Neurosurgical operation scheme optimization method based on digital twinning and causal inference

Abstract

The invention discloses a neurosurgical operation scheme optimization method based on digital twinning and causal inference, belongs to the technical field of medical equipment and information processing, is suitable for radiofrequency thermocoagulation (RF-TC) treatment drug-refractory epilepsy (DRE) operation under the guidance of stereotactic electroencephalogram (SEEG), and comprises the following steps: constructing a patient specific calculation model (PSCM); integrating multi-modal data to establish a digital twinborn model covering anatomical geometry, organizational physical attributes, neurodynamics and radio frequency thermal coagulation multi-physics field coupling; an optimization scheme is generated based on a causal inference generative adversarial network (CI-GAN), the causal effect of candidate schemes is quantitatively evaluated through PSCM, and the curative effect and safety are optimized through adversarial training; fine adjustment of a doctor on an AI scheme is achieved through an interactive platform, a final scheme is determined by combining real-time feedback quantitative indexes of the PSCM, objective identification of an epileptic source area, precise prediction of lesion and man-machine collaborative optimization are achieved, and scientificity and feasibility of an operation scheme are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical equipment and information processing technology, more particularly to a neurosurgical operation scheme optimization method based on digital twinning and causal inference. BACKGROUND

[0002] Drug-resistant epilepsy (DRE) is a severe neurological disease affecting millions of people worldwide. For DRE patients with clear and localized epileptic foci, surgical resection or lesion of epileptic foci is one of the most effective treatment methods. In recent years, radiofrequency thermocoagulation (RF-TC) guided by stereotactic electroencephalography (SEEG) has become an important treatment option due to its small trauma, strong repeatability, and fewer complications.

[0003] The conventional clinical process of this technology includes: 1) implanting SEEG electrodes to monitor brain electrical activity and locating the epileptogenic zone (EZ); 2) releasing radiofrequency current through specific contacts of SEEG electrodes to thermocoagulate target brain tissue. However, existing clinical practice faces many challenges in scheme development and optimization:

[0004] Subjectivity and limitations of epileptogenic zone localization: relying on expert visual interpretation of SEEG signals, there are differences in interpretation, and traditional methods focus on the onset area, which is not necessarily the key driver node of the pathological network, resulting in non-optimal target selection.

[0005] Non-individualized and inaccurate lesion prediction: based on simplified physical models or empirical estimates of lesion range, patient brain tissue attribute differences (such as white matter electrical conduction anisotropy, vascular "heat sink effect") are ignored, resulting in a deviation between predicted and actual lesion, which may cause incomplete lesion or injury.

[0006] Lack of quantitative prediction and optimization capabilities: preoperative objective quantitative evaluation of treatment efficacy and risk is not possible, and differences in network effects of different schemes cannot be compared, leaving treatment scheme development at the "tentative" level.

[0007] Data fusion and analysis process is fragmented: multi-modal data (MRI, DTI, SEEG, etc.) are analyzed independently, information fusion relies on experience, and there is a lack of unified computational framework to mine deep correlations (such as the relationship between anatomical structure and epileptic network dynamics).

[0008] Based on this, the present application proposes a neurosurgical operation scheme optimization method based on digital twinning and causal inference to solve the above problems. SUMMARY

[0009] In order to overcome the above-mentioned defects of the prior art, the present application provides a neurosurgical operation scheme optimization method based on digital twinning and causal inference to solve the problems existing in the background art.

[0010] This invention provides the following technical solution: a method for optimizing neurosurgical procedures based on digital twins and causal inference, comprising the following steps:

[0011] S1: Construct a patient-specific computational model (PSCM), which integrates multimodal patient data, including anatomical geometry model, tissue physical property model, neurodynamic model, and radiofrequency thermocoagulation multiphysics coupling simulation model;

[0012] S2: Generative Adversarial Network (CI-GAN) based on causal inference is used to generate and optimize treatment plans. The PSCM is used to quantitatively evaluate the causal effects of candidate plans. The generator is guided to output optimized plans through adversarial training.

[0013] S3: Interactive verification. Through an interactive platform, clinicians can fine-tune the AI-generated plan. The PSCM provides real-time feedback on the efficacy and safety indicators of the adjusted plan to determine the final plan.

[0014] As a further aspect of the present invention: the construction of the patient-specific computational model (PSCM) in step S1 includes the following sub-steps:

[0015] S1.1: Multimodal data fusion and high-fidelity geometric model establishment: T1-weighted MRI, T2-FLAIR sequences, diffusion tensor imaging (DTI), and magnetic resonance angiography (MRA / MRV) are registered to a unified space. Geometric information of cortical gray matter, white matter, cerebrospinal fluid, deep nuclei, and vascular trees is extracted through the 3DU-Net segmentation network. Regions of interest (ROIs) are then defined using high-resolution brain atlases.

[0016] S1.2: Tissue physical property parameterization and spatial distribution modeling. Based on the segmentation results, an adaptive tetrahedral finite element mesh is generated. Physical parameters such as electrical conductivity, thermal conductivity, specific heat capacity and tissue density are assigned to the mesh elements. Among them, the electrical conductivity of the white matter region is constructed by constructing an anisotropic electrical conductivity tensor through the dispersion tensor derived from DTI data.

[0017] S1.3: Construction and personalization of whole-brain neurodynamic model: Deploy the "epilepsy sub" ordinary differential equation system with ROI as network nodes, compare the virtual SEEG generated by the model with the patient's real SEEG record through Bayesian inference, and optimize the model dynamic parameters;

[0018] S1.4: The multi-physics coupling simulation model of radiofrequency thermocoagulation is integrated. The potential distribution is calculated by solving the steady-state Laplace equation. Combined with the Pennes biological heat conduction equation and the Arrhenius damage integral model, the three-dimensional morphology and extent of the lesion are predicted.

[0019] As a further aspect of the present invention: the calculation formula for the conductivity tensor of the white matter region in step S1.2 is as follows:

[0020] σ WM (x)=σ ⊥ I+(σ ‖ -σ ⊥ v(x)v(x) T

[0021] Where x is the location of the white matter region, σ ⊥ and σ ‖ denoted as scalar conductivity perpendicular to and parallel to the nerve fiber bundle direction, respectively; I is a 3×3 identity matrix; v(x) is the principal eigenvector of the diffusion tensor calculated from DTI data; v(x)v(x) T It is the outer product of vector v(x) itself.

[0022] As a further aspect of the present invention: the set of ordinary differential equations for "epilepsy" in step S1.3:

[0023]

[0024] Where (x1, y1) and (x2, y2) represent the fast-timescale neuronal firing activity and the slow-timescale regulatory process, respectively;

[0025] z i A slower variable whose value determines whether node i is in a normal state or a state susceptible to epilepsy.

[0026] x 0，i A key excitability parameter is the primary goal of model personalization. 0，i The lower the value, the easier it is for node i to be triggered into epileptiform discharge;

[0027] A ij The structural connection strength between nodes i and j is quantified by fiber bundle tracking technology from the DTI data processed in S1.1. It serves as a bridge connecting the anatomical structure and the dynamic model.

[0028] K: Global coupling constant, which regulates the overall synchronization trend of the whole brain network;

[0029] The optimal parameter set is inferred by comparing the virtual SEEG generated by the model with the patient's real SEEG records using Bayesian inference methods. The model reproduces the patient-specific epilepsy pattern.

[0030] As a further aspect of the present invention: the generation and optimization of the CI-GAN-based treatment plan in step S2 includes the following sub-steps:

[0031] S2.1: Construct a CI-GAN framework, which includes a generator (G) and a discriminator (D); the generator receives a random noise vector z and a PSCM feature condition vector c, and outputs a three-dimensional treatment intervention field Φ(x); the discriminator outputs a authenticity score, a predicted efficacy index, and a safety score for candidate solutions.

[0032] S2.2: Based on PSCM, causal effect quantification of treatment options is performed, virtual surgery is conducted on candidate options, and the volume of the lesion V is calculated. lesion (S), after intervention in the neurodynamic model, the virtual SEEG before and after intervention was compared, and the causal treatment effectiveness index (CTEI) and safety score (SS) were calculated;

[0033] S2.3: Adversarial training, alternately optimizing the loss functions of the generator and discriminator; the generator optimizes the scheme by maximizing the efficacy and safety scores predicted by the discriminator, and the discriminator optimizes its evaluation ability by distinguishing between the real and generated schemes and reducing regression prediction errors.

[0034] As a further aspect of the present invention: in step S2.2, the causal treatment effectiveness index (CTEI) is obtained by quantifying the statistical differences in the epileptic seizure initiation patterns and propagation paths of the virtual SEEG before and after intervention; the safety score (SS) is calculated based on the geometric distance between the lesion and key anatomical structures such as blood vessels and functional areas.

[0035] As a further aspect of the present invention: the discriminator loss function in step S2.3 includes WGAN-GP loss and regression loss, wherein the WGAN-GP loss is used to distinguish between the real and generated protocols, and the regression loss is used to reduce the prediction error of efficacy and safety; the generator loss function includes adversarial loss and causal feedback term, the causal feedback term being weighted by a coefficient γ. e γ s Optimize the intensity of control over efficacy and safety.

[0036] As a further aspect of the present invention: interactive verification in step S3 includes scheme generation and visualization, and human-computer collaborative optimization; scheme visualization integrates candidate schemes with three-dimensional anatomical structures for display; in human-computer collaborative optimization, the clinician's fine-tuning triggers real-time PSCM simulation and updates the CTEI and SS prediction values; the final scheme is output as a standardized report containing coordinates, prediction indicators and visualization images.

[0037] As a further aspect of the present invention: the neurosurgical procedure is a radiofrequency thermocoagulation (RF-TC) procedure guided by stereotactic electroencephalography (SEEG) for the treatment of drug-resistant epilepsy (DRE).

[0038] Compared with the prior art, the present invention has the following significant advantages:

[0039] (1) Objectivity and quantification: This invention integrates the patient’s anatomical images, fiber bundle connections, vascular distribution and SEEG data into a high-fidelity “digital twin” brain by constructing a patient-specific computational model (PSCM); on this basis, the dynamic parameters of the epilepsy network are personalized through Bayesian inference, realizing objective and quantitative identification of the epilepsy source area, overcoming the limitations of traditional methods that rely on subjective experience and have interpretation differences.

[0040] (2) Precise physical simulation capability: The PSCM of this invention integrates a multi-physics field coupling simulation module that includes anisotropic conductivity tensor and vascular "heat sink effect"; this makes the simulation of RF-TC process highly individualized and precise, and can accurately predict the three-dimensional morphology and range of the lesion formed after thermal coagulation in a specific patient and at a specific location, which significantly improves the accuracy and safety of surgical planning.

[0041] (3) Prospective prediction of causal efficacy: The greatest innovation of this invention is that it can perform “virtual surgery” on PSCM, that is, simulate the causal effect of the lesion operation on the dynamics of the whole brain epilepsy network; by calculating the causal treatment effectiveness index (CTEI), this method realizes for the first time the quantitative prediction of the expected efficacy of any treatment plan before surgery; this makes clinical decision-making change from “inference based on historical experience” to “optimization based on prospective data simulation”, realizing a fundamental change from “open loop” to “closed loop”;

[0042] (4) Intelligent solution generation and optimization: This invention introduces a generative adversarial network (CI-GAN) based on causal inference, which can conduct efficient and intelligent exploration in a high-dimensional solution space. The network uses the causal evaluation results of PSCM as a supervision signal, and can learn autonomously and generate a treatment plan that achieves the best balance between efficacy (high CTEI) and safety (high SS), which greatly improves the efficiency and quality of solution formulation.

[0043] (5) Interactive decision support through human-machine collaboration: The present invention ultimately provides an interactive platform that combines AI-generated optimization schemes with the professional knowledge and experience of clinicians; the system can provide real-time feedback on the quantitative impact of any minor adjustments made by the doctor on the expected efficacy and safety; this human-machine collaboration model not only utilizes the powerful computing and optimization capabilities of artificial intelligence, but also respects the final decision-making power of clinicians, ensuring the scientific nature and clinical feasibility of the scheme. Attached Figure Description

[0044] The invention will now be further described with reference to the accompanying drawings.

[0045] Fig. 1 This is a flowchart of the method for optimizing neurosurgical procedures based on digital twins and causal inference according to the present invention.

[0046] Fig. 2 This is a block diagram of the PSCM construction module in this invention;

[0047] Fig. 3 This is a diagram of the brain MRI segmentation results using DU-Net in this invention. Detailed Implementation

[0048] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0049] Please see Figs. 1-3 As shown, this invention proposes a method for optimizing neurosurgical procedures based on digital twins and causal inference. This method constructs a high-fidelity patient-specific computational model and applies a generative adversarial network based on causal inference to this model to achieve automated generation and quantitative evaluation of treatment plans. The specific implementation method is as follows:

[0050] S1: Constructing a patient-specific computational model (PSCM)

[0051] This step aims to create a unified computing platform for each patient that integrates multi-source data and accurately reflects individual anatomical, physical, and pathophysiological characteristics. This model forms the basis for all subsequent analyses and optimizations, and its construction process follows a rigorous progressive logic.

[0052] S1.1: Multimodal data fusion and establishment of high-fidelity geometric models

[0053] Input data includes T1-weighted (T1w) MRI, T2-FLAIR sequences, diffusion tensor imaging (DTI), and magnetic resonance angiography (MRA / MRV). First, using T1w images as the anatomical reference standard, a high-precision nonlinear registration tool (such as ANTsSyN) is employed to accurately register images from all other modalities to the T1w space, ensuring spatial information consistency. Subsequently, on the unified image data, a 3DU-Net segmentation network is applied to accurately extract geometric information from cortical gray matter, white matter, cerebrospinal fluid, deep nuclei (such as the thalamus and hippocampus), and the vascular tree. Simultaneously, nonlinear registration is used to map high-resolution brain atlases (such as HCP-MMP1.0) onto the patient space, dividing the cortex into N regions of interest (ROIs). The output of this step is a three-dimensional anatomical geometric model in a unified coordinate system, incorporating fine tissue segmentation and brain region division, providing the geometric basis for subsequent physical attribute assignment and network node definition.

[0054] S1.2: Parameterization and Spatial Distribution Modeling of Tissue Physical Properties. This step transforms the purely geometric model generated in S1.1 into a computational model capable of multiphysics simulation. First, a high-quality adaptive tetrahedral finite element (FEM) mesh is generated based on the segmented tissue geometry. Next, each element in the mesh is assigned corresponding physical parameters according to its tissue type, including electrical conductivity σ(x), thermal conductivity k(x), specific heat capacity c(x), and tissue density ρ(x). For the white matter region, electrical conductivity is not a scalar but a tensor reflecting the orientation of nerve fiber bundles. This conductivity tensor σ... WM (x) is constructed from the diffusion tensor derived from DTI data, and its mathematical form is:

[0055] σ WM (x)=σ ⊥ I+(σ ‖ -σ ⊥ v(x)v(x) T

[0056] The terms in this formula are explained as follows:

[0057] x: represents any spatial location within the white matter region.

[0058] σ WM (x): The 3×3 conductivity tensor at position x.

[0059] σ ⊥ and σ ‖ : These are scalar conductivity values ​​perpendicular to and parallel to the direction of the local nerve fiber bundle, respectively, reflecting the anisotropy of current conduction.

[0060] I: A 3×3 identity matrix.

[0061] v(x): The principal feature vector of the diffusion tensor calculated from DTI data at position x, whose direction represents the average direction of the nerve fiber bundle at that point.

[0062] v(x)v(x) T The outer product of vector v(x) itself is a projection operator.

[0063] The completion of this step transforms the PSCM from a static geometric model into a physical field simulation model containing complete physical parameters. The distribution of its internal electrical and thermal properties provides the necessary conditions for the simulation calculation of S1.4.

[0064] S1.3: Construction and Personalization of the Whole-Brain Neurodynamic Model. This step aims to endow the PSCM with the ability to simulate the brain's spontaneous and evoked electrical activity. We use the N ROIs partitioned in S1.1 as network nodes, and deploy a set of ordinary differential equations, the "epiliptor," to describe epileptic dynamics at each node i:

[0065]

[0066] The key terms in this model are explained as follows:

[0067] (x1, y1) and (x2, y2): represent the fast-timescale neuronal firing activity and the slow-timescale regulatory process, respectively.

[0068] z i A slower variable whose value determines whether node i is in a normal state or a state susceptible to epilepsy.

[0069] x 0，i A key excitability parameter is the primary goal of model personalization. 0，i The lower the value, the easier it is for node i to be triggered into epileptiform discharge.

[0070] A ij The structural connection strength between nodes i and j is quantified by fiber bundle tracking technology from the DTI data processed in S1.1. It serves as a bridge connecting the anatomical structure and the dynamic model.

[0071] K: Global coupling constant, which regulates the overall synchronization trend of the whole brain network.

[0072] To enable this model to reproduce the epilepsy patterns of a specific patient, we employ Bayesian inference by comparing the virtual SEEG generated by the model with the patient's actual SEEG records to infer the optimal parameter set. After this step, PSCM becomes a personalized model capable of simulating the patient's unique epileptic dynamics.

[0073] S1.4: Integration of the Multiphysics Coupled Simulation Model for Radiofrequency Thermocoagulation. To enable the PSCM to evaluate treatment protocols, the final step is to integrate the physical processes of RF-TC. When a treatment protocol (i.e., a set of target coordinates) is proposed, this module first calculates the potential distribution V(x) on the FEM mesh constructed in S1.2, using the defined conductivity distribution σ(x), by solving the steady-state Laplace equation:

[0074]

[0075] The solution to this equation gives the electric field strength. This leads to the heat source term Q generated within the organization. rf (x)=σ(x)|E(x)| 2 Subsequently, this heat source term is substituted into Penners' biological heat conduction equation to solve for the temperature field T(x, t) within the tissue:

[0076]

[0077] Ultimately, the extent of tissue damage was quantified by the Arrhenius damage integral model, which calculates a dimensionless damage exponent Ω(x,t) that incorporates the combined effects of temperature and time.

[0078]

[0079] The terms in this formula are explained as follows:

[0080] Ω(x, t ): Cumulative damage exponent at location x at time t, Ω = 1 is defined as the threshold for complete tissue necrosis. A: Pre-exponential factor (unit: s) -1 ), where is a material-related constant.

[0081] E a Activation energy (unit: J·mol) -1 ), representing the thermodynamic energy barrier required for tissue denaturation.

[0082] R: Universal gas constant (8.314 J·mol⁻¹) -1 ·K -1 ).

[0083] T(x, τ): The absolute temperature of position x at time τ, obtained from the biological heat equation (unit: K).

[0084] By solving this integral, the coordinate input of any treatment plan can be transformed into a precisely predicted three-dimensional lesion volume V through PSCM. lesion At this point, a complete and computable "digital twin" model has been constructed.

[0085] S2: CI-GAN-based treatment regimen generation and optimization

[0086] On the high-fidelity computing platform PSCM, we apply a generative adversarial network (CI-GAN) framework based on causal inference to intelligently generate and iteratively optimize treatment plans.

[0087] S2.1: Composition of the CI-GAN framework. This framework consists of two competing deep neural networks: Generator (G): Its function is to learn a mapping from the latent space to the space of effective treatment options. It takes a random noise vector z and a conditional vector c extracted from PSCM (encoding key features of the patient's epilepsy network) as input and outputs a three-dimensional "treatment intervention field" Φ(x) covering the entire brain, the value of which represents the appropriateness of ablation at each location. Discriminator (D): Its function is as a multi-task judging network. It takes a candidate treatment option (a mask of ablation foci obtained by thresholding Φ(x)) and outputs three values: (1) adversarial score, judging the authenticity of the option; (2) predicted efficacy index; (3) predicted safety score.

[0088] S2.2: Causal Effect Quantification of Treatment Protocols Based on PSCM. This step is the linchpin connecting CI-GAN and PSCM, and its purpose is to provide supervisory signals (i.e., "gold standard" labels) for the training of the discriminator D. For any given treatment protocol S, we perform the following "virtual surgery" procedure: 1. Perform physical intervention: Input protocol S into the simulation module of S1.4 to calculate the corresponding three-dimensional lesion volume V. lesion (S). 2. Perform dynamic intervention: In the neurodynamic model of S1.3, all those located in V lesion ROI nodes within (S) and their relationships in the structural connection matrix A ij The connection weights in the model are reset to zero. This operation is mathematically equivalent to performing a causal intervention do(Ablation = S). 3. Quantify the intervention effect: Run the neurodynamic model after the intervention to generate a "post-operative" virtual SEEG. By comparing the virtual SEEGs before and after the intervention, calculate the Causal Therapeutic Efficacy Index (CTEI). Simultaneously, based on V... lesion (S) is the geometric relationship between the key anatomical structures (such as blood vessels and the argumentative area) established in S1.1, and the safety score (SS) is calculated.

[0089] S2.3: The adversarial training algorithm for CI-GAN. The training process involves alternately optimizing the parameters of the generator G and the discriminator D. Training the discriminator D: The loss function L of the discriminator D. D It consists of two parts:

[0090]

[0091] The terms in this formula are explained as follows:

[0092] L WGAN-GPWasserstein GAN loss with gradient penalty. adv (·) represents the adversarial score output of the discriminator. This is intended to enable D to distinguish between real and generated schemes. λ gp It is the weight coefficient of the gradient penalty. These are random interpolation points between real and generated samples.

[0093] L regression : Regression loss. D eff (S) and D safe (S) is the discriminator's predicted value for efficacy and safety. CTEI(S) and SS(S) are the true labels calculated from S2.2. MSE is the mean squared error. This step aims to train D to accurately predict the efficacy and safety of any regimen. λ e , λ s These are the weights of each regression task.

[0094] Training the generator G: The loss function L of the generator G G It is designed to use the discriminator's evaluation output to guide its optimization direction:

[0095] L G =-E[D adv (G(z|c))]-γ e E[D eff (G(z|c))]-γ s E[D safe (G(z|c))]

[0096] The terms in this formula are explained as follows:

[0097] -E[D adv [G(z|c))]: Adversarial loss. The goal of G is to generate an adversarial loss that allows D to... adv The solution that scores high is given in order to "deceive" the discriminator.

[0098] -γ e E[D eff (G(z|c))] and -γ s E[D safe [G(z|c))]: This is the core innovation of this method, namely the causal feedback term. G is directly guided to generate a treatment regimen that performs better in PSCM by maximizing the efficacy and safety scores predicted by the discriminator. γ e γ s It is a weighting coefficient that controls the strength of this guidance.

[0099] Through this adversarial training, the generator G eventually learned to generate a treatment plan optimized in both efficacy and safety for a specific patient's PSCM.

[0100] S3: Interactive Validation

[0101] The trained CI-GAN system was deployed as a clinical decision support tool.

[0102] Solution generation and visualization: For new patients, the system first constructs their PSCM according to the S1 process. Then, the feature vector of the PSCM is input into the trained generator G to generate candidate solutions, which are then fused and displayed with the patient's three-dimensional anatomical structure (transparent brain, blood vessels, fiber bundles).

[0103] Interactive optimization through human-machine collaboration: The system provides an interactive platform that allows clinicians to fine-tune AI-generated treatment plans based on their expertise. Key consistency lies in the fact that every manual modification by a clinician (such as dragging, adding, or deleting target points) immediately triggers the causal effect assessment process in S2.2. The system utilizes PSCM to perform rapid "virtual surgery" simulations in the background and updates the CTEI and SS prediction values ​​displayed on the interface in real time within seconds. This provides clinicians with a powerful quantitative feedback loop, enabling them to combine clinical experience with data-driven causal analysis to ultimately determine a fully validated optimal treatment plan. The final plan is output in a standardized report format containing precise coordinates, predictive indicators, and visualizations to guide surgical execution.

[0104] The above details the overall framework and technical aspects of a neurosurgical treatment plan optimization method based on digital twins and causal inference. From the construction logic of the patient-specific computational model (PSCM) to the CI-GAN-based plan generation and optimization principle, and finally to the closed-loop process of interactive verification, it fully presents the technical chain of this method from data integration to plan output. To further illustrate the practical application effects and operational details of this method in clinical scenarios, the following will demonstrate its implementation process in optimizing treatment plans for real patients using specific cases:

[0105] Specifically, this embodiment describes the entire process of applying the method of the present invention to a 32-year-old patient A suffering from drug-resistant right temporal lobe epilepsy. Initial clinical suspicion suggests that the epilepsy originated in the right hippocampus and amygdala region.

[0106] S1: Construction of a patient-specific computational model (PSCM) for patient A

[0107] Input data: High-resolution T1-weighted MRI (1mm) 3 Voxel scans. T2-FLAIR sequence showing a slight signal abnormality in the right temporal pole. DTI data (64 diffusion directions). MRA data used to reconstruct the arterial tree. Long-term (72-hour) monitoring data recorded by a SEEG electrode implanted in the right hemisphere, during which 3 recurrent episodes were captured.

[0108] S1.1: Geometric Model Construction First, all T2-FLAIR, DTI, and MRA data of patient A were nonlinearly registered to their T1w image space, forming a multimodal image dataset in a unified coordinate system. Subsequently, the 3DU-Net segmentation network processed this dataset, outputting the following results: 1. Tissue Segmentation: Precise 3D masks of structures such as gray matter, white matter, cerebrospinal fluid, hippocampus, and amygdala. The segmentation results confirmed that the slightly abnormal signal region on the T2-FLAIR sequence corresponded to a localized cortical dysplasia (FCDTypella) in the right temporal pole (Brodmann area 38). 2. Vascular Reconstruction: The geometric model of the middle cerebral artery (MCA) and its main branches was reconstructed from the MRA data. 3. Brain Region Division: By nonlinearly registering the HCP-MMP1.0 atlas to patient A's brain, the cortex was divided into 360 ROIs, and combined with the deep nucleus segmentation results, a total of 379 independent anatomical nodes were obtained.

[0109] S1.2: Physical Properties Assigned Based on the geometric model generated in S1.1, we constructed an adaptive finite element (FEM) mesh containing approximately 3.5 million tetrahedral elements. Physical parameters were assigned to each element based on its organization label, such as the gray matter conductivity σ. GM =0.33 S / m, the anisotropic conductivity parameter of white matter is σ ⊥ =0.15S / m and σ ‖ = 0.45 S / m. Specifically, for each cell in the white matter region, a unique conductivity tensor σ is calculated and assigned based on the DTI data at that location. WM (x). At this point, a static physical model containing precise anatomical structure and anisotropic physical properties has been constructed.

[0110] S1.3: Personalization of the Neurodynamic Model We used the 379 brain regions segmented in S1.1 as network nodes and deployed an "epilepsy" dynamic model on each node. The connection strength matrix between nodes is A. ij Quantized by DTI fiber tract tracking. Next comes the crucial personalization step: Input: 72-hour SEEG data from patient A. Procedure: We run a Bayesian inference algorithm based on MCMC, the goal of which is to adjust the set of excitability parameters {x} in the model. 0，iThis allows the virtual SEEG signal generated by the model to achieve the greatest possible agreement with the statistical characteristics (such as the Epileptogenicity Index, EI) of patient A's actual SEEG record in terms of the onset pattern and propagation path of epileptic seizures. Output: The inference results show that, among 379 nodes, the excitability parameter x of three nodes—the right anterior hippocampus (ROI215), the right amygdala (ROI218), and the right temporal pole cortex (ROI312, corresponding to the FCD region)—is significantly higher. 0，i The values ​​were identified as significantly lower than the average brain values ​​(-2.05, -1.98, -2.10, respectively, compared to a brain average of -1.6). This quantitatively indicates at the model level that the core of patient A's epileptic network is located in these three regions.

[0111] S1.4: Radiofrequency Thermocoagulation Simulation Module Integration. We have integrated the simulation module containing the electro-thermal-damage coupling equations into the personalized model described above. Now, patient A's PSCM has been constructed. It not only includes its unique anatomical and physical properties but also simulates its unique seizure behavior and can predict the thermal ablation effect under any given RF-TC parameters.

[0112] S2: Treatment plan generation and iterative optimization based on CI-GAN

[0113] S2.1 & S2.2: Initialization and First Evaluation of CI-GAN We encode the key features of the patient's epilepsy network obtained in S1.3 (i.e., the low excitability parameters of ROIs 215, 218, and 312) into a conditional vector c. This vector, along with a random noise vector z, is input into a pre-trained generator G. Generator Output: G generates an initial three-dimensional treatment intervention field Φ0(x). By setting a threshold, an initial treatment plan S0 is extracted from this field, which suggests one ablation each at the right hippocampus, amygdala, and temporal pole cortex, for a total of three target points. PSCM Evaluation: The system automatically submits plan S0 to patient A's PSCM for "virtual surgery" evaluation. Module S1.4 calculates that S0 will generate a total of approximately 480 mm. 3 The volume of damage was determined. Subsequently, the three nodes and their connections were "silenced" in the kinetic model of S1.3, and the model was run again. By comparing the virtual SEEG before and after the intervention, the causal treatment effectiveness index CTEI was calculated to be 0.65, and the safety score SS was calculated to be 0.98 (due to the target point being far from the main vascular branches).

[0114] S2.3: The initial scheme S0 and its evaluation results (CTEI = 0.65, SS = 0.98) of the adversarial optimization loop are used in the online optimization process of Cl-GAN. The discriminator D updates its prediction ability based on this "real" data and provides gradient feedback to the generator G. The goal of G is to generate a scheme that allows D to predict higher CTEI and SS. Iterative process: This optimization process is performed automatically in the background. The generator G continuously generates new schemes S1, S2, ..., each of which is quickly evaluated by the PSCM, and the evaluation results in turn guide the next generation of G. Convergence result: After about 200 iterations, the output of the generator G converges, giving an optimized intervention field Φ. opt (x). The optimized solution S extracted from it opt Significantly different from the initial plan: it recommended two more concentrated, irregularly shaped (conformal) lesions, one covering the junction of the right hippocampal head and amygdala, and the other precisely covering the FCD region of the temporal pole. Final assessment: for S opt The final "virtual surgery" evaluation yielded the following results: CTEI = 0.92, SS = 0.95. Compared to the initial protocol, the predictive value for efficacy was significantly improved, while the safety score decreased slightly but remained at a very safe level. This indicates that by more precisely targeting the core nodes of the network, better therapeutic effects can be achieved with less lesion volume.

[0115] S3: Clinical Interaction Validation and Protocol Finalization

[0116] Visual presentation: Optimization solution S opt Presented on a 3D interactive interface, the attending physician can clearly see: a semi-transparent brain model of patient A; the right hippocampus and amygdala highlighted in green; the MRA-reconstructed vascular tree shown in blue; and the brain structure reconstructed by S, shown in red. opt Two suggested areas of conformal damage.

[0117] Human-Machine Collaborative Optimization: During the review of the treatment plan, the doctor discovered that the lesion covering the hippocampal-amygdala junction had an anterior margin only 2.5 mm from a small branch of the middle cerebral artery. Despite a high SS score of 0.95, for maximum safety, the doctor used the mouse to shift the centroid of the lesion 2 mm backward on the 3D interface. Real-time Feedback: After the doctor submitted the modification, the system immediately triggered the PSCM in the background to re-perform the rapid assessment process of S2.2. Updated Results: Approximately 15 seconds later, the predicted values ​​on the interface were dynamically updated to: CTEI = 0.89, SS = 0.99.

[0118] Final decision: The doctor observed that increasing the safe distance by 2mm slightly decreased the predicted efficacy index from 0.92 to 0.89, while improving the safety score from 0.95 to 0.99. The doctor deemed this trade-off acceptable. He ultimately approved the manually tweaked plan.

[0119] Output Report: The system ultimately generates a PDF report, which includes: 1. The stereotactic coordinates (center point and shape parameters) of the two finalized ablation targets. 2. The final predicted efficacy index (CTEI = 0.89) and safety score (SS = 0.99). 3. 3D screenshots from multiple angles, clearly showing the positional relationship between the final ablation targets and surrounding key anatomical structures (hippocampus, amygdala, blood vessels).

[0120] This report will be used directly to guide subsequent surgical planning and navigation systems, completing a closed loop from raw data collection to the development of personalized, intelligent, and clinician-validated treatment plans.

[0121] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for optimizing neurosurgical procedures based on digital twins and causal inference, characterized by: Includes the following steps: S1: Construct a patient-specific computational model (PSCM), which integrates multimodal patient data, including anatomical geometry model, tissue physical property model, neurodynamic model, and radiofrequency thermocoagulation multiphysics coupling simulation model; S2: Generative Adversarial Network (CI-GAN) based on causal inference is used to generate and optimize treatment plans. The PSCM is used to quantitatively evaluate the causal effects of candidate plans. The generator is guided to output optimized plans through adversarial training. S3: Interactive verification. Through an interactive platform, clinicians can fine-tune the AI-generated plan. The PSCM provides real-time feedback on the efficacy and safety indicators of the adjusted plan to determine the final plan.

2. The method for optimizing neurosurgical procedures based on digital twins and causal inference according to claim 1, characterized in that: Step S1, which involves constructing a patient-specific computational model (PSCM), includes the following sub-steps: S1.1: Multimodal data fusion and high-fidelity geometric model establishment: T1-weighted MRI, T2-FLAIR sequences, diffusion tensor imaging (DTI), and magnetic resonance angiography (MRA / MRV) are registered to a unified space. Geometric information of cortical gray matter, white matter, cerebrospinal fluid, deep nuclei, and vascular trees is extracted through the 3DU-Net segmentation network. Regions of interest (ROIs) are then defined using high-resolution brain atlases. S1.2: Tissue physical property parameterization and spatial distribution modeling. Based on the segmentation results, an adaptive tetrahedral finite element mesh is generated. Physical parameters such as electrical conductivity, thermal conductivity, specific heat capacity and tissue density are assigned to the mesh elements. Among them, the electrical conductivity of the white matter region is constructed by constructing an anisotropic electrical conductivity tensor through the dispersion tensor derived from DTI data. S1.3: Construction and personalization of whole-brain neurodynamic model: Deploy the "epilepsy sub" ordinary differential equation system with ROI as network node, compare the virtual SEEG generated by the model with the patient's real SEEG record through Bayesian inference, and optimize the model dynamic parameters; S1.4: The multi-physics coupling simulation model of radiofrequency thermocoagulation is integrated. The potential distribution is calculated by solving the steady-state Laplace equation. Combined with the Pennes biological heat conduction equation and the Arrhenius damage integral model, the three-dimensional morphology and extent of the lesion are predicted.

3. The method for optimizing neurosurgical procedures based on digital twins and causal inference according to claim 2, characterized in that: The formula for calculating the conductivity tensor of the white matter region in step S1.2 is as follows: s WM (x)=σ ⊥ I+(s ‖ -s ⊥ )v(x)v(x) T Where x is the location of the white matter region, σ ⊥ and σ ‖ denoted as scalar conductivity perpendicular to and parallel to the nerve fiber bundle direction, respectively; I is a 3×3 identity matrix; v(x) is the principal eigenvector of the diffusion tensor calculated from DTI data; v(x)v(x) T It is the outer product of vector v(x) itself.

4. The method for optimizing neurosurgical procedures based on digital twins and causal inference according to claim 2, characterized in that: The system of ordinary differential equations for "epilepsy" in step S1.3: Where (x1, y1) and (x2, y2) represent the fast-timescale neuronal firing activity and the slow-timescale regulatory process, respectively; z i A slower variable whose value determines whether node i is in a normal state or a state susceptible to epilepsy. x 0，i A key excitability parameter is the primary goal of model personalization. 0，i The lower the value, the easier it is for node i to be triggered into epileptiform discharge; A ij The structural connection strength between nodes i and j is quantified by fiber bundle tracking technology from the DTI data processed in S1.

1. It serves as a bridge connecting the anatomical structure and the dynamic model. K: Global coupling constant, which regulates the overall synchronization trend of the whole brain network; The optimal parameter set is inferred by comparing the virtual SEEG generated by the model with the patient's real SEEG records using Bayesian inference methods. The model reproduces the patient-specific epilepsy pattern.

5. The method for optimizing neurosurgical procedures based on digital twins and causal inference according to claim 1, characterized in that: Step S2, the generation and optimization of CI-GAN-based treatment plans, includes the following sub-steps: S2.1: Construct a CI-GAN framework, which includes a generator (G) and a discriminator (D); the generator receives a random noise vector z and a PSCM feature condition vector c, and outputs a three-dimensional treatment intervention field Φ(x); the discriminator outputs a authenticity score, a predicted efficacy index, and a safety score for candidate solutions. S2.2: Based on PSCM, causal effect quantification of treatment options is performed, virtual surgery is conducted on candidate options, and the volume of the lesion V is calculated. lesion (S), after intervention in the neurodynamic model, the virtual SEEG before and after intervention was compared, and the causal treatment effectiveness index (CTEI) and safety score (SS) were calculated; S2.3: Adversarial training, alternately optimizing the loss functions of the generator and discriminator; the generator optimizes the scheme by maximizing the efficacy and safety scores predicted by the discriminator, and the discriminator optimizes its evaluation ability by distinguishing between the real and generated schemes and reducing regression prediction errors.

6. The method for optimizing neurosurgical procedures based on digital twins and causal inference according to claim 5, characterized in that: In step S2.2, the Causal Treatment Effectiveness Index (CTEI) is quantified by comparing the statistical characteristics of the epileptic seizure initiation patterns and propagation pathways in the virtual SEEG before and after intervention. The safety score (SS) is calculated based on the geometric distance between the lesion and key anatomical structures such as blood vessels and functional areas.

7. The method for optimizing neurosurgical procedures based on digital twins and causal inference according to claim 5, characterized in that: In step S2.3, the discriminator loss function includes WGAN-GP loss and regression loss. WGAN-GP loss is used to distinguish between real and generated protocols, while regression loss is used to reduce prediction errors in efficacy and safety. The generator loss function includes adversarial loss and a causal feedback term, with the causal feedback term weighted by a coefficient γ. e γ s Optimize the intensity of control over efficacy and safety.

8. The method for optimizing neurosurgical procedures based on digital twins and causal inference according to claim 1, characterized in that: The interactive verification in step S3 includes scheme generation and visualization, and human-computer collaborative optimization; scheme visualization integrates candidate schemes with three-dimensional anatomical structures for display; in human-computer collaborative optimization, clinicians' fine-tuning triggers real-time PSCM simulation and updates CTEI and SS prediction values; the final scheme is output as a standardized report containing coordinates, prediction indicators and visualization images.

9. The method for optimizing neurosurgical procedures based on digital twins and causal inference according to any one of claims 1-8, characterized in that: The neurosurgical procedure was a stereotactic electroencephalogram (SEEG)-guided radiofrequency thermocoagulation (RF-TC) procedure for the treatment of drug-resistant epilepsy (DRE).

Images