Lymphatic system magnetic resonance imaging prediction postoperative brain injury evaluation method and system

Abstract

The invention discloses a lymphatic system magnetic resonance imaging prediction postoperative brain injury evaluation method and system, and relates to the field of brain injury evaluation, and the method comprises the steps: obtaining multi-modal magnetic resonance imaging image data of a joint anatomy region of a target subject; extracting magnetic resonance imaging function parameters based on the multi-mode magnetic resonance imaging image data; fusing the magnetic resonance imaging function parameters, the brain region white matter fiber integrity parameters, the cerebral hemodynamic indexes and the clinical data of the operation process to construct a multi-dimensional feature vector; and inputting the feature vectors into a hybrid neural network model consisting of a graph convolutional network and a long-short-term memory network, and outputting risk prediction values of postoperative encephaledema, cognitive impairment and white matter injury. According to the method, the limitation that an existing evaluation means is highly dependent on a structural image, functional evaluation is lacked and prediction accuracy is insufficient is broken through, and accurate prediction and intervention of postoperative brain injury based on lymphatic-like function imaging are realized.

Description

TECHNICAL FIELD

[0001] The present specification relates to the field of brain injury assessment, and more specifically, to a lymphoid system magnetic resonance imaging prediction method and system for postoperative brain injury assessment. BACKGROUND

[0002] Current imaging evaluation of postoperative brain injury in neurosurgery mainly relies on structural magnetic resonance imaging (MRI) techniques, including T1-weighted, T2-weighted, FLAIR, and other sequences. These methods can be used to observe changes in brain tissue structure, edema range, or hemorrhagic lesions, and other anatomical changes, but are difficult to reflect more critical pathophysiological processes such as postoperative brain metabolism, white matter fiber damage, or lymphoid system dysfunction. For example, traditional structural images cannot dynamically evaluate the diffusion efficiency of liquid in the perivascular space (PVS), the clearance capacity of metabolic waste, and other key functions of the lymphoid system.

[0003] To overcome the limitations of structural images, some studies have attempted to visualize the lymphoid system by intrathecal injection of contrast agents (such as gadolinium) or fluorescent markers, but such methods have significant invasive risks, such as infection, meningeal irritation, and are difficult to repeat in humans. In addition, some animal studies have achieved higher spatial resolution through near-infrared imaging or micro-optical windows, but their application in the clinic is still limited by ethical and equipment issues.

[0004] On the other hand, existing postoperative brain injury prediction models are mostly based on single image features (such as CBV, ADC) or structural changes (such as edema volume), and fail to integrate multi-source information such as lymphoid function, blood perfusion, white matter fiber integrity, and intraoperative operation variables, thus having poor prediction specificity and low timeliness. At the same time, these methods generally rely on manual segmentation of regions of interest (ROI) or manual feature extraction, which is inefficient, highly subjective, and difficult to promote in large-scale patients.

[0005] Notably, previous studies on the evaluation of the lymphoid system have focused on chronic neurodegenerative diseases such as Alzheimer's disease, and in the context of perioperative acute injury, there is a lack of standardized quantitative indicators such as the ALPS index and the CEPS index, as well as corresponding human verification systems, limiting their practical application in predicting postoperative brain injury.

[0006] Therefore, it is necessary to propose a lymphoid system magnetic resonance imaging prediction method for postoperative brain injury assessment to at least solve some of the above problems. SUMMARY

[0007] The summary section introduces a series of simplified concepts, which will be further explained in detail in the detailed description section. This summary section is not intended to limit the key features and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0008] In a first aspect, the present invention proposes a method for predicting postoperative brain injury assessment using magnetic resonance imaging of a lymphatic system, comprising:

[0009] Multimodal magnetic resonance imaging data of the joint anatomical region of the target subject was acquired, wherein the multimodal magnetic resonance imaging data included diffusion tensor imaging data, dynamic contrast-enhanced imaging data and arterial spin labeling magnetic resonance imaging sequence.

[0010] Based on the above multimodal magnetic resonance imaging data, magnetic resonance imaging functional parameters are extracted, including the diffusion index along the perivascular space direction, the contrast enhancement perivascular space index, and the perivascular space volume ratio.

[0011] By integrating the above-mentioned magnetic resonance imaging functional parameters, brain region white matter fiber integrity parameters, cerebral hemodynamic indicators and surgical clinical data, a multi-dimensional feature vector is constructed.

[0012] The aforementioned feature vectors are input into a hybrid neural network model composed of a graph convolutional network and a long short-term memory network, which outputs predicted risk values ​​for postoperative cerebral edema, cognitive impairment, and white matter damage.

[0013] In one feasible implementation, the aforementioned hybrid neural network model incorporates pre-trained weights obtained from Alzheimer's disease lymphoma studies through transfer learning, and is fine-tuned based on postoperative patient images and follow-up data.

[0014] In one feasible implementation, it further includes:

[0015] Intraoperative magnetic resonance imaging terminal to monitor changes in the diffusion index and contrast-enhanced perivascular space index along the perivascular space direction in real time.

[0016] When the diffusion index along the perivascular space direction decreases beyond a preset threshold or the contrast-enhanced perivascular space index clearance delay exceeds a set standard deviation range, an audio-visual warning is triggered and intraoperative intervention suggestions are output.

[0017] In one feasible implementation, the aforementioned key anatomical regions are automatically segmented using a U-Net neural network, and the aforementioned joint anatomical regions include the lateral ventricle body and the basal ganglia.

[0018] In one feasible implementation, it further includes:

[0019] Functional index curves were updated based on longitudinal follow-up magnetic resonance imaging data during the postoperative stage;

[0020] The risk of long-term brain injury can be predicted by using the aforementioned functional index curves in conjunction with Monte Carlo simulations.

[0021] In one feasible implementation, the prediction of long-term brain injury risk using the aforementioned functional index curves combined with Monte Carlo simulations includes:

[0022] The above functional index curves are denoised, outlier removed, and smoothed to generate individualized functional recovery curves.

[0023] Intraoperative operation parameters, physiological baseline variables and blood perfusion parameters are introduced to construct a set of perturbation factors. Based on this set of perturbation factors, the above functional recovery curves are simulated and perturbed to generate multiple sets of fitted curves.

[0024] Monte Carlo simulation was performed based on the above set of multiple fitted curves to obtain the probability distribution and statistical characteristics of postoperative functional status.

[0025] The probability distribution and statistical characteristics of the postoperative functional status are input into the trained brain injury risk prediction model, and the brain injury risk score and risk classification label of the target patient are output.

[0026] In one feasible implementation, the above-mentioned brain injury risk prediction model includes:

[0027] The multi-scale perturbation attention extraction module is used to receive the probability distribution and statistical characteristics of the above-mentioned postoperative functional states, and construct a multi-scale attention map based on the perturbation amplitude and the key inflection point position in the time series, so as to extract dynamic sensitive features that are highly correlated with the risk of injury in the process of brain functional evolution.

[0028] The brain region functional mapping module is used to map the above dynamic sensitive features to a preset human brain anatomical structure atlas;

[0029] The gated temporal memory modeling module includes an enhanced bidirectional long short-term memory network, which is used to perform temporal modeling on the joint features from the aforementioned attention extraction module and graph encoding module, learning the trend changes and phased fluctuations in the function recovery process.

[0030] The risk decoding and output module includes a parallel risk regression branch and a risk classification branch. The regression branch is used to output a continuous brain injury risk score for the target patient, and the classification branch is used to output a discrete risk level label. The output results of the two branches are dynamically weighted and fused through an attention gating mechanism.

[0031] In one feasible implementation, the aforementioned preset human brain anatomical structure atlas uses functional regions obtained from magnetic resonance imaging scans as nodes, and the edge weights between nodes are determined based on the structural connection strength and signal correlation.

[0032] In one feasible implementation, the brain region functional map encoding module introduces intraoperative operation path map as external auxiliary map information to model the graph structure coupling relationship between changes in lymphatic function and intraoperative traumatic stimulation.

[0033] Secondly, the present invention also proposes a lymphatic system-like magnetic resonance imaging system for predicting postoperative brain injury assessment, comprising:

[0034] The acquisition unit is used to acquire multimodal magnetic resonance imaging data of the joint anatomical region of the target subject, wherein the multimodal magnetic resonance imaging data includes diffusion tensor imaging data, dynamic contrast-enhanced imaging data, and arterial spin-labeled magnetic resonance imaging sequence.

[0035] The extraction unit is used to extract magnetic resonance imaging functional parameters based on the above-mentioned multimodal magnetic resonance imaging image data, wherein the above-mentioned magnetic resonance imaging functional parameters include diffusion index along the perivascular space direction, contrast enhancement perivascular space index, and perivascular space volume ratio.

[0036] The construction unit is used to integrate the above-mentioned magnetic resonance imaging functional parameters, brain region white matter fiber integrity parameters, cerebral hemodynamic indicators and surgical clinical data to construct a multi-dimensional feature vector.

[0037] The output unit is used to input the above feature vectors into a hybrid neural network model composed of a graph convolutional network and a long short-term memory network, and output the risk prediction values ​​of postoperative cerebral edema, cognitive impairment and white matter damage.

[0038] In summary, this invention, by fusing DTI, DCE-MRI, and ASL sequences to construct the ALPS and CEPS indices, achieves for the first time quantitative measurement of perivascular space fluid diffusion efficiency and metabolite clearance rate, avoiding invasive procedures such as intrathecal injection and significantly improving repeatability and patient acceptance. Deeply fusing lymphoid system functional indicators with white matter fiber integrity (FA value), cerebral hemodynamics (CBF, CBV), and intraoperative data (such as traction duration and hemorrhage volume), and inputting them into a hybrid neural network model composed of graph convolutional networks and LSTM, effectively establishes a causal relationship between intraoperative trauma, neural structural damage, and lymphoid function. Using longitudinal MRI data to track the evolution trend of postoperative functional indicators, combined with Monte Carlo simulation and multi-scale perturbation modeling, it outputs individualized risk distribution and trend prediction, significantly improving the early warning ability for complications such as cerebral edema, cognitive impairment, and white matter damage within 24–48 hours postoperatively, outperforming traditional static structural indicator prediction methods. Introducing an automatic segmentation model based on U-Net for brain region ROI identification significantly reduces the need for manual intervention and improves efficiency and standardization. Simultaneously, CEPS is proposed as a dedicated quantitative indicator for acute postoperative lymphoid dysfunction, and a cross-validation system is established in beagle dog and human data to enhance the model's generalization and clinical adaptability. This method supports intraoperative MRI dynamic monitoring of ALPS and CEPS changes. When abnormal fluctuations in the indicators exceed the threshold, an audible and visual warning is automatically triggered, providing surgeons with timely intervention suggestions (such as reducing traction force and adjusting perfusion pressure) to achieve proactive control of brain injury risk. In summary, this invention overcomes the limitations of existing assessment methods, such as strong reliance on structural images, lack of functional assessment, and insufficient predictive accuracy, achieving precise prediction and intervention of postoperative brain injury based on lymphoid functional imaging.

[0039] Other advantages, objectives and features of this application will be apparent in part from the description which follows, and in part from what those skilled in the art will understand through study and practice of this application. Attached Figure Description

[0040] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0041] Figure 1 A flowchart illustrating a method for predicting postoperative brain injury using magnetic resonance imaging of a lymphatic system, as provided in an embodiment of this application.

[0042] Figure 2 This is a flowchart illustrating a method for assessing brain injury during magnetic resonance imaging of a lymphoid system, as provided in an embodiment of this application.

[0043] Figure 3 A flowchart illustrating a method for predicting long-term brain injury risk provided in an embodiment of this application;

[0044] Figure 4 A flowchart illustrating a method for predicting long-term brain injury risk using the aforementioned functional index curves combined with Monte Carlo simulation, provided as an embodiment of this application.

[0045] Figure 5 This is a structural schematic diagram of a lymphatic system-like magnetic resonance imaging system for predicting postoperative brain injury assessment, provided in an embodiment of this application. Detailed Implementation

[0046] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this application will now be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0047] Please refer to the figure. Figure 1 This application provides a structural schematic diagram of a method for predicting postoperative brain injury assessment using magnetic resonance imaging of a lymphoid system, which may specifically include:

[0048] S110. Acquire multimodal magnetic resonance imaging data of the joint anatomical region of the target subject, wherein the multimodal magnetic resonance imaging data includes diffusion tensor imaging data, dynamic contrast-enhanced imaging data and arterial spin-labeled magnetic resonance imaging sequence.

[0049] S120. Based on the above multimodal magnetic resonance imaging data, extract magnetic resonance imaging functional parameters, wherein the above magnetic resonance imaging functional parameters include diffusion index along the perivascular space direction, contrast enhancement perivascular space index, and perivascular space volume ratio.

[0050] S130. The above-mentioned magnetic resonance imaging functional parameters, brain region white matter fiber integrity parameters, cerebral hemodynamic indicators and surgical clinical data are integrated to construct a multi-dimensional feature vector.

[0051] S140. Input the above feature vectors into a hybrid neural network model composed of a graph convolutional network and a long short-term memory network, and output the risk prediction values ​​of postoperative cerebral edema, cognitive impairment and white matter damage.

[0052] For example, magnetic resonance imaging data of key anatomical regions (such as the body of the lateral ventricle and the basal ganglia) of the target subject are acquired before and after surgery. The imaging data includes: diffusion tensor imaging (DTI) data, used to quantify the diffusion characteristics of water molecules along a specific direction; dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) sequences, used to observe the dynamic changes of gadolinium in the perivascular space (PVS); and arterial spin labeling (ASL) sequences, used to measure the perfusion blood flow intensity of the brain region.

[0053] Based on the above multimodal images, lymphatic function-related parameters are constructed, including but not limited to:

[0054] 1. The ALPS index (Along Perivascular Space index) is used to represent the degree of opening of the PVS pathway, and is defined as follows:

[0055]

[0056] Among them, D x,proj D represents the diffusion coefficient along the direction of the projected fiber. y,assoc This represents the diffusion coefficient of the vertically joined fibers. A lower ratio indicates greater resistance to lymphatic pathways.

[0057] 2. The CEPS index (Contrast-enhanced Perivascular Space index) is used to measure the efficiency of metabolic waste removal in the perivascular space:

[0058]

[0059] Among them, AUC decay is the area under the curve for the contrast agent signal attenuation segment, and TTP is the signal peak time.

[0060] 3. The PVS volume ratio is calculated by segmenting and analyzing T2-weighted images to determine the volume percentage of PVS within the target region.

[0061]

[0062] The above functional parameters were fused with the following data to construct a multidimensional feature vector: white matter fiber integrity indicators (such as FA value (Fractional Anisotropy), cerebral hemodynamic indicators (such as CBF (Cerebral Blood Flow) and CBV (Cerebral Blood Volume)), and surgery-related clinical information (such as intraoperative blood loss, traction time, and operation area).

[0063] Construct a multidimensional feature vector in the following form:

[0064] X = [ALPS, CEPS, PVS] ratio ,FA,CBF,T surgery ,…]

[0065] The above feature vectors are input into a hybrid neural network model, which consists of the following two substructures:

[0066] 1. The Graph Convolutional Network (GCN) module constructs a node graph structure based on the functional connectivity map of the human brain. It defines the node edge weights by structural connectivity weights and functional relevance, thereby realizing the spatial understanding of lymphoid functional differences between brain regions.

[0067] 2. The Long Short-Term Memory (LSTM) module is used to model the trends of MRI parameters and intraoperative variables over time, and extract dynamic sequence features to predict pathological changes.

[0068] The final output is the following prediction result: Postoperative cerebral edema risk Possibility of cognitive impairment Prediction of white matter fiber damage

[0069] This invention, by fusing DTI, DCE-MRI, and ASL sequences, constructs the ALPS and CEPS indices, achieving for the first time a quantitative measurement of perivascular space fluid diffusion efficiency and metabolite clearance rate. This avoids invasive procedures such as intrathecal injection, significantly improving repeatability and patient acceptance. Deeply fusing lymphoid system functional indicators with white matter fiber integrity (FA value), cerebral hemodynamics (CBF, CBV), and intraoperative data (such as traction duration and hemorrhage volume) and inputting them into a hybrid neural network model composed of graph convolutional networks and LSTM, effectively establishes a causal relationship between intraoperative trauma, neural structural damage, and lymphoid function. Using longitudinal MRI (Magnetic Resonance Imaging) data to track the evolution trend of postoperative functional indicators, combined with Monte Carlo simulation and multi-scale perturbation modeling, it outputs individualized risk distribution and trend prediction, significantly improving the early warning ability for complications such as cerebral edema, cognitive impairment, and white matter damage within 24–48 hours postoperatively, outperforming traditional static structural indicator prediction methods. This invention introduces an automatic segmentation model based on U-Net for brain region (ROI) identification, significantly reducing the need for manual intervention and improving efficiency and standardization. Simultaneously, CEPS is proposed as a dedicated quantitative indicator for acute postoperative lymphoid dysfunction, and a cross-validation system is established on beagle dog and human data to enhance the model's generalization and clinical adaptability. This method supports intraoperative MRI dynamic monitoring of ALPS and CEPS changes. When abnormal fluctuations in indicators exceed thresholds, an audible and visual warning is automatically triggered, providing surgeons with timely intervention suggestions (such as reducing traction force and adjusting perfusion pressure), thus achieving proactive control of brain injury risk. In summary, this invention overcomes the limitations of existing assessment methods, such as strong reliance on structural images, lack of functional assessment, and insufficient predictive accuracy, achieving precise prediction and intervention of postoperative brain injury based on lymphoid functional imaging.

[0070] In one feasible implementation, the aforementioned hybrid neural network model incorporates pre-trained weights obtained from Alzheimer's disease lymphoma studies through transfer learning, and is fine-tuned based on postoperative patient images and follow-up data.

[0071] For example, in the lymphatic system-like magnetic resonance imaging prediction method for postoperative brain injury assessment described in this invention, the hybrid neural network model is initialized using a transfer learning strategy, thereby significantly improving the model's predictive performance and stability on small sample postoperative patient datasets.

[0072] Specifically, the pre-trained model is derived from multimodal magnetic resonance imaging data from the Alzheimer's disease lymphoid cohort, and the training objective is to identify the correlation features between lymphoid dysfunction and cognitive decline. This pre-trained model includes optimized weight parameters for a graph convolutional network (GCN) and a bidirectional long short-term memory network (Bi-LSTM), which can effectively extract the potential coupling patterns between brain region structural connectivity and temporal dynamic changes in function.

[0073] Before being applied to the task of predicting postoperative brain injury risk, the model weights will undergo the following fine-tuning process:

[0074] Define the pre-trained model parameter set as:

[0075]

[0076] in: W represents the weights of the l-th layer graph convolutional network. BiLSTM For bidirectional LSTM network parameters; W output These are the weights for the pre-trained classification output layer.

[0077] This parameter set is used to initialize the backbone structure of the current model, preserving its ability to represent lymphoid functional map structures.

[0078] Introduce a dual loss structure, including the main task loss and the transfer regularization term:

[0079]

[0080] in: Predict the loss of the main task for the current brain injury (such as mean squared error or cross-entropy); λ is the transfer regularization term, used to constrain the parameters of the new model from deviating from the pre-trained model; λ is the weight coefficient, controlling the degree of influence of the transfer regularization term on the total loss; θ new The model parameters were fine-tuned based on postoperative patient data.

[0081] The training inputs received by the model in the target task include: lymphoid system MRI parameters: ALPS index, CEPS index, PVS volume ratio; white matter integrity indicators: FA (fractional anisotropy) value, MD (mean diffusivity) value; blood perfusion parameters: CBF (cerebral blood flow), CBV (cerebral blood volume), TTP (time to peak); surgery-related variables: intraoperative traction duration, hemorrhage, anesthesia duration, etc.; follow-up labels: whether cerebral edema, cognitive impairment, or other clinical records occurred within 3 months after surgery.

[0082] The model performs forward propagation using the aforementioned multimodal input features to extract key graph structure and temporal dynamic features, and outputs a damage risk value. During training, backpropagation is used to jointly optimize the model with the aforementioned loss function, enabling the model to transfer and adapt from the Alzheimer's disease task to the postoperative damage prediction task.

[0083] In one feasible implementation, such as Figure 2 As shown, Figure 2 This is a schematic flowchart of a lymphoid system-like magnetic resonance imaging intraoperative brain injury assessment system provided in an embodiment of this application. The above method further includes:

[0084] S210. Real-time monitoring of changes in the diffusion index and contrast-enhanced perivascular space index along the perivascular space direction in an intraoperative magnetic resonance imaging terminal.

[0085] S220. When the diffusion index along the perivascular space direction is detected to decrease beyond the preset threshold or the contrast-enhanced perivascular space index clearance delay exceeds the set standard deviation range, an audio-visual warning is triggered and intraoperative intervention suggestions are output.

[0086] For example, in one feasible implementation, the present invention also includes an intraoperative real-time monitoring and early warning mechanism for lymphoid function to improve the immediacy and accuracy of intraoperative intervention. This function is achieved through integration with an intraoperative MRI terminal, combining diffusion tensor imaging (DTI) and dynamic contrast-enhanced MRI (DCE-MRI) technologies to collect and analyze lymphoid function status indicators of the perivascular space (PVS) in real time.

[0087] During neurosurgical procedures, the following two types of dynamic indicators are obtained using an intraoperative high-field MRI system:

[0088] 1. Perivascular space diffusion index (ALPS index): This index is based on DTI imaging to calculate the principal diffusion coefficient D of the diffusion tensor in a specific white matter fiber direction (such as the x-axis of the projection fiber). x Its perpendicular direction (joint fiber y-axis) D y The ratio:

[0089]

[0090] When this ratio decreases, it indicates that the PVS functional pathway is restricted and the lymphoid clearance capacity is weakened.

[0091] 2. Contrast-enhanced PVS index (CEPS index): Based on DCE-MRI imaging, after injection of a low dose of gadolinium, its concentration change curve in the perivenous space was dynamically tracked, and the following parameters were extracted: time-to-peak (TTP), area under the curve (AUC), and CEPS index were defined as follows:

[0092]

[0093] The smaller the value, the slower the clearance rate, indicating an increased risk of metabolic waste retention.

[0094] The intraoperative system compares real-time ALPS and CEPS values ​​with the individual's preoperative baseline to determine whether one of the following warning trigger conditions is met: ALPS decreases by more than 15% from the preoperative baseline or TTP delay in CEPS exceeds two standard deviations from the preoperative mean.

[0095] If the above threshold is triggered, the system will immediately issue an audible and visual alarm signal through the intraoperative navigation terminal, and mark high-risk areas (such as areas with low ALPS values ​​or delayed CEPS values) on the visualization interface, and output intervention suggestions, including but not limited to: reducing the brain traction amplitude, temporarily stopping the traction operation, reducing local perfusion pressure, or strengthening intraoperative dehydration treatment.

[0096] The method proposed in this embodiment forms a complete chain from real-time acquisition of MRI parameters and index calculation to alarm output. Through dynamic monitoring of ALPS and CEPS, it can identify potential PVS dysfunction in advance, reduce the risk of postoperative cerebral edema and cognitive impairment, provide surgeons with data-driven adjustment basis, and improve the safety and individualized precision of surgical operations.

[0097] In one feasible implementation, the aforementioned key anatomical regions are automatically segmented using a U-Net neural network, and the aforementioned joint anatomical regions include the lateral ventricle body and the basal ganglia.

[0098] For example, this invention employs deep learning methods to automatically segment key anatomical regions of the brain, improving the accuracy and efficiency of extracting MRI functional parameters of the lymphoid system. Specifically, a U-Net neural network structure is introduced to accurately identify and segment key lymphoid-related anatomical regions, including the lateral ventricle body and basal ganglia, in magnetic resonance imaging images.

[0099] The system first selects structural images containing T1, T2, and FLAIR sequences, along with corresponding manually annotated regions, from preoperative or postoperative MRI scan data as the training dataset. The manually annotated regions of interest (ROIs) include, but are not limited to: the body of the lateral ventricle (near the choroid plexus, with significant convergence of PVS) and the basal ganglia (a region with dense PVS distribution, intersecting with various white matter fiber tracts).

[0100] The U-Net model uses the following structural parameters:

[0101] The encoder contains 4 levels of convolutional-pooling layers, with each convolutional kernel having a size of 3×3 and the number of channels increasing from 32 to 256.

[0102] Decoder: Performs upsampling operations at the corresponding level and skips connections to preserve spatial details;

[0103] Loss function: Combines Dice coefficient loss and cross-entropy loss to balance contour accuracy and pixel classification accuracy;

[0104] Optimizer: Adam, initial learning rate 1×10 -4 Dynamic decay.

[0105] During the deployment phase, structural MRI images (including T1 or T2 sequences) of the target patient are input into the trained U-Net model. The system automatically outputs mask images of the following anatomical regions: lateral ventricle body ROI mask and basal ganglia ROI mask.

[0106] The above mask image is registered and resampled with functional MRI parameter images (such as anisotropy images generated by DTI and DCE-MRI enhancement curves) to extract the corresponding water molecule diffusion parameters (for ALPS index calculation), gadolinium concentration change parameters (for CEPS index calculation) and PVS morphological features (such as volume, thickness, and extension direction) of the region.

[0107] The U-Net model proposed in this invention has been trained with a large amount of preoperative and postoperative data, exhibiting strong adaptability to structural variations. Its automatic segmentation achieves a Dice coefficient exceeding 0.90, and the automatic segmentation process for a single MRI data sample takes less than 2 seconds, significantly reducing time compared to manual operation (which requires 5-10 minutes). It ensures consistency in ROI extraction under multi-center, multi-sequence scanning conditions, providing a unified basis for subsequent parameter comparison and model input. It avoids the poor reproducibility issues introduced by manual zoning and supports stable extraction of lymphoid marker curves during longitudinal follow-up.

[0108] In one feasible implementation, it further includes:

[0109] S310. Update the functional index curve based on longitudinal follow-up magnetic resonance imaging data in the postoperative stage;

[0110] S320. Predict the risk of long-term brain injury using the above functional index curves combined with Monte Carlo simulation.

[0111] In one feasible implementation, step S320 above predicts the long-term risk of brain injury using the aforementioned functional index curves and in conjunction with Monte Carlo simulations, including:

[0112] S3201. The above functional index curves are denoised, outlier removed and smoothed to generate individualized functional recovery curves.

[0113] S3202. Introduce intraoperative operation parameters, physiological baseline variables and blood perfusion parameters to construct a set of perturbation factors. Based on this set of perturbation factors, simulate and perturb the above-mentioned functional recovery curves to generate multiple sets of fitting curves.

[0114] S3203. Input the probability distribution and statistical characteristics of the above postoperative functional status into the trained brain injury risk prediction model, and output the brain injury risk score and risk classification label of the target patient.

[0115] For example, this invention provides a method for predicting long-term brain injury risk based on postoperative longitudinal follow-up magnetic resonance imaging data. This method aims to dynamically assess the recovery process of lymphoid function and achieve a prospective grading assessment of long-term brain injury risk. It fully integrates imaging functional parameters, physiological variables, and surgical perturbation factors, combining Monte Carlo simulations and deep learning models to improve the individualized adaptability and reliability of the prediction.

[0116] First, multimodal magnetic resonance imaging (MRI) data of the target patients were collected at different time points after surgery, and indicators closely related to the function of the lymphoid system were extracted from them, such as the perivascular diffusion index (ALPS), the contrast-enhanced perivascular space index (CEPS), and the rate of change in lymphoid channel volume. These indicators constituted a longitudinal functional index curve over time to reflect the postoperative functional recovery trajectory.

[0117] Next, the system processes the raw curve data, including noise suppression, outlier removal, and curve smoothing, ultimately generating an individualized functional recovery curve C(t). To simulate the potential impact of different physiological states and surgical procedures on lymphoid regeneration, the system incorporates intraoperative operational parameters (such as resection extent and blood loss), baseline physiological variables (such as age and cognitive level), and blood perfusion parameters to construct a set of perturbation factors.

[0118] Subsequently, the system performs nonlinear perturbation simulation on the functional recovery curve C(t) based on the perturbation factor set, generating multiple sets of fitted curves in the following form:

[0119]

[0120] Where, ε k (t) represents the relationship with the disturbance factor f k The relevant time-varying offset, The variance of the disturbance intensity is used to generate M fitting curves.

[0121] Next, the system conducts Monte Carlo simulations based on these perturbation curve sets, statistically analyzes the functional change characteristics under different curves, and extracts statistical features such as peak recovery rate, recovery lag time, early clearance rate, and fluctuation amplitude to construct a probability distribution vector S representing the individual's postoperative functional status.

[0122] [s1,s2,…,s k ].

[0123] Finally, the statistical feature vector is input into a pre-trained brain injury risk prediction model. This model consists of multiple functional modules, including:

[0124] Multi-scale perturbation attention extraction module: assigns different attention weights to time-series inflection points and perturbation amplitude-sensitive regions in the simulation curve to extract key dynamic features;

[0125] Brain region functional map convolutional coding module: embeds the above dynamic features into a preset brain region functional map, and performs graph modeling by combining structural connectivity strength and functional signal differences;

[0126] Gated time memory modeling module: Based on an enhanced bidirectional LSTM network, time series modeling of graph encoding output is performed to capture and recover trends and fluctuation rhythms;

[0127] Risk decoding and output module: includes a regression branch (outputting continuous risk scores) and a hierarchical classification branch (outputting low / medium / high risk level labels), and merges the results of the two output paths through an attention mechanism.

[0128] Ultimately, the system can output the postoperative brain injury continuous risk score R for the target patient. s The values ​​∈[0,1] and discrete risk level labels L∈{low risk, medium risk, high risk} provide precise support for postoperative follow-up intervention.

[0129] This implementation method overcomes the shortcomings of traditional single-index prediction models in terms of timeliness and specificity by introducing multiple sets of perturbation simulation and probabilistic modeling mechanisms. By combining multimodal imaging, surgical parameters and follow-up data, it realizes dynamic modeling and individual risk assessment of the lymphoid function recovery process, effectively improving the ability to identify early postoperative brain injury and the level of precise intervention.

[0130] In one feasible implementation, the above-mentioned brain injury risk prediction model includes:

[0131] The multi-scale perturbation attention extraction module is used to receive the probability distribution and statistical characteristics of the above-mentioned postoperative functional states, and construct a multi-scale attention map based on the perturbation amplitude and the key inflection point position in the time series, so as to extract dynamic sensitive features that are highly correlated with the risk of injury in the process of brain functional evolution.

[0132] The brain region functional mapping module is used to map the above dynamic sensitive features to a preset human brain anatomical structure atlas;

[0133] The gated temporal memory modeling module includes an enhanced bidirectional long short-term memory network, which is used to perform temporal modeling on the joint features from the aforementioned attention extraction module and graph encoding module, learning the trend changes and phased fluctuations in the function recovery process.

[0134] The risk decoding and output module includes a parallel risk regression branch and a risk classification branch. The regression branch is used to output a continuous brain injury risk score for the target patient, and the classification branch is used to output a discrete risk level label. The output results of the two branches are dynamically weighted and fused through an attention gating mechanism.

[0135] In one feasible implementation, the aforementioned preset human brain anatomical structure atlas uses functional regions obtained from magnetic resonance imaging scans as nodes, and the edge weights between nodes are determined based on the structural connection strength and signal correlation.

[0136] In one feasible implementation, the brain region functional map encoding module introduces intraoperative operation path map as external auxiliary map information to model the graph structure coupling relationship between changes in lymphatic function and intraoperative traumatic stimulation.

[0137] For example, the present invention provides an advanced brain injury risk prediction model, the core of which is to introduce a multi-scale perturbation modeling mechanism and a graph neural network architecture, combined with human brain anatomy knowledge and time modeling capabilities, to achieve a high-precision mapping between postoperative functional status and potential brain injury risk.

[0138] First, in the multi-scale perturbation attention extraction module, an attention weight calculation formula is introduced:

[0139]

[0140] in, denoted as the attention score at time point i in layer s, representing the importance of damage prediction at that moment during functional recovery; It is the state vector at the i-th time point under the disturbance simulation, reflecting the response intensity of the function curve under different disturbance conditions; The inflection point indicator corresponding to this time point, such as derivative mutation or trend reversal, is usually obtained through curve analysis; f(·) is the attention scoring function, commonly used forms include dot product, bilinear mapping or feedforward neural network, and its function is to measure the contribution of each time point to the overall trend.

[0141] Next, in the brain region functional mapping module, the graph structure connection weights between nodes are calculated as follows:

[0142] w ij =λ1·Corr(x) i ,x j )+λ2·FA ij

[0143] In addition, w ij Corr(x) represents the connection strength between brain region i and brain region j in the functional map. i ,x j ) is the signal correlation coefficient between these two brain regions in time-series MRI data, measuring the degree of synchronization of their functional activities; FA ij It is based on the average anisotropy extracted from diffusion tensor imaging, which reflects the integrity and connectivity density of white matter fibers between the two brain regions; λ1 and λ2 are adjustable weighting coefficients used to balance the proportion of structural information and functional relevance in graph modeling.

[0144] Subsequently, in the gated temporal memory modeling module, an enhanced bidirectional long short-term memory network (Bi-LSTM) is used for temporal modeling, and its basic structure is as follows:

[0145] f t =σ(W f x t +U f h t-1 +b f )

[0146] i t =σ(W i x t +U i h t-1 +b i )

[0147] o t =σ(W o x t +U o h t-1 +b o )

[0148] c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +U c h t-1 +b c )

[0149] h t =o t ⊙tanh(c t )

[0150] In the above formula, x t It is a joint feature of the input at the current time point, containing multi-source information such as perturbation response and spectral coding; h t-1 It is the hidden state of the previous moment; f t i t o t These are the forget gate, input gate, and output gate, used to control the retention and updating of information flow; c t The state of the memory cell is fused and depends on the current input for a long time; σ(·) is the Sigmoid function, and ⊙ represents the element-wise multiplication operation.

[0151] Finally, in the risk decoding and output module, the following formula is used to achieve the fusion of continuous and hierarchical outputs:

[0152]

[0153] in, The final comprehensive score representing the risk of brain injury; R s The continuous risk regression value output by the model, typically located in the interval [0, 1], is used to characterize the degree of risk; L s γ is a discrete risk level label mapping value (e.g., low = 0.2, medium = 0.5, high = 0.8, etc.); γ is a weighting factor dynamically calculated by the attention gating mechanism, which adjusts the fusion ratio of the two outputs according to the model's prediction confidence and uncertainty for the current sample, thereby taking into account both refined prediction and level judgment.

[0154] Through the synergistic operation of the above parameters and formulas, this brain injury risk prediction model can fully integrate dynamic perturbation sensitivity, spatial structural connectivity, and temporal evolution trends, effectively capturing the potential contribution of postoperative lymphoid function changes to brain injury risk, and improving the sensitivity and clinical applicability of the prediction.

[0155] The second aspect, such as Figure 5 As shown, Figure 5 This is a structural schematic diagram of a lymphoid system-like magnetic resonance imaging (MRI) system for predicting postoperative brain injury assessment, provided in an embodiment of this application. The present invention also proposes a lymphoid system-like MRI system for predicting postoperative brain injury assessment, comprising:

[0156] The acquisition unit 21 is used to acquire multimodal magnetic resonance imaging data of the joint anatomical region of the target subject, wherein the multimodal magnetic resonance imaging data includes diffusion tensor imaging data, dynamic contrast enhancement imaging data and arterial spin labeling magnetic resonance imaging sequence.

[0157] Extraction unit 22 is used to extract magnetic resonance imaging functional parameters based on the above-mentioned multimodal magnetic resonance imaging image data, wherein the above-mentioned magnetic resonance imaging functional parameters include diffusion index along the perivascular space direction, contrast enhancement perivascular space index, and perivascular space volume ratio.

[0158] Construction unit 23 is used to integrate the above-mentioned magnetic resonance imaging functional parameters, brain region white matter fiber integrity parameters, cerebral hemodynamic indicators and surgical clinical data to construct a multi-dimensional feature vector;

[0159] Output unit 24 is used to input the above feature vectors into a hybrid neural network model composed of graph convolutional network and long short-term memory network, and output the risk prediction values ​​of postoperative cerebral edema, cognitive impairment and white matter damage.

[0160] It is understood that the present invention also proposes a lymphoid system magnetic resonance imaging prediction system for assessing postoperative brain injury, which can also perform the method described in any of the first aspects.

[0161] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting postoperative brain injury using magnetic resonance imaging of a lymphatic system, characterized in that, include: Acquire multimodal magnetic resonance imaging data of the joint anatomical region of the target subject, wherein the multimodal magnetic resonance imaging data includes diffusion tensor imaging data, dynamic contrast-enhanced imaging data, and arterial spin-labeled magnetic resonance imaging sequence; Based on the multimodal magnetic resonance imaging data, magnetic resonance imaging functional parameters are extracted, including the diffusion index along the perivascular space direction, the contrast enhancement perivascular space index, and the perivascular space volume ratio. The magnetic resonance imaging functional parameters, brain region white matter fiber integrity parameters, cerebral hemodynamic indicators, and surgical clinical data are fused together to construct a multi-dimensional feature vector. The feature vector is input into a hybrid neural network model composed of a graph convolutional network and a long short-term memory network, which outputs the risk prediction values ​​of postoperative cerebral edema, cognitive impairment and white matter damage.

2. The method for predicting postoperative brain injury using magnetic resonance imaging of a lymphatic system according to claim 1, characterized in that, The hybrid neural network model incorporates pre-trained weights obtained from Alzheimer's disease lymphoma studies through transfer learning, and is fine-tuned based on postoperative patient images and follow-up data.

3. The method for predicting postoperative brain injury using magnetic resonance imaging of a lymphoid system according to claim 1, characterized in that, Also includes: Intraoperative magnetic resonance imaging terminal to monitor changes in the diffusion index and contrast-enhanced perivascular space index along the perivascular space direction in real time. When the diffusion index along the perivascular space direction decreases beyond a preset threshold or the contrast-enhanced perivascular space index clearance delay exceeds a set standard deviation, an audible and visual warning is triggered and intraoperative intervention suggestions are output.

4. The method for predicting postoperative brain injury using magnetic resonance imaging of a lymphoid system according to claim 1, characterized in that, The key anatomical regions are automatically segmented using a U-Net neural network, and the joint anatomical regions include the lateral ventricle body and the basal ganglia.

5. The method for predicting postoperative brain injury using magnetic resonance imaging of a lymphoid system according to claim 1, characterized in that, Also includes: Functional index curves were updated based on longitudinal follow-up magnetic resonance imaging data during the postoperative stage; The long-term risk of brain injury is predicted using the aforementioned functional index curves combined with Monte Carlo simulations.

6. The method for predicting postoperative brain injury using magnetic resonance imaging of a lymphoid system according to claim 5, characterized in that, The prediction of long-term brain injury risk using the functional index curve in conjunction with Monte Carlo simulation includes: The functional index curves are denoised, outlier removed, and smoothed to generate individualized functional recovery curves. Intraoperative operation parameters, physiological baseline variables, and blood perfusion parameters are introduced to construct a set of perturbation factors. Based on this set of perturbation factors, the functional recovery curve is simulated and perturbed to generate multiple sets of fitted curves. Monte Carlo simulation was performed based on the set of multiple fitted curves to obtain the probability distribution and statistical characteristics of postoperative functional status. The probability distribution and statistical characteristics of the postoperative functional status are input into the trained brain injury risk prediction model, and the brain injury risk score and risk classification label of the target patient are output.

7. The method for predicting postoperative brain injury using magnetic resonance imaging of a lymphoid system according to claim 6, characterized in that, The brain injury risk prediction model includes: The multi-scale perturbation attention extraction module is used to receive the probability distribution and statistical characteristics of the postoperative functional state, and construct a multi-scale attention map based on the perturbation amplitude and the key inflection point position in the time series, so as to extract dynamic sensitive features that are highly correlated with the risk of injury during the brain function evolution process. The brain region functional mapping module is used to map the dynamic sensitive features to a preset human brain anatomical structure atlas; The gated temporal memory modeling module includes an enhanced bidirectional long short-term memory network, used to perform temporal modeling of the joint features from the attention extraction module and the graph encoding module, learning the trend changes and phased fluctuations in the function recovery process; The risk decoding and output module includes a parallel risk regression branch and a risk classification branch. The regression branch is used to output a continuous brain injury risk score for the target patient, and the classification branch is used to output a discrete risk level label. The output results of the two branches are dynamically weighted and fused through an attention gating mechanism.

8. The method for predicting postoperative brain injury using magnetic resonance imaging of a lymphoid system according to claim 7, characterized in that, The preset human brain anatomical structure atlas uses functional regions obtained from magnetic resonance imaging scans as nodes, and the edge weights between nodes are determined based on the structural connection strength and signal correlation.

9. The method for predicting postoperative brain injury using magnetic resonance imaging of a lymphoid system according to claim 1, characterized in that, The brain region functional map encoding module introduces intraoperative operation path map as external auxiliary map information to model the graph structure coupling relationship between changes in lymphatic function and intraoperative traumatic stimulation.

10. A lymphatic system-like magnetic resonance imaging system for predicting postoperative brain injury assessment, characterized in that, include: The acquisition unit is used to acquire multimodal magnetic resonance imaging data of the joint anatomical region of the target subject, wherein the multimodal magnetic resonance imaging data includes diffusion tensor image data, dynamic contrast enhancement image data, and arterial spin labeling magnetic resonance imaging sequence. The extraction unit is used to extract magnetic resonance imaging functional parameters based on the multimodal magnetic resonance imaging image data, wherein the magnetic resonance imaging functional parameters include diffusion index along the perivascular space direction, contrast enhancement perivascular space index, and perivascular space volume ratio. The construction unit is used to fuse the magnetic resonance imaging functional parameters, brain region white matter fiber integrity parameters, cerebral hemodynamic indicators and surgical clinical data to construct a multi-dimensional feature vector. The output unit is used to input the feature vector into a hybrid neural network model composed of a graph convolutional network and a long short-term memory network, and output the risk prediction values ​​of postoperative cerebral edema, cognitive impairment and white matter damage.

Images