Virtual reality-based liver failure risk progression prediction method and system

Abstract

The application discloses a liver failure risk process prediction method and system based on virtual reality, and relates to the technical field of virtual reality; the method comprises the following steps: acquiring multi-modal medical image data, generating a fusion image data set after standardization preprocessing, obtaining a liver region mask after segmenting a target region, and constructing a blood vessel network topology structure based on the liver region mask; converting the obtained liver region mask into a three-dimensional surface grid model, completing a setting action and a texture mapping action on liver material, adopting a point-spring model to construct a biomechanical model of the liver, and mapping clinical examination indexes onto the biomechanical model; the technical key points are that a dynamic architecture of AI prediction-interactive correction-model relearning is constructed, the inherent problems of poor interpretability and insufficient flexibility of a traditional AI model in a complex clinical scene are fundamentally solved, and a paradigm shift from a black box to human-computer cooperation is realized.

Description

Technical Field

[0001] This invention relates to the field of virtual reality technology, specifically to a method and system for predicting the risk progression of liver failure based on virtual reality. Background Technology

[0002] Virtual reality (VR) is a technology that uses computers to generate a three-dimensional immersive environment and provides users with a sense of presence through multi-sensory feedback such as vision and hearing. Its core characteristics include immersion, interactivity, and imaginative possibilities. When applied to the medical field, for example in predicting the progression of risks in existing internal organs, VR technology is often used as an important visualization and interaction tool for digital twins or virtual surgical systems. The general process is as follows: First, by collecting the patient's medical images and clinical data, a patient-specific three-dimensional organ anatomical model is constructed in the computer. Then, virtual surgical technology is used to simulate the physiological or pathological evolution of the organ. Doctors can use VR headsets and spatial interaction controllers to intuitively observe organ morphological changes and simulate different intervention plans in virtual space, and combine this with 3D printed physical models for physical verification, thereby predicting risks such as anatomical variations and complications.

[0003] Existing technologies for predicting the risk of liver failure mainly have the following problems:

[0004] First, traditional segmentation models often miss segments when processing liver CT / MRI due to blurred boundaries or background noise. Furthermore, multimodal data is often simply stitched together, making it difficult to automatically filter out invalid information. Second, existing human-computer interaction corrections are mostly isolated operations; the system only records the final result and lacks memory and logical connection to the doctor's intervention process, such as the magnitude and direction of adjustments, leading to a break in contextual information. For example, if a doctor increases the weight of a liver lobe, the system cannot intelligently continue this intervention intention, easily resulting in reverse parameter distortion. Traditional AI models are mostly black-box outputs, lacking flexibility. Doctors cannot intuitively inject difficult-to-quantify clinical experience, such as functional abnormalities in specific areas, into the prediction process. Moreover, mechanical parameter stacking easily leads to drastic fluctuations in prediction results, making it difficult to achieve true human-computer collaboration and continuous iterative optimization of the model, thus affecting the sustainable development of clinical decision-making. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A virtual reality-based method for predicting the progression of liver failure risk includes the following steps:

[0007] After acquiring multimodal medical image data and performing standardized preprocessing, a fused image dataset is generated. The target region is segmented to obtain a liver region mask, and a vascular network topology is constructed based on the liver region mask.

[0008] The obtained liver region mask is converted into a three-dimensional surface mesh model. The liver material is set and texture mapping is performed. A mass-spring model is used to construct a biomechanical model of the liver to realize real-time deformation simulation of liver tissue. Clinical examination indicators are mapped onto the biomechanical model to achieve information visualization.

[0009] Multimodal features of the target are extracted and feature fusion is performed using a gated attention fusion mechanism; a multi-task deep learning prediction model is constructed and preliminary prediction results are output synchronously; based on the preliminary prediction results, a risk dynamic correction mechanism based on virtual reality interaction is triggered and the final prediction results are output; the prediction results include: the predicted probability of liver failure, the progression time window, and the severity grading; the risk dynamic correction mechanism based on virtual reality interaction includes setting primary correction conditions, setting secondary correction conditions, and judging convergence of multiple corrections;

[0010] Based on the final prediction results, construct an interactive form for the virtual reality risk process;

[0011] Risk stratification and early warning actions will be carried out based on the final forecast results.

[0012] Furthermore, multimodal medical imaging data includes at least: CT images, MR images, and LS data;

[0013] The standardized preprocessing involves image registration and fusion of multimodal medical image data, spatial registration, establishment of a unified three-dimensional coordinate system, and the use of a multimodal image registration algorithm based on mutual information (MI) to achieve spatial alignment by maximizing the mutual information value. After registration, the multimodal medical image data is fused at the pixel level to obtain a fused image dataset.

[0014] Furthermore, the process of obtaining the liver region mask is as follows: an improved 3D U-Net deep learning model is used to segment the target region of the fused image dataset; wherein, the improved 3D U-Net deep learning model means: an attention gating mechanism is introduced on the basis of the existing U-Net, and after segmenting the target region, the improved 3D U-Net deep learning model is used for model training and inference to obtain the liver region mask.

[0015] The basis for constructing the vascular network topology is as follows: based on the liver region mask, the intrahepatic vascular structure is extracted using a vascular enhancement filtering algorithm based on the Hessian matrix, the response function of the vascular network topology is defined, the vascular enhancement degree (Vesselness) is obtained, and the vascular system is extracted and constructed through threshold segmentation.

[0016] Furthermore, when converting to obtain a three-dimensional surface mesh model, the moving cube algorithm is used to extract the liver isosurface and generate triangular facet meshes; when setting the liver material, physically based rendering technology is applied to set the material; when performing texture mapping on the liver material: the liver surface texture is generated by extracting grayscale information from CT images to generate normal maps and roughness maps; clinical examination indicators include at least: liver function reserve parameters, LS data, and hemodynamic parameters.

[0017] Furthermore, the extracted target multimodal features include at least: image feature vectors, clinical examination indicator feature vectors, and genomic feature vectors; when constructing a multi-task deep learning prediction model, its backbone network adopts a Transformer encoder structure, which contains 6 encoder layers, each of which contains a multi-head self-attention mechanism and a feedforward neural network.

[0018] Among them, the multi-head self-attention mechanism is to compute the attention of multiple subspaces in parallel;

[0019] The feedforward neural network consists of two fully connected layers with the GELU activation function in between;

[0020] The output layer of the multi-task deep learning prediction model contains three parallel prediction heads: a risk occurrence prediction head, which uses the Sigmoid activation function to output the probability of liver failure, with a value range of [0, 1]; a progression time prediction head, which uses a linear activation function to output the expected time to progress to liver failure, corresponding to the progression time window; and a severity grading prediction head, which uses the Softmax activation function to output the probability distribution of mild, moderate, and severe liver failure, corresponding to the severity grading.

[0021] Furthermore, the setting of the first correction condition means: setting the trigger condition for the first correction, namely condition A: when the output predicted probability of liver failure p̂ is greater than the preset first threshold θ1, the first correction mode is entered.

[0022] The preset first threshold θ1 has a value range of 0 to 1; in the single correction mode, medical staff can choose whether to select the target area on the liver model through the virtual reality controller and adjust the functional parameters of the target area.

[0023] The first correction formula is: p̂'=p̂+Δp; where p̂' represents the corrected risk probability, Δp represents the risk correction amount, and the formula is: Δp=a1×(w_region-w_baseline)×f_region; where a1 is the first correction coefficient, with a value range of [0.1, 0.3], w_region represents the weight coefficient of the target region, with a value range greater than 1, w_baseline represents the baseline weight coefficient, and f_region represents the functional abnormality score of the target region, with a value range of [0, 1].

[0024] Furthermore, the secondary correction condition setting means: setting the trigger condition for secondary correction, namely condition B: when the change in risk probability after the first correction |Δp| ​​is greater than the preset second threshold θ2, and the corrected risk probability p̂' is still greater than the preset first threshold θ1, the directional index of the first correction to obtain the first risk correction amount satisfies: sign(Δp)=sign(ΔROI_score), where ΔROI_score represents the amount of adjustment change of the degree of functional abnormality of the region of interest by the medical staff;

[0025] If three sub-conditions in condition B above are met, then the secondary correction mode is entered.

[0026] The preset second threshold θ2 has a value range of 0~1. In the secondary correction mode, medical staff can choose whether to adjust the liver hemodynamic parameters, including at least: portal vein pressure (PVP) and hepatic artery blood flow (HAF).

[0027] The secondary correction formula is: p̂''=p̂'+Δp'; where p̂'' represents the risk probability after secondary correction, and Δp' represents the secondary correction amount, based on the formula: Δp'=a2×(ΔPVP / PVP_baseline+ΔHAF / HAF_baseline); where a2 is the second correction coefficient, with a value greater than 0, equipped with a dynamic adjustment mechanism for the second correction coefficient, based on: a2=a2_0×(1+t0×|Δp| / θ2); where a2_0 is the baseline coefficient, with a value range of 0~1, t0 is the sensitivity coefficient, with a value range of [0.2, 0.5]; ΔPVP represents the change in portal vein pressure, PVP_baseline represents the baseline portal vein pressure, ΔHAF represents the change in hepatic artery blood flow, and HAF_baseline represents the baseline hepatic artery blood flow.

[0028] Furthermore, the multiple correction convergence judgment means that the correction convergence condition is set: when the risk correction amount of two consecutive times is less than the preset third threshold θ3, the correction process is considered to have converged, and the final prediction result is output; where the preset third threshold θ3 is set to a value range of 0~1; after the second correction mode is completed, if further correction is still needed, the correction will continue, without needing to satisfy condition A or condition B again; the operation method is to randomly select the adjustment object: adjust the functional parameters of the target area again, corresponding to the first correction mode; adjust the liver hemodynamic parameters again, corresponding to the second correction mode; or adjust any one of the multiple parameters in a mixed manner.

[0029] Among these parameters are functional parameters of the target region and hepatic hemodynamic parameters.

[0030] Furthermore, the construction of a virtual reality risk process interaction format includes at least: a timeline of the liver failure risk process, a dynamic evolution simulation of the liver state, and a virtual reality interactive interface setting; wherein, the timeline of the liver failure risk process means: the timeline of the liver failure risk process is constructed starting from the current time of the target patient and extending into the future to the predicted time of liver failure; the dynamic evolution simulation of the liver state: based on the timeline, the dynamic evolution process of the liver state from the current state to the state of liver failure is simulated, and the evolution simulation includes at least: changes in liver morphology and changes in liver color; the virtual reality interactive interface setting includes a virtual reality interactive interface, including a main view, a risk dashboard, a time controller, a parameter adjustment panel, and annotation tools;

[0031] The risk stratification and early warning actions are based on the following: according to the probability of liver failure in the final prediction results, different patients are classified into risk levels and the corresponding decision suggestion texts in the decision template library are retrieved.

[0032] A virtual reality-based system for predicting the progression of liver failure risk, comprising:

[0033] The data acquisition and topology construction module acquires multimodal medical image data, performs standardized preprocessing, generates a fused image dataset, segments the target region to obtain a liver region mask, and constructs a vascular network topology based on the liver region mask.

[0034] The modeling and mapping module converts the obtained liver region mask into a three-dimensional surface mesh model, completes the setting action and texture mapping action for the liver material, and uses a mass-spring model to construct a biomechanical model of the liver to realize real-time deformation simulation of liver tissue. Clinical examination indicators are mapped onto the biomechanical model to realize information visualization.

[0035] A feature fusion and risk correction module, which extracts target multimodal features and completes feature fusion using a gated attention fusion mechanism; constructs a multi-task deep learning prediction model to output preliminary prediction results synchronously; triggers a virtual reality interaction-based dynamic risk correction mechanism according to the preliminary prediction results, and outputs final prediction results; wherein the prediction results include the predicted occurrence probability of liver failure, progression time window and severity grade; the virtual reality interaction-based dynamic risk correction mechanism includes primary correction condition setting, secondary correction condition setting and multiple correction convergence judgment;

[0036] A virtual reality interaction module, which constructs a virtual reality risk progression interaction form based on the final prediction results;

[0037] A layered early warning support module, which completes risk stratification and early warning actions according to the final prediction results.

[0038] The present invention provides a virtual reality-based liver failure risk progression prediction method and system, which have the following beneficial effects:

[0039] 1. This solution introduces a three-dimensional U-Net improved by an attention gating mechanism, which effectively solves the under-segmentation problem of traditional segmentation models when processing liver regions with blurred boundaries and heavy background noise interference. It uses gating signals to filter irrelevant features in skip connections, enabling the model to focus on liver edges and lesion areas, and cooperates with a composite loss function including boundary loss to greatly improve segmentation accuracy. The adopted gated attention fusion mechanism is not a simple concatenation, but can automatically learn the contribution of different modal data to liver failure prediction, suppress the interference of invalid information, and achieve the effect of providing high-purity, high-correlation feature vectors for subsequent risk prediction through refined feature extraction and fusion.

[0040] 2. This solution converts the process data of primary correction rather than the result data into the decision basis for secondary correction, which is in sharp contrast with traditional solutions. It takes the magnitude of |Δp| as the scale for measuring the intervention intensity of medical staff, dynamically adjusts the second correction coefficient a2, so that the correction intensity is proportional to the intervention intensity of medical staff, realizing self-adaptation of the correction process; at the same time, the solution introduces directionality verification of sign(Δp) to ensure that the secondary correction is consistent with the initial intention of medical staff, and avoids reverse distortion of prediction results caused by parameter adjustment;

[0041] This deep reuse of process data solves the technical problem of broken context information in traditional human-computer interaction, and realizes coherent intelligent correction with memory and logic, so that the system supporting this solution is no longer a tool that passively responds to instructions, but an intelligent collaborator that can understand and continue the decision-making ideas of medical staff, significantly improving the prediction accuracy and clinical credibility.

[0042] 3. The core of this solution lies in: by constructing a dynamic architecture of AI prediction-interactive correction-model relearning, it fundamentally solves the inherent problems of poor interpretability and insufficient flexibility of traditional AI models in complex clinical scenarios, and realizes a paradigm shift from black box to human-machine collaboration.

[0043] On the one hand, the solution does not stop at the initial prediction of multimodal feature fusion and Transformer model, but innovatively introduces a risk dynamic correction mechanism based on virtual reality interaction. When the initial judgment of risk is high, the system does not give a conclusion directly, but opens up an interactive correction space for medical staff. Medical staff can use their clinical experience to directly adjust the functional weights or hemodynamic parameters of specific liver lobes on the virtual liver model. This design seamlessly injects the prior knowledge that is difficult for medical staff to quantify into the prediction process through an intuitive interactive method, so that the final result is no longer a cold algorithm output, but a precise judgment that integrates medical wisdom.

[0044] On the other hand, this correction mechanism is not a simple parameter superposition; its internal logic is rigorous and adaptive.

[0045] By setting tiered triggering conditions and convergence criteria, the system can intelligently identify the intervention intentions of medical staff and scientifically quantify the correction intensity using dynamically adjusted correction coefficients. This avoids drastic fluctuations in prediction results caused by human intervention. This design not only ensures the stability and rationality of the correction process but also allows the AI ​​model to indirectly learn decision-making logic through interaction with medical staff. This provides a valuable data foundation for the model's future self-iteration and optimization, truly achieving complementary advantages between humans and machines and significantly improving the credibility and practicality of clinical decision-making. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the method for predicting the progression of liver failure risk based on virtual reality in this invention.

[0047] Figure 2 This is a modular schematic diagram of the virtual reality-based liver failure risk progression prediction system of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1: Please refer to Figure 1This embodiment provides a method for predicting the progression of liver failure risk based on virtual reality. The specific implementation steps of this prediction method are as follows: S1. Acquire multimodal medical image data, perform standardized preprocessing, generate a fused image dataset, segment the target region to obtain a liver region mask, and construct a vascular network topology based on the liver region mask; this step aims to acquire complete liver medical image data of the target patient and perform corresponding processing, providing effective data as a foundation for subsequent 3D modeling and risk prediction; the detailed explanation of S1 is as follows:

[0050] S1.1 Acquisition of Multimodal Medical Imaging Data:

[0051] Abdominal enhanced computed tomography (CT) images of the target patient were acquired and denoted as CT, magnetic resonance imaging (MRI) images as MR, and ultrasound elastography (LS) images as LS. Among them, CT images were acquired using scanning parameters of 0.5 mm slice thickness and 0.5 mm slice spacing to obtain three-phase enhanced images of the arterial phase, portal venous phase, and delayed phase; MR images were acquired using T1-weighted, T2-weighted, and diffusion-weighted imaging sequences, i.e., DWI; LS data were acquired using liver stiffness values ​​(LSM data). Therefore, the multimodal medical imaging data in this embodiment includes: CT images, MR images, and LS data.

[0052] S1.2 Standardized preprocessing:

[0053] Specifically, this involves image registration and fusion of multimodal medical image data. Different modalities of multimodal medical image data are spatially registered to establish a unified three-dimensional coordinate system. A multimodal image registration algorithm based on mutual information (MI) is employed, with the mathematical expression: MI(I1, I2) = H(I1) + H(I2) - H(I1, I2); where MI represents the mutual information value; I1 represents the reference image (in this embodiment, a CT image is selected as the reference image); I2 represents the image to be registered, which can refer to data other than the CT image in the multimodal medical image data; H(I1) represents the entropy value of the reference image; H(I2) represents the entropy value of the image to be registered; and H(I1, I2) represents the joint entropy value. By maximizing the mutual information value, spatial alignment of different modalities of multimodal medical image data is achieved. After registration, the multimodal medical image data is fused at the pixel level to generate or obtain the corresponding fused image dataset.

[0054] S1.3, Target Region Segmentation (In this embodiment, the target region refers to the liver region):

[0055] An improved 3D U-Net deep learning model is used to segment target regions on a fused image dataset. The improved 3D U-Net deep learning model is represented by introducing an attention gating mechanism on the basis of the traditional U-Net to enhance the model's ability to identify liver boundary regions. In the traditional U-Net architecture, the deep semantic features extracted by the encoder are directly concatenated with the corresponding layers of the decoder through skip connections. However, this direct concatenation will also pass a large amount of background noise and irrelevant tissue features from the encoder into the decoder, affecting the actual segmentation accuracy of fine structures such as liver edges in the liver region. Therefore, the core of introducing the attention gating mechanism is to filter the feature information in the skip connections.

[0056] The specific process of introducing an attention gating mechanism to derive an improved 3D U-Net deep learning model is explained as follows: First, determine the input signal: After each layer upsampling operation of the decoder and before concatenation with the encoder features, an attention gating module is inserted. This module receives two inputs: skip connection features and a gating signal. The skip connection features are shallow, high-resolution feature maps from the corresponding layer of the autoencoder, containing rich spatial details but also background noise. The gating signal comes from a deep, low-resolution feature map from the layer above the decoder, containing high-level semantic information that guides the model to focus on the target region. Second, linear transformation and dimensionality reduction: The skip connection features and the gating signal are subjected to 1... The first step is a 1×1×1 3D convolution operation, which maps the number of channels to the same intermediate dimension; for example, it reduces the dimensionality to the intermediate number of channels C_inter for subsequent fusion calculations. The second step is feature fusion and activation: the two feature maps after convolution are added element-wise in space, and the result is passed through a ReLU activation function to introduce non-linearity and filter out some negative irrelevant responses. The third step is to generate attention coefficients: the feature map after ReLU activation is passed through a 1×1×1 convolution layer to compress the number of channels to 1, and the value is mapped to the interval 1 through a Sigmoid activation function. The output of this step is the attention coefficient map.

[0057] Regions with coefficients close to 1: These represent areas that the model considers to be key regions of the liver, such as borders or lesions;

[0058] Regions with coefficients close to 0: These represent areas that the model considers to be irrelevant background or noise.

[0059] The final step, the fifth step, is feature weighting and output: the generated attention coefficient map is multiplied element-wise with the original skip connection features; through this weighting operation, the key regions in the original feature map are magnified and preserved, while the background noise is suppressed and cleared to zero; finally, this feature map purified by the attention mechanism is sent to the decoder and concatenated with the upsampled features.

[0060] The loss function of the improved 3D U-Net deep learning model is defined as:

[0061] L = w1 × Dice_loss + w2 × Cross_entropy_loss + w3 × Boundary_loss; where L represents the total loss function; Dice_loss represents the Dice similarity coefficient loss, used to measure the overlap between the segmentation result and the ground truth annotation; in the application scenario of this embodiment, Dice_loss directly focuses on the overlap ratio of the foreground region, has strong robustness to class imbalance, can ensure the integrity of the liver body, and prevent the model from missing large areas of segmentation; Cross_entropy_loss represents the cross-entropy loss, used to optimize pixel classification accuracy; Boundary_loss represents the boundary loss, used to enhance the segmentation accuracy of the liver boundary. Combining Dice_loss with the introduction of cross-entropy loss can provide the network with a stable and dense... Gradient signals accelerate model convergence and optimize classification accuracy for each voxel. Dice_loss and Cross_entropy_loss often focus only on whether it's a liver, neglecting edge accuracy. In medical images, the liver's grayscale values ​​are very similar to neighboring organs like the kidneys and heart, resulting in extremely blurred edges. Therefore, the added Boundary_loss, through a distance map, imposes a larger penalty on incorrect predictions that are further from the true boundary, forcing the predicted contour to approach the true boundary infinitely. w1, w2, and w3 represent the weight coefficients of the Dice similarity coefficient loss, cross-entropy loss, and boundary loss, respectively, with values ​​ranging from 0 to 1. In this embodiment, w1=0.6, w2=0.3, and w3=0.1. These loss functions, through the established triple constraints, achieve high-precision segmentation.

[0062] After segmenting the target region, an improved 3D U-Net deep learning model is used for model training and inference to obtain a liver region mask. Model training involves inputting the fused image dataset and corresponding real liver annotations into the improved 3D U-Net deep learning model. The predicted mask is calculated through forward propagation, and the error is calculated using a loss function. Backpropagation is then used to update the network parameters, and training is iteratively until the loss converges. Model inference involves inputting the fused image dataset of the patient to be tested into the trained model. The network directly outputs a 3D probability map of the liver region through forward propagation. After threshold segmentation, the final 3D segmentation mask of the liver is obtained, i.e., the aforementioned liver region mask. Threshold segmentation here is the process of converting the probability map output by the model into a binary mask. A threshold, such as 0.5, is typically set. If the probability value of a pixel is greater than or equal to this threshold, it is determined to be liver (foreground) and assigned a value of 1; otherwise, it is determined to be background and assigned a value of 0. Example: If the predicted probability of a pixel is 0.85, since 0.85 > 0.5, the point is retained as liver tissue; if the probability is 0.3, it is filtered out as background.

[0063] By adopting the above technical solutions and introducing an improved 3D U-Net with an attention gating mechanism, the problem of missed segmentation in traditional segmentation models when dealing with blurred liver boundaries and large background noise interference is effectively solved. By using gating signals to filter irrelevant features in skip connections, the model can focus on the liver edge and lesion area. Combined with a composite loss function that includes boundary loss, the segmentation accuracy is greatly improved. The gating attention fusion mechanism is not a simple splicing, but can automatically learn the contribution of different modalities of data to liver failure prediction, suppress the interference of invalid information, and achieve the effect of providing high-purity and high-correlation feature vectors for subsequent risk prediction through refined feature extraction and fusion, thus ensuring the robustness of the prediction model from the source.

[0064] S1.4 Extraction of vascular network topology:

[0065] S1.4.1. Based on the liver region mask obtained from segmentation, a Hessian matrix-based vascular enhancement filtering algorithm is used to extract intrahepatic vascular structures; for each voxel point (x, y, z) in the 3D image, its Hessian matrix H is calculated:

[0066] H = [∂²I / ∂x² ∂²I / ∂x∂y ∂²I / ∂x∂z][∂²I / ∂y∂x ∂²I / ∂y² ∂²I / ∂y∂z][∂²I / ∂z∂x ∂²I / ∂z∂y ∂²I / ∂z²];

[0067] Where I represents the image grayscale value; ∂²I / ∂x² represents the second-order partial derivative in the x-direction, and the meanings of the other symbols are analogous and will not be elaborated here; Eigenvalue decomposition of the H matrix yields three eigenvalues ​​λ1, λ2, and λ3, satisfying |λ1|≤|λ2|≤|λ3|; Note: The Hessian matrix is ​​composed of second-order partial derivatives, which can accurately characterize the local grayscale curvature changes of the image; Liver vessels appear as tubular structures with rapidly changing local grayscale in CT images. The Hessian matrix can accurately extract this anisotropic geometric feature through eigenvalue decomposition, thereby effectively distinguishing blood vessels from surrounding tissues.

[0068] S1.4.2 The response function of the vascular network topology is defined as follows:

[0069] Vesselness={0, if λ2>0 or λ3>0;

[0070] 1-exp(-R_α² / 2α²)×[1-exp(-R_ β² / 2β²)]×[1-exp(-S² / 2c²)], otherwise};

[0071] In this embodiment, Vesselness is a technical term that can be understood as vascular enhancement. R_α = |λ1| / |λ2| represents the anisotropy measure of tubular structures; R_β = |λ2| / |λ3| represents the distinguishing measure of plate-like structures; S = √(λ1² + λ2² + λ3²) represents structural strength; α, β, and c are all adjustable parameters, where α and β range from 0 to 1, and c ranges from greater than 0. In this embodiment, α = 0.5, β = 0.5, and c = 500. This function locks blood vessels through multi-condition constraints; λ2 > 0 or λ3 > 0 is used to filter background and non-tubular structures; R_α uses anisotropy to suppress plaque-like tissue, R_β is used to distinguish between plate-like structures and linear blood vessels, and S ensures that the response has sufficient local structural strength, thus comprehensively ensuring the accuracy of vascular enhancement.

[0072] S1.4.3. Through threshold segmentation, the three major vascular systems of hepatic artery, portal vein, and hepatic vein are extracted to construct a vascular network topology. The threshold segmentation here is to compare Vesselness with a preset response value threshold. In this embodiment, the response value threshold can be set to 0.6. Therefore, when Vesselness exceeds 0.6, it is determined that the point belongs to the intrahepatic blood vessel and is retained; otherwise, it is directly filtered, and then the vascular network topology is finally constructed.

[0073] S2. The obtained liver region mask is converted into a three-dimensional surface mesh model. The liver material is set and texture mapping is performed. A mass-spring model is used to construct a biomechanical model of the liver to realize real-time deformation simulation of liver tissue. Clinical examination indicators are mapped onto the biomechanical model to realize information visualization.

[0074] The specific explanation of S2 is as follows: S2.1, Three-dimensional liver geometric modeling:

[0075] When converting to obtain a 3D surface mesh model, the existing moving cubes algorithm is used to extract liver isosurfaces and generate triangular facet meshes. Specifically, the moving cubes algorithm divides the 3D voxel data of the liver into countless small cubes. The 3D voxel data of the liver is a 3D density matrix composed of stacked CT slices. By setting a preset CT density threshold, the relationship between the eight vertices of each cube and the liver surface is determined, that is, the vertex values ​​are compared with the CT density threshold. When one of the two endpoints of the cube edge exceeds the CT density threshold and the other does not exceed the CT density threshold, linear interpolation is used to calculate the intersection point between the cube edge and the surface, and these intersection points are connected to form triangular facets. Finally, all cubes are traversed and stitched together to generate a complete liver isosurface mesh, thus obtaining the 3D surface mesh model. The specific CT density threshold setting can be determined according to the actual situation and will not be elaborated here. Subsequently, to improve the smoothness of the model surface, the Laplacian smoothing algorithm can be applied to optimize the mesh vertices: vᵢ'=vᵢ+λ×(1 / |N(i)|)×Σⱼ∈N(i)(vⱼ - vᵢ); where vᵢ represents the original coordinates of the i-th vertex, vᵢ' represents the smoothed vertex coordinates, λ represents the smoothing coefficient, with a value range of [0, 1], and in this embodiment λ=0.3; N(i) represents the set of vertices adjacent to vertex i, |N(i)| represents the number of adjacent vertices; vⱼ represents the coordinates of adjacent vertices; after smoothing, the number of triangular faces of the liver surface mesh is controlled between 50,000 and 100,000, which ensures both model detail and meets the performance requirements of real-time rendering.

[0076] S2.2, Liver Material and Texture Mapping:

[0077] Instructions for setting up the liver material: To enhance the realism of the virtual liver model, physically based rendering (PBR) technology is used for material settings. The method is as follows: Physically based rendering (PBR) mainly simulates realistic optical properties by adjusting material parameters and textures. The settings typically include the following core parameters:

[0078] Base color: Sets the inherent color of the liver tissue, such as dark red, without including lighting and shadow information;

[0079] Metallicity: The liver is a non-metallic biological tissue, so this value should be set to 0;

[0080] Roughness: Simulates the moist, slightly raised texture of the liver surface, and is usually set to a medium to low value. For example, in this embodiment, it is usually set in the range of 0.2 to 0.4 to present a soft, moist highlight.

[0081] Normal mapping: Adds microscopic bump details to the surface, enhancing the sense of three-dimensionality.

[0082] The optical properties of liver tissue are described by the bidirectional reflectance distribution function (BRDF): fᵣ(ωᵢ,ωᵣ) = f_d(ωᵢ,ωᵣ) + f_s(ωᵢ,ωᵣ); where fᵣ represents the total reflectance distribution function, ωᵢ represents the incident light direction, ωᵣ represents the reflected light direction, f_d represents the diffuse component (using the existing Lambertian model), and f_s represents the specular component (using the existing Cook-Torrance model). It should be noted that BRDF is the underlying mathematical core of PBR technology. PBR material settings essentially provide physical parameters to the BRDF formula. Diffuse and specular reflection in the above formula are the two main components of BRDF. Setting the base color provides color data for the diffuse component; setting roughness and metallicity controls the intensity of light scattering and reflection from micro-surfaces in the specular component. In short, BRDF is the physical law for calculating light and shadow, while PBR material settings are the input parameters that conform to the real characteristics of the liver, driving this law to render realistic visual effects.

[0083] The explanation of the texture mapping operation for the liver material is as follows: The liver surface texture is generated by extracting grayscale information from CT images to generate normal maps and roughness maps; the generation of the normal map uses the Sobel operator to calculate the image gradient: N_x=G_x×I; N_y=G_y×I; N_z=1-√(N_x²+N_y²); where N_x, N_y, and N_z represent the x, y, and z components of the normal vector, respectively, and G_x and G_y represent the Sobel convolution kernels in the x and y directions, respectively;

[0084] The above description describes the preliminary material generation steps for texture mapping. Ordinary CT grayscale images can only provide color and cannot represent the texture of bumps and shadows. The normal map generated by the above calculation records the bump direction and roughness map (this is a conventional technology and will not be elaborated on further), and records the surface gloss. It is an indispensable core texture in the PBR material system, which directly determines how the rendering engine calculates the reflection and scattering of light on the liver surface. It should be noted that the core logic of this formula is to derive the three-dimensional surface normal from the two-dimensional grayscale changes. The Sobel operator is convolved with the image grayscale value I to calculate the grayscale gradient of the pixel in the horizontal and vertical directions, that is, N_x and N_y, which reflect the steepness of the surface. Since the normal vector is a unit vector with a length of 1, the Z-axis component perpendicular to the surface can be calculated from the above formula. Note: Standard physical calculations usually involve taking the square root and normalizing, but this is an approximate mapping in engineering. Finally, these three components are mapped to RGB color values ​​to generate a normal map that records the bump direction of the surface.

[0085] Texture mapping is the process of wrapping a 2D image onto a 3D model. First, using UV unwrapping technology, two-dimensional texture coordinates (U, V) are assigned to each vertex of the liver's 3D mesh. Then, in the rendering engine's material system (i.e., setting actions corresponding to the liver material), the generated normal map is connected to the normal channel, the roughness map is connected to the roughness channel, and the color map extracted by CT is connected to the base color channel. During rendering, the GPU reads the texture pixels according to the UV coordinates, combines the bump data of the normal map with the gloss data of the roughness map, and solves the lighting equation in real time to finally present a liver surface with realistic texture.

[0086] S2.3, Liver biomechanical modeling:

[0087] To simulate the deformation characteristics of liver tissue, a mass-spring model is used to construct the biomechanical model of the liver. Each vertex of the mesh in the three-dimensional surface model corresponding to the liver is considered a mass, and adjacent vertices are connected by springs. The motion of the mass follows Newton's second law: mᵢ_1×aᵢ_1=F_ext+Σⱼ_1∈N(i_1)k×(‖xⱼ_1-xᵢ_1‖-L0)×(xⱼ_1-xᵢ_1) / ‖xⱼ_1-xᵢ_1‖+Σⱼ_1∈N(i_1)c×(vⱼ_1-vᵢ_1)×(xⱼ_1-xᵢ_1) / ‖xⱼ_1-xᵢ_1‖; In the equation, mᵢ_1 represents the mass of the i_1th particle, aᵢ_1 represents the acceleration of the particle, F_ext represents the external force, which in this embodiment can be represented as the contact force of the surgical instrument, k represents the spring constant, L0 represents the natural length of the spring, xᵢ_1 and xⱼ_1 represent the position vectors of particles i_1 and j_1 respectively, vᵢ_1 and vⱼ_1 represent the velocity vectors of the particles respectively, c represents the damping coefficient, and N(i_1) represents the set of particles connected to particle i_1. The above differential equation is solved by numerical integration methods, such as Verlet integral, to realize real-time deformation simulation of liver tissue.

[0088] It should be noted that this formula is based on Newton's second law, discretizing liver tissue into a point mass spring model. The left side of the equation represents the inertial force of the point mass, i.e., mass multiplied by acceleration; the right side contains three resultant forces: external force, spring elastic restoring force, which is determined by the elastic coefficient k and deformation, giving the tissue its rebound characteristics, and damping force, which is determined by the damping coefficient c and relative velocity, simulating the viscosity and energy dissipation of the tissue. These three forces together determine the acceleration of the point mass, thereby precisely driving the entire soft tissue mesh to produce real-time deformation conforming to biomechanical characteristics. In this embodiment, Verlet product is used. The Verlet integral is used to solve the above differential equation. Since this is a current technique, only a brief overview is given: The Verlet integral updates the position at the next moment by using the historical state of the position, avoiding the cumulative error of directly calculating the velocity. Its core is to use the current position and the previous position, combined with the acceleration calculated by the current force and the square of the time step, to deduce the new position at the next moment. Then, the velocity is estimated by dividing the difference between the old and new positions (i.e., the absolute difference) by twice the time step. This method has extremely high numerical stability and time reversibility, and is suitable for high-frequency iterative real-time deformation simulation.

[0089] S2.4, Mapping of Clinical Examination Indicators:

[0090] Specifically, the clinical examination indicators mentioned in this embodiment include: liver function reserve parameters, LS data included in the multimodal medical imaging data in S1.1, and hemodynamic parameters. The clinical laboratory examination indicators are mapped onto a biomechanical model to visualize functional information. Specifically, this includes: liver function reserve parameter mapping: converting indicators such as serum albumin (ALB) and prothrombin time (PT) into overall liver color codes, using a gradient chromatogram from green to red, corresponding to the progression from normal to failure; it should be noted that liver function reserve parameters include serum albumin (ALB) and prothrombin time (PT); liver stiffness parameter mapping: mapping LS data onto the liver surface and displaying it in the form of a heatmap, with higher stiffness values ​​leaning more towards red; and hemodynamic parameter mapping: based on vascular structure, simulating parameters such as portal vein blood flow velocity and hepatic artery blood flow, displaying them in the vascular network topology using streamline visualization technology; therefore, in this embodiment, hemodynamic parameters include portal vein blood flow velocity and hepatic artery blood flow.

[0091] Example: In the case of a 56-year-old male patient, a liver surface mesh model containing 86,420 triangular facets was generated through S2.1. After setting the material in S2.2, the liver surface exhibited realistic tissue texture and optical properties. In S2.3, the mass mᵢ_1=0.01kg, elastic coefficient k=50N / m, and damping coefficient c=5N·s / m were set, successfully simulating the soft characteristics of liver tissue. Finally, through S2.4, the patient's parameters such as ALB=32g / L, PT=16.5s, and LSM=18.6kPa were mapped onto the virtual liver model, intuitively displaying the decline in liver function reserve and the degree of liver fibrosis.

[0092] S3. Extract target multimodal features and use a gated attention fusion mechanism to complete feature fusion; construct a multi-task deep learning prediction model and output preliminary prediction results simultaneously; based on the preliminary prediction results, trigger a risk dynamic correction mechanism based on virtual reality interaction and output the final prediction results; among which, the prediction results (i.e., whether it is the preliminary prediction result or the final prediction result) include: predicting the probability of liver failure, the progression time window and the severity grade.

[0093] S3 is the core of this invention. It designs a risk dynamic correction mechanism based on virtual reality interaction, allowing clinicians or other medical staff to adjust corresponding parameters through a virtual reality interface, observe risk change trends, and perform iterative optimization. Specific details are as follows:

[0094] S3.1 Multimodal Feature Extraction and Fusion:

[0095] The extracted target multimodal features include: image feature vectors, clinical examination indicator feature vectors, and genomic feature vectors. Specifically, regarding image feature vector extraction: the existing 3D ResNet-50 network is used to extract the depth features of CT / MR images contained in multimodal medical image data. This network contains 5 residual blocks, and the output of each residual block is: y_1=F(x_1,{Wᵢ_2})+x_1; where x_1 represents the input feature, y_1 represents the output feature, F represents the residual function, which is a nonlinear mapping composed of several convolutional layers, used to learn the difference between the input and the expected output, i.e., fine-tuning the perturbation term, which will not be elaborated here. Wᵢ_2 represents the weight parameter of the i_2th layer, with a value range of 0~1; after a global average pooling layer, a 2048-dimensional image feature vector is obtained.

[0096] It should be noted that this network extracts features through a residual learning mechanism. The input data is propagated forward layer by layer. In each residual block, the convolutional layer extracts a nonlinear feature map F(x) and directly superimposes the original input x_1 onto the output through skip connections, thereby effectively alleviating the gradient vanishing problem and preserving the original information. After deep processing through multiple residual blocks, the network captures rich three-dimensional spatial and semantic features. Finally, the global average pooling layer performs spatial dimension aggregation and compression on the high-dimensional feature map, flattening it into a fixed-length feature vector of 2048 dimensions.

[0097] Regarding the extraction of feature vectors for clinical examination indicators: After standardizing the clinical examination indicators and demographic features, the data are input into a fully connected neural network to obtain 128-dimensional feature vectors for the clinical examination indicators. Specifically, demographic features include age, gender, etc. During standardization, due to the different data types, Z-score standardization is used for continuous numerical data to eliminate dimensional differences. For example, serum albumin (ALB), prothrombin time (PT), and LS data in clinical examination indicators, as well as age in demographic features, are all continuous numerical data. One-hot encoding is used for categorical data to facilitate input into the subsequent fully connected neural network.

[0098] The following is an explanation of fully connected neural networks:

[0099] In this embodiment, the fully connected neural network (FCN) fuses multi-source clinical examination indicator features into a high-dimensional semantic vector through multi-layer linear transformation and non-linear activation. The specific derivation process is as follows: 1. Input layer construction: Standardized continuous numerical data is concatenated with one-hot encoded categorical features to form the initial input vector; 2. Linear weighted transformation: Data enters the hidden layer, and each neuron in the hidden layer is fully connected to the previous layer. Assuming the weight matrix of the L1 layer is Q... [L1] The bias vector is b [L1] Then the linear output Z of the Lth layer[L1] Z [L1] =Q [L1] ×A [L1-1] +b [L1] A represents the initial input vector, and the weights Q represent the contribution of each clinical examination indicator to the final feature; 3. Nonlinear activation: The result Z after linear transformation [L1] By introducing nonlinear factors through activation functions such as ReLU, complex feature interactions can be extracted: A [L1] =Activation(Z [L1] 4. Feature Compression and Output: After processing through several hidden layers, the data is mapped to a 128-dimensional latent space. This layer acts as a feature bottleneck, forcing the network to retain 128 core feature values ​​that best represent the liver's functional state, such as the degree of cirrhosis and hemodynamic abnormalities. The final output vector is A. [final] That is, a 128-dimensional feature vector of clinical examination indicators.

[0100] Regarding genomic feature vector extraction: Under the condition that gene testing data exists in the target patient, single nucleotide polymorphism (SNP) site information related to liver failure is extracted and converted into a 64-dimensional genomic feature vector through an embedding layer;

[0101] Multimodal feature fusion employs a gated attention fusion mechanism, as described above, based on the following formula: F_fused = σ(W_g × [F_img; F_cli; F_gen]) ⊙ [F_img; F_cli; F_gen]; where F_fused represents the fused feature vector, F_img, F_cli, and F_gen represent the image feature vector, clinical examination indicator feature vector, and genomic feature vector, respectively, [·; ·; ·] represents the vector concatenation operation, W_g represents the gate weight matrix, σ represents the Sigmoid activation function, and ⊙ represents element-wise multiplication. It should be noted that target patients with gene testing data refer to those who have undergone gene testing and have complete results. SNPs are mapped to dense vectors through the embedding layer and participate in feature fusion as one of the input modalities of the gated attention mechanism; this will not be elaborated upon further here.

[0102] S3.2 Construction of a multi-task deep learning prediction model:

[0103] The backbone network of the model adopts a Transformer encoder structure, which contains 6 encoder layers. Each encoder layer contains a multi-head self-attention mechanism and a feedforward neural network.

[0104] The calculation formula for the multi-head self-attention mechanism is as follows:

[0105] MultiHead(Q, K, V) = Concat(head1, ..., head_H) × W^O; where MultiHead represents a multi-head self-attention mechanism, Q, K, and V represent the query, key, and value matrices, respectively, Concat represents concatenation, and H represents the number of attention heads. In this embodiment, head_h = Attention(Q × W_i^Q, K × W_i^K, V × W_i^V); where Attention represents attention, W_i^Q, W_i^K, and W_i^V represent the h-th attention head, i.e., the query, key, and value projection of head_h. The matrix W^O represents the output projection matrix, so h = 1, 2, ..., H, where H is the number of attention heads and also the maximum value of h; Attention(Q, K, V) = softmax(Q×K^T / √d_k)×V; where d_k represents the dimension of the key vector, and this formula represents the weighted aggregation of information based on similarity, and T represents the matrix transpose, used to calculate the dot product of Q and K to measure relevance; it should be noted that the calculation formula of the multi-head self-attention mechanism enables the model to capture complex dependencies of multiple dimensions such as syntax and semantics simultaneously by computing the attention of multiple subspaces in parallel, and the final concatenation and fusion can significantly enhance the expressive power and robustness of the model.

[0106] The feedforward neural network consists of two fully connected layers, with the GELU activation function in between.

[0107] FFN(xt) = W2×GELU(W1×xt+b1)+b2; where xt represents the previous sublayer, usually the output vector after multi-head self-attention mechanism and residual connection and layer normalization operation, W1 and W2 represent the weight matrices of the first and second layers respectively, b1 and b2 represent the bias terms respectively, and GELU represents the Gaussian error linear unit activation function, which is well known in the field and will not be described in detail here.

[0108] The output layer of a multi-task deep learning prediction model contains three parallel prediction heads (corresponding to preliminary prediction results):

[0109] (1) Risk occurrence prediction head: adopts the Sigmoid activation function to output the probability of liver failure, with a value range of [0, 1]; (2) Progression time prediction head: adopts the linear activation function to output the expected time to progress to liver failure, in days; (3) Severity grading prediction head: adopts the Softmax activation function to output the probability distribution of mild, moderate, and severe liver failure; It should be noted that the activation functions used in the above three prediction heads are all mature existing technologies in the field of deep learning, and there is no need to introduce their underlying mathematical principles in detail. Here is a brief summary. The model output layer adopts a multi-task parallel architecture, and matches the physical characteristics of different tasks. The model employs standard activation strategies for risk prediction: for binary classification problems, the Sigmoid function is used to map the output to intervals, directly representing the independent probability of liver failure; for progression time prediction, a linear activation function is used for regression tasks to maintain the continuity of the output and ensure accurate prediction of specific number of days; for severity grading, a Softmax function is used to normalize the output vector for multi-class classification problems, making the sum of the probabilities of each level equal to 1, thus providing a clear distribution of disease states. This design based on standard functions not only conforms to existing technical paradigms but also ensures that the model has extremely high stability and interpretability when considering the prediction of multiple clinical indicators.

[0110] The model's total loss function is:

[0111] L_total = w4 × BCE(p, p̂) + w5 × MAE(t, t̂) + w6 × CE(s, ŝ); where BCE represents the binary cross-entropy loss, used for risk occurrence prediction; MAE represents the mean absolute error loss, used for progression time prediction; CE represents the cross-entropy loss, used for severity grading prediction; p, t, and s represent the true predicted probability label of liver failure, progression time window label, and severity grading label, respectively; p̂, t̂, and ŝ represent the corresponding predicted values, respectively; and w4, w5, and w6 are all loss weight coefficients, with values ​​ranging from 0 to 1.

[0112] S3.3, Risk dynamic correction mechanism based on virtual reality interaction:

[0113] This step is the core of this solution. By designing the aforementioned risk dynamic correction mechanism based on virtual reality interaction, clinicians are allowed to adjust prediction parameters, observe risk change trends, and perform iterative optimization through a virtual reality interface. This risk dynamic correction mechanism based on virtual reality interaction includes setting primary correction conditions, setting secondary correction conditions, and judging convergence of multiple corrections. The prediction parameters are the functional parameters of the liver, liver hemodynamic parameters, etc.

[0114] The specific breakdown is as follows:

[0115] S3.3.1 Setting the conditions for a single correction:

[0116] The trigger condition for a single correction is set, namely condition A: when the predicted probability of liver failure p̂ is greater than the preset first threshold θ1, the single correction mode is entered; where the preset first threshold θ1 ranges from 0 to 1, and in this embodiment θ1=0.6; in the single correction mode, clinicians can select target areas on the liver model, such as the right lobe, left lobe, caudate lobe, etc., using virtual reality controllers, and adjust the functional parameters of the target areas, such as liver stiffness values, etc.

[0117] The specific correction formula is: p̂'=p̂+Δp; where p̂' represents the corrected risk probability, Δp represents the risk correction amount, based on the formula: Δp=a1×(w_region-w_baseline)×f_region; where a1 is the first correction coefficient, with a value range of [0.1, 0.3], and in this embodiment a1=0.2; w_region represents the weight coefficient of the target region, with a value range usually greater than 1; specifically, the weight coefficient w_region of the target region is determined in the following way:

[0118] The first step is initial setup: the system (which corresponds to the system in the subsequent embodiment 2) is based on liver anatomy knowledge and clinical data to preset initial weights for the target region, such as 1.2 for the right lobe, 1.1 for the left lobe, and 1.3 for the caudate lobe. The second step is dynamic adjustment: doctors can make real-time adjustments by dragging sliders or inputting values ​​in the virtual reality interface. The third step is ensuring that the baseline is exceeded: the system interface sets a minimum value constraint. When w_region≤w_baseline, Δp≤0, and risk adjustment cannot be achieved. Therefore, when doctors need to improve risk assessment, they will naturally choose a value of w_region>w_baseline. The system can also set the input range to [1,2] at the UI level to force the effectiveness of the adjustment.

[0119] w_baseline represents the baseline weighting coefficient, which is usually set to 1 in this embodiment. f_region represents the functional abnormality score of the target region, with a value range of [0, 1]. Its specific value is obtained by calculating the percentage of liver fat content and normalizing it to the range of [0, 1] based on magnetic resonance liver fat quantification (MR-PDFF) technology, or by using the Child-Pugh grading system score through linear mapping transformation. The specific value can be selected according to the actual needs.

[0120] Example: Continuing with the case of the 56-year-old male patient mentioned above, the predicted probability of liver failure p̂=0.72, which is greater than θ1=0.6, prompts the system to enter a correction mode. The doctor selects the right lobe of the liver using a virtual reality controller and observes that the liver stiffness value in this region is relatively high (LSM=22.3kPa). The weight coefficient of this region is adjusted to w_region=1.3, and the functional abnormality score is set to f_region=0.8. The calculated Δp=0.2×(1.3-1)×0.8=0.048, and the corrected risk probability is p̂'=0.72+0.048=0.768.

[0121] S3.3.2, Setting the Secondary Correction Conditions:

[0122] Set the trigger condition for the second correction, namely condition B: when the change in risk probability after the first correction |Δp| ​​is greater than the preset second threshold θ2, and the corrected risk probability p̂' is still greater than the preset first threshold θ1, the directional index of the first correction amount (i.e. the risk correction amount) satisfies: sign(Δp)=sign(ΔROI_score), that is, the direction of risk change after the first correction is consistent with the direction of change of the region of interest score adjusted by the doctor.

[0123] It should be noted that ΔROI_score represents the amount of adjustment made by healthcare professionals to the degree of functional abnormality in the region of interest within the virtual reality interface. It is obtained by the system monitoring in real time the doctor's actions of dragging a slider with a controller or inputting a numerical value to score the region of interest, calculating the difference between the adjusted score f_new and the initial score f_old, i.e., ΔROI_score = f_new - f_old. This indicator quantifies the subjective intervention direction of healthcare professionals, i.e., doctors, in this region. When ΔROI_score > 0, it indicates that the doctor believes the condition in that region has worsened; conversely, it indicates that the condition has improved. By comparing whether sign(Δp) and sign(ΔROI_score) are consistent, the system can verify whether the direction of risk change in the model is consistent with the doctor's clinical judgment. The initial score can be obtained based on magnetic resonance fat quantification or Child-Pugh classification normalization, while the adjusted score is input by the doctor in the virtual reality interface.

[0124] If all three sub-conditions in condition B above are met, then entering the secondary correction mode is permitted;

[0125] The preset second threshold θ2 ranges from 0 to 1, and in this embodiment, θ2 = 0.03. It should be noted that the correlation between condition B and condition A is that condition B requires that the first correction produces a significant change in risk, i.e., |Δp| ​​> θ2. This indicates that there is a large difference between the doctor's functional assessment of the target area and the initial model prediction, and other relevant parameters need to be further adjusted to obtain more accurate prediction results. In the second correction mode, the doctor can adjust liver hemodynamic parameters, such as portal vein pressure (PVP) and hepatic artery blood flow (HAF).

[0126] The secondary correction formula is: p̂''=p̂'+Δp'; where p̂'' represents the risk probability after secondary correction, and Δp' represents the secondary correction amount (i.e., the secondary risk correction amount), according to the formula: Δp'=a2×(ΔPVP / PVP_baseline+ΔHAF / HAF_baseline); where a2 is the second correction coefficient, with a value greater than 0, and in this embodiment a2=0.1; ΔPVP represents the change in portal vein pressure, PVP_baseline represents the baseline portal vein pressure, usually taken as 10 mmHg; ΔHAF represents the change in hepatic artery blood flow, HAF_baseline represents the baseline hepatic artery blood flow, usually taken as 300 ml / min;

[0127] For a2, unlike the traditional fixed coefficient, it is set to be dynamic in this application, with a second correction coefficient dynamic adjustment mechanism, based on: a2=a2_0×(1+t0×|Δp| / θ2); where a2_0 is the base coefficient, with a value range of 0~1, and in this embodiment the value can be set to 0.1; t0 is the sensitivity coefficient, with a value range of [0.2, 0.5], and in this embodiment t0=0.3;

[0128] Table 1. Comparison of Optimization Effects by Dimension:

[0129] Conditional judgment Based on threshold only Add directional consistency verification Avoid incorrect corrections and improve correction accuracy. coefficient adjustment Fixing a2 is equivalent to a2_0 Dynamic a2 and secondary utilization of Δp More scientific correction Clinical significance Mechanical correction Combining the doctor's intentions and the desired effect It aligns better with clinical decision-making logic and enhances credibility.

[0130] Table 1 above shows the specific operational effect of the dynamic adjustment mechanism for the second correction coefficient.

[0131] In summary, compared with traditional solutions, the technical advantages of this solution are as follows:

[0132] Directional consistency verification ensures that the doctor's correction intention is consistent with the model response, avoiding prediction distortion caused by reverse correction; dynamic coefficient adjustment makes the amount of secondary correction proportional to the significance of the primary correction, making the correction force more scientific; the comprehensive judgment mechanism raises the trigger threshold for secondary correction, reduces unnecessary correction operations, and thus improves system stability; this scheme, while maintaining the original traditional scheme framework, significantly improves the rationality and clinical applicability of secondary correction by introducing directional verification and dynamic coefficient adjustment.

[0133] Example: Continuing with the case in S3.3.1, the change in risk probability after the first correction is |Δp|=0.048>θ2=0.03, and p̂'=0.768>θ1=0.6, which also meets the directional index, satisfying condition B. The system enters the second correction mode. The doctor observes through the virtual reality interface that the patient's portal vein blood flow velocity has slowed down, and adjusts the portal vein pressure to PVP=15mmHg, ΔPVP=5mmHg, and the hepatic artery blood flow to HAF=250ml / min, ΔHAF=-50ml / min. The calculated value is Δp'=0.1×(5 / 10+(-50) / 300)=0.1×(0.5-0.167)=0.0333.

[0134] Therefore, the risk probability after the second correction is p̂''=0.768+0.0333=0.8013.

[0135] The core advantage of this solution in terms of data reusability lies in transforming the process data of a first correction, rather than the result data, into the decision-making basis for a second correction. This is in stark contrast to traditional solutions. In traditional methods, each correction is isolated, and the system only records the final value after correction, while process information such as the magnitude and direction of previous corrections is completely discarded.

[0136] This solution creatively reuses key process data, namely the risk change Δp of a single correction;

[0137] The system uses the magnitude of |Δp| ​​as a benchmark to measure the intensity of intervention by medical staff, and dynamically adjusts the second correction coefficient a2 to ensure that the correction intensity is proportional to the intensity of intervention, thus achieving self-adaptation in the correction process. Simultaneously, the scheme introduces directional verification of sign(Δp) to ensure that the secondary correction remains consistent with the initial intent of the medical staff, avoiding the reverse distortion of prediction results caused by parameter adjustments. This deep reuse of process data solves the technical problem of contextual information fragmentation in traditional human-computer interaction, achieving coherent intelligent correction with memory and logic. The operation of this scheme transforms the system from a passively responding tool into an intelligent collaborator that understands and continues the decision-making process of medical staff, significantly improving the accuracy and clinical reliability of predictions.

[0138] S3.3.3, Multiple corrections to convergence judgment:

[0139] Set the convergence condition: When the risk correction amount is less than the preset third threshold θ3 for two consecutive times, the correction process is considered to have converged, and the final prediction result is output. The preset third threshold θ3 is also set to a value range of 0~1. In this embodiment, θ3=0.01. For multiple corrections more than twice, the distinction between first-time and second-time correction types is no longer strictly made. Instead, a unified interactive correction process is used: after the second-time correction mode is completed, if further correction is still needed, the correction will continue without having to satisfy condition A or condition B again. The operation method is to randomly select the adjustment object: ① readjust the functional parameters of the target area (similar to first-time correction); ② readjust the liver hemodynamic parameters (similar to second-time correction); ③ mix and adjust any one of the multiple parameters.

[0140] Example: During the third correction, the parameters of the left lobe of the liver were adjusted, and Δp''=0.008<θ3=0.01 was calculated; during the fourth correction, Δp'''=0.006<θ3=0.01 was calculated; since the changes in the two consecutive corrections, i.e., the risk corrections, are both less than θ3, the correction process converges, and the final output is the probability of liver failure, with an estimated progression time of 45 days and a severity grade of severe; the specific explanations of the progression time prediction head and the severity grade prediction head are not detailed here, and you can set your own standards and requirements according to the situation to meet the actual needs.

[0141] The core effect of this solution is that it fundamentally solves the inherent problems of poor interpretability and insufficient flexibility of traditional AI models in complex clinical scenarios by constructing a dynamic architecture of AI prediction-interactive correction-model relearning, and realizes a paradigm shift from black box to human-machine collaboration.

[0142] On the one hand, the solution does not stop at the initial prediction of multimodal feature fusion and Transformer model, but innovatively introduces a risk dynamic correction mechanism based on virtual reality interaction. When the initial judgment of risk is high, the system does not give a conclusion directly, but opens up an interactive correction space for medical staff. Medical staff can use their clinical experience to directly adjust the functional weights or hemodynamic parameters of specific liver lobes on the virtual liver model. This design seamlessly injects the prior knowledge that is difficult for medical staff to quantify into the prediction process through an intuitive interactive method, so that the final result is no longer a cold algorithm output, but a precise judgment that integrates medical wisdom.

[0143] On the other hand, this correction mechanism is not a simple parameter superposition; its internal logic is rigorous and adaptive.

[0144] By setting tiered trigger conditions (conditions A / B) and convergence criteria, the system can intelligently identify the intervention intentions of medical staff and scientifically quantify the correction intensity using dynamically adjusted correction coefficients (such as a2). This avoids drastic fluctuations in prediction results caused by human intervention. This design not only ensures the stability and rationality of the correction process but also allows the AI ​​model to indirectly learn decision-making logic through interaction with medical staff. This provides a valuable data foundation for the model's future self-iteration and optimization, truly achieving complementary advantages between humans and machines and significantly improving the credibility and practicality of clinical decision-making.

[0145] S4. Based on the final prediction results, construct a virtual reality risk process interaction form, including: a timeline of the liver failure risk process, simulation of the dynamic evolution of liver status, and virtual reality interactive interface settings.

[0146] Specifically, the solution described in S4 is used to immerse and visualize the final prediction results obtained in S3 using virtual reality technology, providing an intuitive display and interactive experience of the risk process. A detailed breakdown is provided below:

[0147] S4.1 Construction of the timeline of liver failure risk progression:

[0148] A timeline for the progression of liver failure risk is constructed, starting from the target patient's current time and extending into the future to the predicted time of liver failure onset; key time nodes are marked on the timeline, including:

[0149] (1) Current time point T0: Displays the current liver function status of the target patient;

[0150] (2) Warning time point T1: The time point when the predicted probability of liver failure reaches 0.5;

[0151] (3) High-risk time point T2: The time point when the predicted probability of liver failure reaches 0.8;

[0152] (4) Predicted time point T3: The time of liver failure predicted by the model.

[0153] S4.2 Simulation of the dynamic evolution of liver status:

[0154] Based on the time axis, the dynamic evolution of the liver state from the current state to the state of liver failure is simulated. The evolution simulation includes changes in liver morphology and liver color.

[0155] The liver morphological changes were simulated using a liver fibrosis progression model, which depicted the shrinkage of liver volume and the formation of surface nodules. It should be noted that the liver fibrosis progression model is an existing experimental research model, primarily including chemically induced animal models such as those using carbon tetrachloride and thioacetamide, used to simulate the pathological process of liver fibrosis; these will not be elaborated upon here. The rate of change in liver volume adopted an exponential decay model: V(t) = V0 × exp(-k_t × t); where V(t) represents the liver volume at time t, V0 represents the initial liver volume, k_t represents the decay rate constant, which is positively correlated with the risk probability and ranges from 0 to 1, and t represents time. Using a continuous exponential decay model, the true pathological process of liver fibrosis progressing to cirrhosis can be simulated very accurately: in the early and middle stages of the disease, massive hepatocyte necrosis and fibrous tissue proliferation lead to rapid volume shrinkage; however, in the late stage, the liver structure is severely damaged, and the rate of volume shrinkage slows significantly, gradually approaching a stable minimum value. This perfectly matches the mathematical curve characteristics of rapid decline in the early stage and gradual flattening in the later stage of exponential decay.

[0156] Liver color change: Based on the degree of decline in liver function reserve, the process of liver color changing from normal reddish-brown to dark red and grayish-white is simulated; the color change is expressed by linear interpolation: C(t) = C_normal + (C_failure - C_normal) × p(t); where C(t) represents the liver color at time t, C_normal represents the normal liver color, C_failure represents the color of the liver failure state, and p(t) represents the probability of liver failure at time t;

[0157] Of course, evolution simulation can also include other things such as changes in blood flow in blood vessels, but we will not elaborate on them here.

[0158] S4.3 Virtual Reality Interactive Interface Settings:

[0159] Equipped with an immersive virtual reality interactive interface, including a main view: displaying a 3D liver model and its dynamic evolution process, supporting users to rotate, zoom, and pan using a controller; a risk dashboard: displaying key indicators such as the current probability of liver failure, progression time window, and severity grade, corresponding to the final prediction result; a time controller: allowing users to drag the time slider to observe the liver state at different time points; a parameter adjustment panel: providing sliders for adjusting liver function parameters and hemodynamic parameters, supporting users to perform hypothetical analysis, i.e., What-If Analysis; and an annotation tool: allowing users to add annotations to the liver model to record key observation points.

[0160] Example: For the aforementioned 56-year-old male patient, a liver model was observed in a virtual reality headset to gradually evolve from its current state (volume 1250ml, color reddish-brown) to a state of liver failure 45 days later (volume 980ml, color dark red). The risk dashboard showed a current probability of liver failure of 0.8093, an estimated progression time of 45 days, and a severity grade of severe. The doctor used the time controller to rewind to the state 30 days prior and observed that the probability of liver failure at that time was only 0.45, verifying the accelerating trend of disease progression.

[0161] S5. Complete risk stratification and early warning actions based on the final forecast results;

[0162] The risk stratification and early warning actions are based on the following: according to the probability of liver failure in the final prediction results, different patients are classified into risk levels and the corresponding decision suggestion texts in the decision template library are retrieved.

[0163] The specific process is as follows: Based on the probability of liver failure in the final prediction results, patients are divided into four risk levels:

[0164] (1) Low risk: The probability of liver failure is <0.3. The corresponding decision recommendation text is: It is recommended to have routine follow-up and re-examination every 6 months. (2) Medium risk: 0.3≤The probability of liver failure is <0.6. The corresponding decision recommendation text is: It is recommended to closely monitor and re-examine every 3 months. (3) High risk: 0.6≤The probability of liver failure is <0.85. The corresponding decision recommendation text is: It is recommended to actively intervene, re-examine monthly, and consider liver transplantation evaluation. (4) Very high risk: The probability of liver failure is ≥0.85. The corresponding decision recommendation text is: It is recommended to intervene urgently, re-examine weekly, and prioritize liver transplantation evaluation. In summary, the corresponding threshold settings for each risk level can also be changed according to actual needs, which will not be elaborated here.

[0165] Example 2: Based on Example 1, with reference to Figure 2 As shown, this embodiment also provides a liver failure risk progression prediction system based on virtual reality. The system includes: a data acquisition and topology construction module, which acquires multimodal medical image data, performs standardized preprocessing, generates a fused image dataset, segments the target region to obtain a liver region mask, and constructs a vascular network topology based on the liver region mask.

[0166] The modeling and mapping module converts the obtained liver region mask into a three-dimensional surface mesh model, completes the setting action and texture mapping action for the liver material, and uses a mass-spring model to construct a biomechanical model of the liver to realize real-time deformation simulation of liver tissue. Clinical examination indicators are mapped onto the biomechanical model to realize information visualization.

[0167] The feature fusion and risk correction module extracts multimodal features of the target and uses a gated attention fusion mechanism to complete feature fusion; it constructs a multi-task deep learning prediction model and outputs preliminary prediction results simultaneously; based on the preliminary prediction results, it triggers a risk dynamic correction mechanism based on virtual reality interaction and outputs the final prediction results; the prediction results include: the predicted probability of liver failure, the progression time window, and the severity grade; the risk dynamic correction mechanism based on virtual reality interaction includes setting primary correction conditions, setting secondary correction conditions, and judging convergence of multiple corrections;

[0168] The virtual reality interaction module constructs an interactive form for the virtual reality risk process based on the final prediction results;

[0169] The tiered early warning support module completes risk tiering and early warning actions based on the final prediction results.

[0170] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0171] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0172] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes 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.

Claims

1. A method for predicting the progression of liver failure risk based on virtual reality, characterized in that, Includes the following steps: After acquiring multimodal medical image data and performing standardized preprocessing, a fused image dataset is generated. The target region is segmented to obtain a liver region mask, and a vascular network topology is constructed based on the liver region mask. The obtained liver region mask is converted into a three-dimensional surface mesh model. The liver material is set and texture mapping is performed. A mass-spring model is used to construct a biomechanical model of the liver to realize real-time deformation simulation of liver tissue. Clinical examination indicators are mapped onto the biomechanical model to achieve information visualization. Multimodal features of the target are extracted, and a gated attention fusion mechanism is used to complete feature fusion; Construct a multi-task deep learning prediction model and output preliminary prediction results simultaneously; Based on the preliminary prediction results, a risk dynamic correction mechanism based on virtual reality interaction is triggered, and the final prediction results are output. The prediction results include: the predicted probability of liver failure, the progression time window, and the severity classification. The risk dynamic correction mechanism based on virtual reality interaction includes setting the first correction condition, setting the second correction condition, and judging the convergence of multiple corrections. Based on the final prediction results, construct an interactive form for the virtual reality risk process; Risk stratification and early warning actions will be carried out based on the final forecast results.

2. The method for predicting the progression of liver failure risk based on virtual reality according to claim 1, characterized in that, Multimodal medical imaging data includes at least: CT images, MR images, and LS data; The standardized preprocessing involves image registration and fusion of multimodal medical image data, spatial registration, establishment of a unified three-dimensional coordinate system, and the use of a multimodal image registration algorithm based on mutual information (MI) to achieve spatial alignment by maximizing the mutual information value. After registration, the multimodal medical image data is fused at the pixel level to obtain a fused image dataset.

3. The method for predicting the progression of liver failure risk based on virtual reality according to claim 1, characterized in that, The process of obtaining the liver region mask is as follows: an improved 3D U-Net deep learning model is used to segment the target region of the fused image dataset; the improved 3D U-Net deep learning model means that an attention gating mechanism is introduced on the basis of the existing U-Net. After segmenting the target region, the improved 3D U-Net deep learning model is used for model training and inference to obtain the liver region mask. The basis for constructing the vascular network topology is as follows: based on the liver region mask, the intrahepatic vascular structure is extracted using a vascular enhancement filtering algorithm based on the Hessian matrix, the response function of the vascular network topology is defined, the vascular enhancement degree (Vesselness) is obtained, and the vascular system is extracted and constructed through threshold segmentation.

4. The method for predicting the progression of liver failure risk based on virtual reality according to claim 2, characterized in that, When converting to a 3D surface mesh model, the moving cube algorithm is used to extract the liver isosurface and generate triangular meshes; when setting the liver material, physically based rendering technology is applied to set the material; when performing texture mapping on the liver material: the liver surface texture is generated by extracting grayscale information from CT images to generate normal maps and roughness maps; clinical examination indicators include at least: liver function reserve parameters, LS data, and hemodynamic parameters.

5. The method for predicting the progression of liver failure risk based on virtual reality according to claim 2, characterized in that, The extracted target multimodal features include at least: image feature vectors, clinical examination indicator feature vectors, and genomic feature vectors; when constructing a multi-task deep learning prediction model, its backbone network adopts a Transformer encoder structure, which contains 6 encoder layers, each of which contains a multi-head self-attention mechanism and a feedforward neural network; Among them, the multi-head self-attention mechanism is to compute the attention of multiple subspaces in parallel; The feedforward neural network consists of two fully connected layers with the GELU activation function in between; The output layer of the multi-task deep learning prediction model contains three parallel prediction heads: a risk occurrence prediction head, which uses the Sigmoid activation function to output the probability of liver failure, with a value range of [0, 1]; a progression time prediction head, which uses a linear activation function to output the expected time to progress to liver failure, corresponding to the progression time window; and a severity grading prediction head, which uses the Softmax activation function to output the probability distribution of mild, moderate, and severe liver failure, corresponding to the severity grading.

6. The method for predicting the progression of liver failure risk based on virtual reality according to claim 1, characterized in that, The setting of the first correction condition means: setting the trigger condition for the first correction, i.e., condition A: when the output predicted probability of liver failure p̂ is greater than the preset first threshold θ1, the first correction mode is entered. The preset first threshold θ1 has a value range of 0 to 1; in the single correction mode, medical staff can choose whether to select the target area on the liver model through the virtual reality controller and adjust the functional parameters of the target area. The first correction formula is: p̂'=p̂+Δp; where p̂' represents the corrected risk probability, Δp represents the risk correction amount, and the formula is: Δp=a1×(w_region-w_baseline)×f_region; where a1 is the first correction coefficient, with a value range of [0.1, 0.3], w_region represents the weight coefficient of the target region, with a value range greater than 1, w_baseline represents the baseline weight coefficient, and f_region represents the functional abnormality score of the target region, with a value range of [0, 1].

7. The method for predicting the progression of liver failure risk based on virtual reality according to claim 6, characterized in that, The secondary correction condition setting means: set the trigger condition for secondary correction, namely condition B: when the change in risk probability after the first correction |Δp| ​​is greater than the preset second threshold θ2, and the corrected risk probability p̂' is still greater than the preset first threshold θ1, the directional index of the first correction to obtain the first risk correction amount satisfies: sign(Δp)=sign(ΔROI_score), where ΔROI_score represents the amount of adjustment change of the degree of functional abnormality of the region of interest by the medical staff; If three sub-conditions in condition B above are met, then the secondary correction mode is entered. The preset second threshold θ2 has a value range of 0~1. In the secondary correction mode, medical staff can choose whether to adjust the liver hemodynamic parameters, including at least: portal vein pressure (PVP) and hepatic artery blood flow (HAF). The secondary correction formula is: p̂''=p̂'+Δp'; where p̂'' represents the risk probability after secondary correction, and Δp' represents the secondary correction amount, based on the formula: Δp'=a2×(ΔPVP / PVP_baseline+ΔHAF / HAF_baseline); where a2 is the second correction coefficient, with a value greater than 0, equipped with a dynamic adjustment mechanism for the second correction coefficient, based on: a2=a2_0×(1+t0×|Δp| / θ2); where a2_0 is the baseline coefficient, with a value range of 0~1, t0 is the sensitivity coefficient, with a value range of [0.2, 0.5]; ΔPVP represents the change in portal vein pressure, PVP_baseline represents the baseline portal vein pressure, ΔHAF represents the change in hepatic artery blood flow, and HAF_baseline represents the baseline hepatic artery blood flow.

8. The method for predicting the progression of liver failure risk based on virtual reality according to claim 7, characterized in that, The convergence judgment of multiple corrections indicates that the correction convergence condition is set: when the risk correction amount of two consecutive times is less than the preset third threshold θ3, the correction process is considered to have converged, and the final prediction result is output; the preset third threshold θ3 is set to a value range of 0~1; after the second correction mode is completed, if further correction is still needed, the correction will continue, without needing to satisfy condition A or condition B again; the operation method is to randomly select the adjustment object: adjust the functional parameters of the target area again, corresponding to the first correction mode; adjust the liver hemodynamic parameters again, corresponding to the second correction mode; or adjust any one of the multiple parameters in a mixed manner. Among these parameters are functional parameters of the target region and hepatic hemodynamic parameters.

9. The method for predicting the progression of liver failure risk based on virtual reality according to claim 1, characterized in that, The construction of a virtual reality risk process interaction format includes at least: a timeline of the liver failure risk process, a dynamic evolution simulation of the liver state, and a virtual reality interactive interface setting. The timeline of the liver failure risk process is defined as follows: it starts from the target patient's current time and extends into the future to the predicted time of liver failure. The dynamic evolution simulation of the liver state simulates the dynamic evolution of the liver state from its current state to the state of liver failure, including at least changes in liver morphology and color. The virtual reality interactive interface setting includes a main view, a risk dashboard, a time controller, a parameter adjustment panel, and annotation tools. The risk stratification and early warning actions are based on the following: according to the probability of liver failure in the final prediction results, different patients are classified into risk levels and the corresponding decision suggestion texts in the decision template library are retrieved.

10. A virtual reality-based liver failure risk progression prediction system, applied to any one of the virtual reality-based liver failure risk progression prediction methods according to claims 1-9, characterized in that, The system includes: The data acquisition and topology construction module acquires multimodal medical image data, performs standardized preprocessing, generates a fused image dataset, segments the target region to obtain a liver region mask, and constructs a vascular network topology based on the liver region mask. The modeling and mapping module converts the obtained liver region mask into a three-dimensional surface mesh model, completes the setting action and texture mapping action for the liver material, and uses a mass-spring model to construct a biomechanical model of the liver to realize real-time deformation simulation of liver tissue. Clinical examination indicators are mapped onto the biomechanical model to realize information visualization. The feature fusion and risk correction module extracts multimodal features of the target and uses a gated attention fusion mechanism to complete feature fusion; it constructs a multi-task deep learning prediction model and outputs preliminary prediction results simultaneously; based on the preliminary prediction results, it triggers a risk dynamic correction mechanism based on virtual reality interaction and outputs the final prediction results; the prediction results include: the predicted probability of liver failure, the progression time window, and the severity grade; the risk dynamic correction mechanism based on virtual reality interaction includes setting primary correction conditions, setting secondary correction conditions, and judging convergence of multiple corrections; The virtual reality interaction module constructs an interactive form for the virtual reality risk process based on the final prediction results; The tiered early warning support module completes risk tiering and early warning actions based on the final prediction results.

Images