cerebral small vessel disease (CSVD)[1] is a disease that occurs in the small perforating arteries, capillaries, and venules of the brain. It can cause small subcortical infarcts, white matter hyperintensities, lacunar lesions, cerebral microbleeds, enlarged perivascular spaces, and brain atrophy. These lesions can lead to cognitive impairment, Alzheimer's disease, mood disorders, and other diseases[2]. PM10 is an aerodynamic particle with a diameter of less than or equal to 10 microns, also known as inhalable particulate matter or floating dust. Long-term exposure to industrial dust can cause tiny dust particles to enter the lungs, nose, and eyes, bypass the blood-brain barrier through breathing, olfactory nerves, and blood, and spread directly to deep brain regions through intersynaptic transmission. Long-term exposure to dust environments has a significant impact on abnormal brain functional networks and the resulting CSVD[3].

[0003] In recent years, near-infrared light therapy has been used both domestically and internationally to alleviate the progression of CSVD and improve cognitive function. However, due to the lack of clarity regarding the mechanisms underlying cognitive decline or dementia caused by CSVD and the lack of effective and accurate non-invasive assessment methods, the search for an effective and accurate non-invasive assessment method is crucial for the clinical diagnosis and treatment of CSVD.

[0004] Currently, using brain diffusion tensor imaging (DTI) and resting-state functional magnetic resonance (RS-fMRI) data to explore the pathogenesis of cognitive impairment caused by CSVD from the structural and functional perspectives of brain connectivity patterns and the clinical symptoms caused by them has become a hot research direction. Studies have found that CSVD patients have abnormal functional activity in the prefrontal cortex, subcortical area, cingulate gyrus and hippocampus, and the degree of abnormality in these brain networks is significantly correlated with the severity of the disease [4]. Studies have compared cerebral microvascular load scores based on traditional imaging markers with brain network indicators and found that brain network change indicators are more correlated with cognition [5]. Studies have also shown that the relationship between these CSVD imaging markers and cognitive decline is at least partially caused by the interruption of brain functional networks [6]. These studies have confirmed that structural and functional analysis of brain connectivity patterns can reflect the pathogenesis of CSVD.

[0005] Clinical diagnosis of CSVD generally relies on imaging examinations to detect white matter lesions, enlarged perivascular spaces (EPVSs), lacunar infarcts (LIs), and cerebral microbleeds [7]. These imaging markers often appear simultaneously or sequentially but cannot assess the severity of CSVD. Recent studies have combined the characteristics of CSVD to create a score, with the severity of CSVD determined based on the score. However, the scoring mechanism is easily influenced by subjective factors and the prognostic significance of the score cannot be determined. Therefore, it is still not the best method for evaluating CSVD.

[0006] Studies have found that similar changes exist in the retinal blood vessels and intracranial blood vessels of stroke patients, and the function and composition of the retinal barrier are similar to those of the blood-brain barrier. There are certain similarities between the anatomical characteristics, barrier composition, and vascular pathology of small brain vessels and fundus microvessels. Related animal experimental studies have also shown that PM2.5 is associated with changes in the visual neural pathway [8]. Therefore, color fundus images can be used as a risk stratification tool for diseases such as dementia and stroke, and can also be used as a clinical monitoring tool to monitor asymptomatic dementia and stroke patients.

[0007] With the rapid development of computer science and technology, machine learning, represented by deep learning, has played an important role in the field of image recognition and classification. The introduction of the U-net network [9] in 2015 has enabled people to use the powerful computing power of computers to process and analyze medical images to obtain regions of interest (ROIs). In addition, radiomics methods can extract a large amount of image information from ROIs, and by mining, predicting, and analyzing image features, they can better assist doctors in making the most accurate diagnosis. U-net is a deep learning segmentation network based on convolutional connections.

[0008] References: [1] Power MC, Lamichhane AP, Liao D, et al. The Association of Long-Term Exposure to Particulate Matter air Pollution withBrain MRI Finding: The APIC Study[J]. Environ Health Perspect,2018,126(2):027009

[0009] [2] Shi Y, Wardlaw J M. Update on cerebral small vessel disease: adynamic whole-brain disease[J]. Stroke and vascular neurology,2016,1(3):83-92.

[0010] [3] Suades D E, Gascon M, Guxens M, et al. Air Pollution andNeurpsychological Development: A Review of the Latest Evidence[J].Endocrinology,2015,156(10):3473-3482.

[0011] [4] Petersen M, Frey B M, Schlemm E, et al. Network localisation ofwhite matter damage in cerebral small vessel disease[J]. Scientific reports,2020,10(1):1-9.

[0012] [5] Petersen M, Frey B M, Schlemm E, et al. Network localisation ofwhite matter damage in cerebral small vessel disease[J]. Scientific reports,2020,10(1):1-9.

[0013] [6] Liu R, Wu W, Ye Q, et al. Distinctive and pervasive alterationsof functional brain networks in cerebral small vessel disease with andwithout cognitive impairment[J]. Dementia and geriatric cognitive disorders,2019,47(1-2):55-67.

[0014] [7] Hypertension-Induced Cerebral Small Vessel Disease Leading to Cognitive Impairment[J]. Yang Liu,Yan-Hong Dong,Pei-Yuan Lyu,Wei-Hong Chen,Rui Li.Chinese Medical Journal. 2018 (5)

[0015] [8] Zhang Ying, Zhang Min, Jia Yanwen, Huang Kuankuan, Huang Shan, Zhang Zhixiang, Yun Wenwei. Research progress on the application of retinal vascular imaging technology in cerebral small vessel disease[J]. Chinese Journal of Neurology, 2021, 54(01): 64-70.

[0016] [9] Ronneberger, Olaf, et al. U-Net: Convolutional Networks forBiomedical Image Segmentation. ArXiv Preprint ArXiv:1505.04597, 2015.

[0017] The Chinese invention patent document with publication number CN114305387A discloses a method, device and medium for classifying cerebral small vessel lesions images based on magnetic resonance imaging. The method first constructs and trains an image classification model based on ensemble learning, then obtains a brain MRI image to be classified, and applies the image classification model to obtain the lesion category corresponding to the brain MRI image to be classified; training the image classification model based on ensemble learning specifically includes the following steps: S101, obtaining a brain MRI data set, preprocessing the image of the brain MRI data set, and obtaining a preprocessed image; S102, performing computational analysis on the preprocessed image to obtain multiple corresponding functional metrics, and screening and obtaining several metric features for classification based on the multiple functional metrics; S103, training the image classification model based on the metric features.

[0018] At present, it has been proven that light stimulation is effective in slowing down Alzheimer's disease. The light stimulation parameters include light stimulation time, light stimulation frequency and light stimulation wavelength. Due to the different degrees of Alzheimer's disease and cognitive decline in brain function, personalized selection and testing are needed. However, there are currently no relevant tests and light stimulation parameter selections. It is currently believed that the stimulation conditions are the same, and individual differences are not taken into account.

[0019] With respect to the above-mentioned related technologies, the inventors believe that the above-mentioned methods lack consideration of individual differences and thus lack relevant stimulation control means, especially lack the selection and control of light stimulation parameters. Summary of the Invention

[0020] In view of the defects in the prior art, the purpose of the present invention is to provide a method and system for selecting patient light stimulation parameters based on brain functional network.

[0021] According to the present invention, a method for selecting light stimulation parameters for a patient based on a brain functional network is provided, comprising the following steps:

[0022] Preprocessing step: Acquire and preprocess brain diffusion tensor modality data and resting-state functional modality data of patients with cerebral small vessel disease;

[0023] Attribute and edge acquisition steps: Based on the preprocessed resting-state functional modal data, graph theory is used to construct a brain functional network. Statistical analysis is then performed based on the brain functional network to obtain multiple different global network attributes and multiple edges with different connection weights.

[0024] Radiomics feature acquisition steps: Use a convolutional neural network to segment the hippocampus region on the preprocessed brain diffusion tensor modality data to obtain the region of interest, extract image features from the region of interest, quantify the information contained in the image, and obtain multiple radiomics features;

[0025] Vector output step: The acquired global network properties, edges with different connection weights, and imaging omics features are integrated to form a feature matrix, multiple basic classifiers are constructed and trained using ensemble learning, and finally a vector is output; the output vector represents the selection result of the patient's photostimulation treatment parameters.

[0026] Preferably, the pretreatment step comprises the following steps:

[0027] Data acquisition steps: Acquire brain diffusion tensor modality data and resting-state functional modality data from multiple patients with cerebral small vessel disease;

[0028] Mark the regions of interest of the fundus and hippocampus on the fundus retinal image and brain diffusion tensor modality respectively;

[0029] The collected data were divided into a healthy control group and a disease group; each case of data included fundus images, brain diffusion tensor modality, resting-state functional modality images, whether brain connectivity was normal, and whether the patient had cerebral small vessel disease;

[0030] Data desensitization steps: Standardize the collection of different clinical data and perform data desensitization; desensitize personally identifiable information and sensitive information in the data, and re-examine and verify the collected original data;

[0031] Data processing steps: The data without brain structural images for each subject in the brain diffusion tensor modality and resting-state functional modality were discarded, and all image data were time-domain rectified, spatially aligned, and Gaussian filtered; the image data were unified in size;

[0032] Label making steps: The light stimulation therapy parameters for patients with cerebral small vessel disease include illumination time, wavelength, frequency, oxygen inhalation time and pressure; the light stimulation therapy parameters are standardized according to the clinical treatment choice to make light stimulation parameter labels.

[0033] Preferably, the attribute and edge acquisition step includes the following steps:

[0034] Network construction steps: Based on resting-state functional modality data, graph theory is used to construct functional brain networks;

[0035] First, based on the automatic anatomical labeling template, the software was used to divide the brain of each subject in the healthy control group and the diseased group into multiple regions of interest based on the resting-state functional modality data.

[0036] Then, the Pearson correlation function is used to calculate the correlation coefficient of the time series of the region of interest to obtain the correlation coefficient matrix, and each correlation coefficient matrix is divided into multiple matrices using sparsity; where sparsity is the ratio of the edges existing in the network to the edges existing in the network;

[0037] Subnetwork division step: multiple brain regions are divided into multiple subnetworks, including the default network, attention network, subcortical network, auditory network, visual network and sensorimotor network, and divided into matrices using sparsity;

[0038] Indicator acquisition steps: Multiple network attributes were calculated for each matrix to compare brain function differences between patients with SVD and healthy controls. Network attributes included clustering coefficient, efficiency, transfer coefficient, characteristic path length, and small-world properties. The area under the curve of each attribute was used as a scale for intergroup comparisons. Differential attributes were used as network features to determine whether patients had SVD. Multiple global network attributes were obtained as imaging-based indicators related to SVD.

[0039] A one-tailed test analysis was performed on multiple pairs of functional connections between the healthy group and the diseased group, and finally multiple edges with different connection weights were obtained. Finally, the imaging genomics feature matrix of brain functional connections for each patient with cerebral small vessel disease was obtained.

[0040] Preferably, the imaging feature acquisition step comprises the following steps:

[0041] Segmentation step: Use the model to segment the hippocampal lesion area and retinal lesion area on the obtained brain diffusion tensor modality image and fundus image;

[0042] Omic feature acquisition steps: Image feature extraction is performed on the region of interest and retinal lesion area respectively to obtain multiple omic features, including shape features, grayscale features and texture features.

[0043] Preferably, the vector output step includes the following steps:

[0044] Reconstruction step: reconstructing the radiomics feature matrix and radiomics features to obtain the feature matrix of the cerebral small vessel disease light stimulation parameter selection model;

[0045] Training steps: Train multiple basic classifiers and weight their respective results to obtain the final output; basic learners include support vector machine classifiers, logistic regression classifiers, and naive Bayes classifiers; light stimulation parameter selection is a multi-classification problem;

[0046] Prediction step: After the training of multiple basic classifiers is completed, the output value is predicted using a combination strategy of weighted average method to obtain a vector; the model output result is compared with the clinical parameter selection to calculate the accuracy.

[0047] According to the present invention, a system for selecting light stimulation parameters for patients based on a brain function network is provided, comprising the following modules:

[0048] Preprocessing module: Acquire and preprocess brain diffusion tensor modality data and resting-state functional modality data of patients with cerebral small vessel disease;

[0049] Attribute and edge acquisition module: Based on the preprocessed resting-state functional modal data, graph theory is used to construct a brain functional network. Statistical analysis is then performed based on the brain functional network to obtain multiple different global network attributes and multiple edges with different connection weights.

[0050] Radiomics feature acquisition module: Uses convolutional neural networks to segment the hippocampus region on preprocessed brain diffusion tensor modality data to obtain regions of interest, extracts image features from these regions of interest, quantifies the information contained in the image, and obtains multiple radiomics features;

[0051] Vector output module: The acquired global network properties, edges with different connection weights, and imaging omics features are integrated to form a feature matrix. Multiple basic classifiers are constructed and trained using ensemble learning, and finally a vector is output. The output vector represents the selection result of the patient's photostimulation treatment parameters.

[0052] Preferably, the preprocessing module includes the following modules:

[0053] Data acquisition module: Acquire brain diffusion tensor modal data and resting-state functional modal data of multiple patients with cerebral small vessel disease;

[0054] Mark the regions of interest of the fundus and hippocampus on the fundus retinal image and brain diffusion tensor modality respectively;

[0055] The collected data were divided into a healthy control group and a disease group; each case of data included fundus images, brain diffusion tensor modality, resting-state functional modality images, whether brain connectivity was normal, and whether the patient had cerebral small vessel disease;

[0056] Data desensitization module: Standardizes the collection of different clinical data and performs data desensitization; desensitizes personally identifiable information and sensitive information in the data, and re-examines and verifies the collected raw data;

[0057] Data processing module: discard the data of each subject without brain structural images in the brain diffusion tensor modality and resting-state functional modality, and perform time domain correction, spatial domain alignment and Gaussian filtering on all image data; the image data is unified in size;

[0058] Label making module: The light stimulation therapy parameters for patients with cerebral small vessel disease include illumination time, wavelength, frequency, oxygen inhalation time and pressure; light stimulation therapy parameters are standardized and light stimulation parameter labels are produced according to the clinical treatment options.

[0059] Preferably, the attribute and edge acquisition module includes the following modules:

[0060] Network construction module: Based on resting-state functional modality data, graph theory is used to construct functional brain networks;

[0061] First, based on the automatic anatomical labeling template, the software was used to divide the brain of each subject in the healthy control group and the diseased group into multiple regions of interest based on the resting-state functional modality data.

[0062] Then, the Pearson correlation function is used to calculate the correlation coefficient of the time series of the region of interest to obtain the correlation coefficient matrix, and each correlation coefficient matrix is divided into multiple matrices using sparsity; where sparsity is the ratio of the edges existing in the network to the edges existing in the network;

[0063] Subnetwork partitioning module: divides multiple brain regions into multiple subnetworks, including the default network, attention network, subcortical network, auditory network, visual network and sensorimotor network, and divides them into matrices using sparsity;

[0064] Indicator acquisition module: Multiple network attributes are calculated for each matrix to compare brain function differences between patients with SVD and healthy controls. Network attributes include clustering coefficient, efficiency, transfer coefficient, characteristic path length, and small-world properties. The area under the curve of each attribute is used as a scale for inter-group comparisons. Differential attributes are used as network features to determine whether patients have SVD. Multiple global network attributes are obtained as relevant imaging omics indicators for SVD.

[0065] A one-tailed test analysis was performed on multiple pairs of functional connections between the healthy group and the diseased group, and finally multiple edges with different connection weights were obtained. Finally, the imaging genomics feature matrix of brain functional connections for each patient with cerebral small vessel disease was obtained.

[0066] Preferably, the imaging feature acquisition module includes the following modules:

[0067] Segmentation module: Use the model to segment the hippocampal lesion area and retinal lesion area on the obtained brain diffusion tensor modality image and fundus image;

[0068] Omic feature acquisition module: Image features of the region of interest and retinal lesion area are extracted respectively to obtain multiple omics features, including shape features, grayscale features and texture features.

[0069] Preferably, the vector output module includes the following modules:

[0070] Reconstruction module: reconstructs the radiomics feature matrix and radiomics features to obtain the feature matrix of the cerebral small vessel disease light stimulation parameter selection model;

[0071] Training module: trains multiple basic classifiers and weights their results to obtain the final output; basic learners include support vector machine classifiers, logistic regression classifiers, and naive Bayes classifiers; light stimulation parameter selection is a multi-classification problem;

[0072] Prediction module: After the training of multiple basic classifiers is completed, the output value is predicted using a combination strategy of weighted average method to obtain a vector; the model output result is compared with the clinical parameter selection to calculate the accuracy.

[0073] Compared with the prior art, the present invention has the following beneficial effects:

[0074] 1. This invention uses brain network to evaluate the condition of CSVD patients, using advanced segmentation model U 2 -net segments the hippocampal lesion area and combines it with radiomics to extract highly correlated features for classification, thereby achieving automated selection of CSVD light stimulation parameters and overcoming the many limitations of complex scoring used in clinical practice;

[0075] 2. The method for selecting light stimulation parameters for CSVD patients based on brain functional networks provided by the present invention has high sensitivity and accuracy, which can help guide the rapid diagnosis of CSVD patients and accelerate the selection of treatment options;

[0076] 3. The present invention has better clinical practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0078] Figure 1 Flowchart of the present invention;

[0079] Figure 2 This is the network structure diagram of the present invention. DETAILED DESCRIPTION

[0080] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0081] The embodiment of the present invention discloses a method for selecting light stimulation parameters for patients based on brain function network, such as Figure 1 and Figure 2 As shown, the following steps are included: Preprocessing step: obtaining and preprocessing brain diffusion tensor modal data and resting-state functional modal data of patients with cerebral small vessel disease.

[0082] The preprocessing steps include the following steps: data acquisition step: obtaining brain diffusion tensor modal data and resting-state functional modal data of multiple patients with cerebral small vessel disease; marking the fundus and hippocampus regions of interest on the fundus retinal images and brain diffusion tensor modalities respectively; dividing the collected data into a healthy control group and a disease group; each data set contains fundus images, brain diffusion tensor modalities, resting-state functional modal images, whether brain connections are normal, and whether the patient has cerebral small vessel disease.

[0083] Data desensitization steps: Standardize the collection of different clinical data and perform data desensitization; desensitize personally identifiable information and sensitive information in the data, and re-examine and verify the collected original data.

[0084] Data processing steps: The data without brain structural images for each subject in the brain diffusion tensor modality and resting-state functional modality were discarded, and all image data were time-domain rectified, spatially aligned, and Gaussian filtered; the image data were unified in size.

[0085] Label making steps: The light stimulation therapy parameters for patients with cerebral small vessel disease include illumination time, wavelength, frequency, oxygen inhalation time and pressure; the light stimulation therapy parameters are standardized according to the clinical treatment choice to make light stimulation parameter labels.

[0086] Attribute and edge acquisition steps: Based on the preprocessed resting-state functional modal data, graph theory is used to construct a brain functional network, and statistical analysis is performed based on the brain functional network to obtain multiple different global network attributes and multiple edges with different connection weights.

[0087] The attribute and edge acquisition steps include the following steps: Network construction step: Based on the resting-state functional modal data, graph theory is used to construct a functional brain network; first, based on the automatic anatomical label template, the software is used to divide the brain of each subject in the healthy control group and the diseased group into multiple regions of interest; then, the Pearson correlation function is used to calculate the correlation coefficient of the time series of the region of interest to obtain the correlation coefficient matrix, and each correlation coefficient matrix is divided into multiple matrices using sparsity; among them, sparsity is the ratio of edges existing in the network to edges existing in the network.

[0088] Subnetwork division step: multiple brain regions are divided into multiple subnetworks, including the default network, attention network, subcortical network, auditory network, visual network and sensorimotor network, and divided into matrices using sparsity.

[0089] Indicator acquisition steps: Multiple network attributes were calculated for each matrix to compare brain function differences between patients with SVD and healthy controls. These network attributes included clustering coefficient, efficiency, transfer coefficient, characteristic path length, and small-world properties. The area under the curve of each attribute was used as a metric for intergroup comparisons. Attributes with differences were used as network features to identify patients with SVD. Multiple global network attributes were then derived as radiomics indicators of SVD. Multiple pairs of functional connections between the healthy and diseased groups were analyzed using one-tailed tests. Finally, multiple edges with different connection weights were identified, resulting in a radiomics feature matrix for brain functional connectivity in each SVD patient.

[0090] Steps for acquiring radiomics features: Use a convolutional neural network to segment the hippocampus region of the preprocessed brain diffusion tensor modality data to obtain the region of interest, extract image features from the region of interest, quantify the information contained in the image, and obtain multiple radiomics features.

[0091] The imaging genomics feature acquisition steps include the following steps: Segmentation step: Use the model to segment the obtained brain diffusion tensor modality image and fundus image into hippocampal lesion areas and retinal lesion areas.

[0092] Omic feature acquisition steps: Image feature extraction is performed on the region of interest and retinal lesion area respectively to obtain multiple omic features, including shape features, grayscale features and texture features.

[0093] Vector output step: The acquired global network properties, edges with different connection weights, and imaging omics features are integrated to form a feature matrix, multiple basic classifiers are constructed and trained using ensemble learning, and finally a vector is output; the output vector represents the selection result of the patient's photostimulation treatment parameters.

[0094] The vector output step includes the following steps: reconstruction step: reconstructing the imaging omics feature matrix and the imaging omics features to obtain the feature matrix of the cerebral small vessel disease light stimulation parameter selection model.

[0095] Training steps: Train multiple basic classifiers and weight their respective results to obtain the final output; basic learners include support vector machine classifiers, logistic regression classifiers and naive Bayes classifiers; light stimulation parameter selection is a multi-classification problem.

[0096] Prediction step: After the training of multiple basic classifiers is completed, the output value is predicted using a combination strategy of weighted average method to obtain a vector; the model output result is compared with the clinical parameter selection to calculate the accuracy.

[0097] The embodiment of the present invention also discloses a patient light stimulation parameter selection system based on brain function network, such as Figure 1 and Figure 2 As shown, it includes the following modules: Preprocessing module: acquires and preprocesses the brain diffusion tensor modal data and resting-state functional modal data of patients with cerebral small vessel disease.

[0098] The preprocessing module includes the following modules: data acquisition module: obtain brain diffusion tensor modal data and resting-state functional modal data of multiple patients with cerebral small vessel disease; mark the fundus and hippocampus regions of interest on the fundus retinal images and brain diffusion tensor modalities respectively; divide the collected data into a healthy control group and a disease group; each data set contains fundus images, brain diffusion tensor modalities, resting-state functional modal images, whether the brain connections are normal, and whether the patient has cerebral small vessel disease.

[0099] Data desensitization module: standardizes the collection of different clinical data and performs data desensitization; desensitizes personally identifiable information and sensitive information in the data, and re-examines and verifies the collected original data.

[0100] Data processing module: The data without brain structural images of each subject in the brain diffusion tensor modality and resting-state functional modality are discarded, and all image data are subjected to time domain correction, spatial domain alignment and Gaussian filtering; the image data are unified in size.

[0101] Label making module: The light stimulation therapy parameters for patients with cerebral small vessel disease include illumination time, wavelength, frequency, oxygen inhalation time and pressure; light stimulation therapy parameters are standardized and light stimulation parameter labels are produced according to the clinical treatment options.

[0102] Attribute and edge acquisition module: Based on the preprocessed resting-state functional modal data, graph theory is used to construct a brain functional network, and statistical analysis is performed based on the brain functional network to obtain multiple different global network attributes and multiple edges with different connection weights.

[0103] The attribute and edge acquisition module includes the following modules: Network construction module: Based on resting-state functional modal data, graph theory is used to construct a functional brain network; first, based on the automatic anatomical label template, the software is used to divide the brain of each subject in the healthy control group and the diseased group into multiple regions of interest; then, the Pearson correlation function is used to calculate the correlation coefficient of the time series of the region of interest to obtain the correlation coefficient matrix, and each correlation coefficient matrix is divided into multiple matrices using sparsity; among them, sparsity is the ratio of edges existing in the network to edges existing in the network.

[0104] Subnetwork division module: multiple brain regions are divided into multiple subnetworks, including the default network, attention network, subcortical network, auditory network, visual network and sensorimotor network, and divided into matrices using sparsity.

[0105] Indicator Acquisition Module: Multiple network attributes are calculated for each matrix to compare brain function differences between patients with SVD and healthy controls. These network attributes include clustering coefficient, efficiency, transfer coefficient, characteristic path length, and small-world properties. The area under the curve of each attribute is used as a scale for intergroup comparisons. Differential attributes are used as network features to determine whether patients have SVD. Multiple global network attributes are obtained as relevant imaging genomics indicators for SVD. Multiple pairs of functional connections between the healthy and diseased groups are analyzed using one-tailed tests. Finally, multiple edges with different connection weights are obtained, resulting in an imaging genomics feature matrix for brain functional connectivity in each SVD patient.

[0106] Radiomics feature acquisition module: Use convolutional neural networks to segment the hippocampus region of the preprocessed brain diffusion tensor modality data to obtain the region of interest, extract image features from the region of interest, quantify the information contained in the image, and obtain multiple radiomics features.

[0107] The imaging genomics feature acquisition module includes the following modules: Segmentation module: Use the model to segment the obtained brain diffusion tensor modality image and fundus image into the hippocampal lesion area and retinal lesion area.

[0108] Omic feature acquisition module: Image features of the region of interest and retinal lesion area are extracted respectively to obtain multiple omics features, including shape features, grayscale features and texture features.

[0109] Vector output module: The acquired global network properties, edges with different connection weights, and imaging omics features are integrated to form a feature matrix. Multiple basic classifiers are constructed and trained using ensemble learning, and finally a vector is output. The output vector represents the selection result of the patient's photostimulation treatment parameters.

[0110] The vector output module includes the following modules: reconstruction module: reconstructs the imaging genomics feature matrix and imaging genomics features to obtain the feature matrix of the cerebral small vessel disease light stimulation parameter selection model.

[0111] Training module: train multiple basic classifiers and weight their respective results to obtain the final output; basic learners include support vector machine classifiers, logistic regression classifiers and naive Bayes classifiers; light stimulation parameter selection is a multi-classification problem.

[0112] Prediction module: After the training of multiple basic classifiers is completed, the output value is predicted using a combination strategy of weighted average method to obtain a vector; the model output result is compared with the clinical parameter selection to calculate the accuracy.

[0113] The present invention also discloses a method for selecting light stimulation parameters for CSVD patients based on brain functional network. Figure 1 and Figure 2 As shown, the method includes the following steps: Step S1: Obtain DTI, RS-fMRI imaging data and clinical data of 300 CSVD patients and perform preprocessing. Create light stimulation parameter labels according to the clinical treatment plan.

[0114] Step S1 specifically includes the following steps: S11: Acquire 300 dust environment-fundus-brain functional connectivity-CSVD imaging data and corresponding clinical data. Ophthalmologists and radiologists annotate the fundus and hippocampus ROIs on the fundus retinal images and DTI modalities, respectively. Senior radiologists divide the acquired data into two groups: a healthy control group (no SVD) and a diseased group (to detect SVD). The data inclusion criteria and the specific images and labels included are shown in Table 1. Each data set includes multiple labels, including fundus images, DTI, RS-fMRI images, and whether brain connectivity is normal and whether SVD is present. MRI stands for Magnetic Resonance Imaging (MRI). WMH stands for White Matter Hyperintensities (WMH).

[0115] Table 1 Inclusion criteria for multimodal MRI imaging, pathology, and cognitive data of CSVD patients

[0116]

[0117] S12: Standardize the collection and desensitization of different clinical data. Desensitize personally identifiable information and sensitive information in the data, and have professional physicians review and verify the collected raw data to improve data consistency and annotation accuracy.

[0118] S13: For each subject, the pre-existing brain structural images for DTI and RS-fMRI modalities were discarded. All image data were then subjected to temporal correction, spatial alignment, and Gaussian filtering. All image data were resized to 1024 × 1024 pixels.

[0119] S14: Parameters for photostimulation therapy for CSVD patients include illumination duration (0-120 minutes), wavelength (600-900 nm), frequency (10-100 Hz), oxygen inhalation duration (0-30 minutes), and pressure (0-0.3 MPa). These parameters are obtained from an external photoacoustic therapy device. Based on clinical treatment options, these five parameters are standardized to create a photostimulation parameter label. The corresponding relationships are shown in Table 2. The photostimulation parameter label corresponds to each parameter in Table 2. Different values are selected for different situations, forming the stimulation parameter label.

[0120] Table 2 Light stimulation parameter label table

[0121]

[0122] Step S2: Based on the RS-fMRI data, a functional brain network is constructed using graph theory. Statistical analysis is then performed to identify n significantly different global network attributes and m edges with significantly different connection weights as early-stage radiomics biomarkers for CSVD. Biomarkers are interchangeable imaging markers that can be used to quantify and assess abnormal brain imaging signals.

[0123] Step S2 specifically includes: S21: Constructing functional brain networks based on RS-fMRI data using graph theory. First, using the Anatomical Automatic Labeling (AAL) template provided by the MNI, the RS-fMRI data of each subject in the healthy control and disease groups were divided into 90 regions of interest (excluding the cerebellum) using GRETNA (http: / / www.nitrc.org / projects / gretna) software. Then, the Pearson correlation function was used to calculate the correlation coefficients (absolute values) between the time series of each pair of regions of interest to obtain a correlation coefficient matrix. Each correlation coefficient matrix was then partitioned into multiple binary matrices based on sparsity S. Sparsity S is the ratio of edges present in the network to the maximum possible number of edges in the network, and ranges from [0.08 to 0.52]. MNI (Montreal Neurological Institute) is a standard brain template provided by the Montreal Neurological Institute, and different brains are registered to this standard template.

[0124] S22: The 90 brain regions are divided into 6 subnetworks, namely the default mode network (DMN), attention network (ATT), subcortical network (SUB), auditory network (AUD), visual network (VIS) and sensorimotor network (SEN), and divided into binary matrices with the same sparsity S (S∈[0.08, 0.52]). The names of the corresponding subnetworks and brain region (node) names are shown in Table 3.

[0125] Table 3 Brain functional network subnet division table

[0126]

[0127] S23: Five network attributes were calculated for each binary matrix to compare the differences in brain function between the CSVD group and the healthy group, namely the clustering coefficient C, efficiency E, transfer coefficient I, characteristic path length L, and small-world attribute S. At the same time, the area under the curve (AUC) of each attribute was used as a scale for inter-group comparison. Attributes with significant differences can be used as network features to determine whether or not the patient has CSVD. A total of n global network attributes were obtained as early-stage related imaging genomics biomarker indicators for CSVD. Similarly, a one-tailed t-test analysis was performed on a total of 4005 pairs of edges of the same functional connection between the healthy group and the diseased group (p≤0.05), and finally m edges with significantly different connection weights were obtained. Finally, the imaging genomics biomarker feature matrix of early brain functional connections for each CSVD patient was obtained. :

[0128]

[0129] in, for Global network attributes, for The connection weight of the edge. This is the area of the brain network, It is the global network attribute representing different brain partitions. Generally, the number of brain partitions can be set to 28-64. The connection weights of 28-64 brain regions can be understood as 28 points in front and the lines connecting the 28 points in the back. Indicates the Global network attributes, Indicates the The connection weight of the edge.

[0130] Step S3: Apply U to the pre-processed DTI data 2 -net convolutional neural network for hippocampal region segmentation to obtain the region of interest ,right Image feature extraction was performed, and the information contained in the image was quantified to obtain 314 imaging omics features.

[0131] Step S3 specifically includes: S31: Use U 2 -net model is used to segment the hippocampal lesion area and retinal lesion area using the obtained DTI image and fundus image. 2-net is a two-level nested U-shaped structure consisting of three parts: (1) a six-layer encoder, (2) a five-layer decoder, and (3) a feature fusion layer. Each layer consists of an RSU module (residual U-block), which can fuse features of different receptive fields to capture more information. The RSU module heights of encoding layers 1-4 and decoding layers 4-1 are 7, 6, 5, and 4, respectively. (7 represents two convolution-normalization operations plus five downsampling-convolution-normalization operations; 6, 5, and 4 reduce one downsampling operation respectively.) See the network structure. Figure 2 , the specific steps of segmentation include S311-S313. 2 -net composition introduction, including RSU is U 2 -net, not a standalone parameter. RSU is a residual convolution module, DownSample represents downsampling, UpSample represents upsampling, and Concatenation represents concatenating multiple tensor matrices of the same dimension. En_1 through En_6 represent the encoder, and De_1 through De_5 represent the decoder, performing image feature representation. represents convolution; represents upsampling, Indicates downsampling.

[0132] S311: DTI original image The feature maps are obtained by 6 encodings. , Size , each time the feature map size is halved. Decode 5 times to get , gradually restore the feature map size to Each time decoding is performed, the feature map of the same layer encoding output is input into the RSU module to obtain the segmentation result. .

[0133] S312: and Get in series ( represents the feature map obtained by convolution), the size is (channel size). Perform convolution + Sigmoid activation operation to obtain the final segmentation result . Sigmoid represents an activation function.

[0134] S313: When the model is trained, each RSU module outputs a loss function. The overall loss function is the sum of the five RSU loss functions plus the loss function after feature fusion, which is expressed as , It is the abbreviation of Loss function.

[0135]

[0136] in, W represents the corresponding weight, m Indicates the corresponding number of layers and images; Indicates the maximum number of layers. is the loss function for each downsampling RSU module. is the weight of the loss function for each downsampling RSU module. is the loss function of the last RSU module for upsampling, is the weight of the loss function of the last RSU module for upsampling. , using the standard binary cross entropy loss function. By minimizing A higher segmentation accuracy can be obtained .

[0137] S314: Fundus image Repeat S311-S313 operations to obtain the corresponding retinal lesion area ;

[0138] S32: Use the WORC toolkit method to and Image feature extraction yielded 314 omics features, including 10 shape features, 16 grayscale features, 25 texture features, and 263 wavelet features. In addition, 16 clinical record features were obtained from S12. The categories and numbers of extracted features are shown in Table 4. WORC is a radiomics computational toolkit included in the open-source Python software.

[0139] Table 4. Characteristics and numbers of hippocampal and fundus lesions

[0140]

[0141] Step S4: The n+m features representing changes in brain network properties and connection weights obtained in Step 2 are combined with the radiomics features from Step 3 to form a feature matrix. Three basic classifiers are constructed and trained using ensemble learning to output a one-dimensional vector. This vector represents the selected parameters for the patient's photostimulation therapy.

[0142] Step S4 specifically includes: S41: reconstructing the feature matrix obtained in S23 and S32 to obtain the feature matrix of the CSVD light stimulation parameter selection model , is the feature matrix obtained by S23 and S32, and the fusion parameter selection matrix is obtained by merging:

[0143] ;

[0144] in, represents the feature matrix set obtained by S23, All are characteristic matrices of the S23 characteristic matrix set; Represents the feature matrix set obtained by S32, Indicates the The feature matrix obtained by S32.

[0145] S42: Using the AdaBoost ensemble learning algorithm, three basic classifiers are trained and their results are weighted to obtain the final output. The three basic learners are SVM classifier, logistic regression classifier and naive Bayes classifier. The selection of light stimulation parameters is a multi-classification problem, defined as . ( ) The matrix Input three basic classifiers to calculate belong The probability of , the maximum value of which is The category to which it belongs. SVM is called Support Vector Machines in English and its Chinese translation is Support Vector Machine Represents the light stimulation parameter label, corresponding to the different stimulation parameters predicted in the previous text.

[0146] S423: After the three basic classifiers are trained, the output value is predicted using a weighted average method to obtain a one-dimensional column vector The formula is ,in are the weights of the three basic classifiers, set to (0.4, 0.4, 0.2), Represents three basic classifiers. The model output was compared with the clinical parameter selection to calculate the accuracy. The accuracy of the CSVD light stimulation parameter selection method of this model was 90%, which is of great significance for the clinical diagnosis and treatment selection of CSVD. and It is a collection of 3 classifiers and integrated output weights.

[0147] Those skilled in the art will appreciate that, in addition to implementing the system and its various devices, modules, and units provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same functions of the system and its various devices, modules, and units provided by the present invention in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; the devices, modules, and units for implementing various functions can also be considered as both software modules implementing the method and structures within the hardware component.

[0148] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A method for selecting light stimulation parameters for patients based on brain functional network, characterized in that: The steps include: Preprocessing step: obtaining multiple sample data and preprocessing the brain diffusion tensor modality data and resting-state functional modality data of each sample data; Attribute and edge acquisition steps: Based on the preprocessed resting-state functional modal data, graph theory is used to construct a brain functional network. Statistical analysis is then performed based on the brain functional network to obtain multiple different global network attributes and multiple edges with different connection weights. Radiomics feature acquisition steps: Use a convolutional neural network to segment the hippocampus region on the preprocessed brain diffusion tensor modality data to obtain the region of interest, extract image features from the region of interest, quantify the information contained in the image, and obtain multiple radiomics features; Vector output step: The acquired global network attributes, edges with different connection weights, and radiomics features are integrated to form a feature matrix. Multiple basic classifiers are constructed and trained using ensemble learning. Finally, a vector is output. The output vector represents the selection result of the light stimulation treatment parameters. The pre-processing step comprises the following steps: Data acquisition steps: Acquire brain diffusion tensor modality data and resting-state functional modality data of multiple sample data; Mark the fundus and hippocampus regions of interest for fundus retinal images and brain diffusion tensor modality data respectively; The collected data were divided into a healthy control group and a disease group; each case of data included fundus images, brain diffusion tensor modality data, resting-state functional modality data, whether brain connectivity was normal, and whether the patient had cerebral small vessel disease; Data desensitization steps: Standardize the collection of different clinical data and perform data desensitization; desensitize personally identifiable information and sensitive information in the data, and re-examine and verify the collected original data; Data processing steps: The data without brain structural images for each subject in the brain diffusion tensor modality data and resting-state functional modality data were discarded, and all image data were subjected to time domain correction, spatial domain alignment and Gaussian filtering; the image data were unified in size; Labeling steps: The light stimulation therapy parameters of the sample data include illumination time, wavelength, frequency, oxygen uptake time and pressure; standardize the light stimulation therapy parameters according to the clinical treatment choice to create light stimulation parameter labels; The attribute and edge acquisition step includes the following steps: Network construction steps: Based on resting-state functional modality data, graph theory is used to construct functional brain networks; First, based on the automatic anatomical labeling template, the software was used to divide the brain of each subject in the healthy control group and the diseased group into multiple regions of interest based on the resting-state functional modality data. Then, the Pearson correlation function is used to calculate the correlation coefficient of the time series of the region of interest to obtain a correlation coefficient matrix, and each correlation coefficient matrix is divided into multiple matrices using sparsity; where sparsity is the ratio of the edges existing in the network to the maximum possible edges in the network; Subnetwork division step: multiple brain regions are divided into multiple subnetworks, including the default network, attention network, subcortical network, auditory network, visual network and sensorimotor network, and divided into matrices using sparsity; Indicator acquisition steps: Multiple network attributes were calculated for each matrix to compare brain function differences between the diseased and healthy control groups. Network attributes included clustering coefficient, efficiency, transfer coefficient, characteristic path length, and small-world properties. The area under the curve of each attribute was used as a scale for intergroup comparisons. Differential attributes were used as network features to determine whether patients had SVD. Multiple global network attributes were obtained as relevant imaging omics indicators for SVD. A one-tailed test analysis was performed on multiple pairs of functional connections between the healthy control group and the diseased group, and finally multiple edges with different connection weights were obtained, and finally the imaging genomics feature matrix of brain functional connections for each sample data was obtained.

2. The method for selecting patient light stimulation parameters based on brain functional network according to claim 1, characterized in that: The imaging genomics feature acquisition step comprises the following steps: Segmentation step: Use the model to segment the hippocampal lesion area and retinal lesion area on the obtained brain diffusion tensor modality data and fundus images; Omic feature acquisition steps: Image feature extraction is performed on the hippocampal lesion area and the retinal lesion area respectively to obtain multiple imaging omics features, including shape features, grayscale features and texture features.

3. The method for selecting patient light stimulation parameters based on brain functional network according to claim 1, characterized in that: The vector output step comprises the following steps: Reconstruction step: reconstructing the radiomics feature matrix and radiomics features to obtain the feature matrix of the cerebral small vessel disease light stimulation parameter selection model; Training steps: training multiple basic classifiers; basic classifiers include support vector machine classifier, logistic regression classifier and naive Bayes classifier; The selection of light stimulation parameters is a multi-classification problem; Prediction step: After the training of multiple basic classifiers is completed, the output of multiple basic classifiers is combined with the weighted average method to obtain a vector; the model output results are compared with the clinical parameter selection to calculate the accuracy.

4. A system for selecting light stimulation parameters for patients based on brain functional network, characterized in that: Includes the following modules: Preprocessing module: acquires multiple sample data and preprocesses the brain diffusion tensor modality data and resting state functional modality data of each sample data; Attribute and edge acquisition module: Based on the preprocessed resting-state functional modal data, graph theory is used to construct a brain functional network. Statistical analysis is then performed based on the brain functional network to obtain multiple different global network attributes and multiple edges with different connection weights. Radiomics feature acquisition module: Uses convolutional neural networks to segment the hippocampus region on preprocessed brain diffusion tensor modality data to obtain regions of interest, extracts image features from these regions of interest, quantifies the information contained in the image, and obtains multiple radiomics features; Vector output module: This module integrates the acquired global network attributes, edges with different connection weights, and radiomics features to form a feature matrix. It then constructs multiple basic classifiers and trains them using ensemble learning. Finally, it outputs a vector representing the selected parameters for photostimulation therapy. The pre-processing module includes the following modules: Data acquisition module: acquires brain diffusion tensor modality data and resting-state functional modality data of multiple sample data; Mark the fundus and hippocampus regions of interest for fundus retinal images and brain diffusion tensor modality data respectively; The collected data were divided into a healthy control group and a disease group; each case of data included fundus images, brain diffusion tensor modality data, resting-state functional modality data, whether brain connectivity was normal, and whether the patient had cerebral small vessel disease; Data desensitization module: Standardizes the collection of different clinical data and performs data desensitization; desensitizes personally identifiable information and sensitive information in the data, and re-examines and verifies the collected raw data; Data processing module: discard the data without brain structural images from each subject in the brain diffusion tensor modal data and resting-state functional modal data, and perform temporal correction, spatial alignment, and Gaussian filtering on all image data; the image data are unified in size; Labeling module: The light stimulation therapy parameters of the sample data include illumination time, wavelength, frequency, oxygen uptake time and pressure; light stimulation therapy parameters are standardized and labeled according to the clinical treatment options; The attribute and edge acquisition module includes the following modules: Network construction module: Based on resting-state functional modality data, graph theory is used to construct functional brain networks; First, based on the automatic anatomical labeling template, the software was used to divide the brain of each subject in the healthy control group and the diseased group into multiple regions of interest based on the resting-state functional modality data. Then, the Pearson correlation function is used to calculate the correlation coefficient of the time series of the region of interest to obtain a correlation coefficient matrix, and each correlation coefficient matrix is divided into multiple matrices using sparsity; where sparsity is the ratio of the edges existing in the network to the maximum possible edges in the network; Subnetwork partitioning module: divides multiple brain regions into multiple subnetworks, including the default network, attention network, subcortical network, auditory network, visual network and sensorimotor network, and divides them into matrices using sparsity; Indicator Acquisition Module: Multiple network attributes are calculated for each matrix to compare brain function differences between the diseased group and the healthy control group. Network attributes include clustering coefficient, efficiency, transfer coefficient, characteristic path length, and small-world properties. The area under the curve of each attribute is used as a scale for inter-group comparison. Differential attributes are used as network features to determine whether patients have SVD. Multiple global network attributes are obtained as relevant imaging omics indicators for SVD. A one-tailed test analysis was performed on multiple pairs of functional connections between the healthy control group and the diseased group, and finally multiple edges with different connection weights were obtained, and finally the imaging genomics feature matrix of brain functional connections for each sample data was obtained.

5. The patient light stimulation parameter selection system based on brain function network according to claim 4, characterized in that: The imaging feature acquisition module includes the following modules: Segmentation module: Use the model to segment the hippocampal lesion area and retinal lesion area based on the obtained brain diffusion tensor modality data and fundus images; Omic feature acquisition module: Image features are extracted from the hippocampal lesion area and the retinal lesion area respectively to obtain multiple imaging omics features, including shape features, grayscale features, and texture features.

6. The patient light stimulation parameter selection system based on brain function network according to claim 4, characterized in that: The vector output module includes the following modules: Reconstruction module: reconstructs the radiomics feature matrix and radiomics features to obtain the feature matrix of the cerebral small vessel disease light stimulation parameter selection model; Training module: trains multiple basic classifiers; basic classifiers include support vector machine classifier, logistic regression classifier and naive Bayes classifier; The selection of light stimulation parameters is a multi-classification problem; Prediction module: After the training of multiple basic classifiers is completed, the output of multiple basic classifiers is combined with the weighted average method to obtain a vector; the model output results are compared with the clinical parameter selection to calculate the accuracy.

Images