Artificial intelligence-based method for constructing a multi-stage screening model for hepatobiliary surgery solutions

Abstract

The application discloses a liver and gall surgery scheme multi-level screening model construction method based on artificial intelligence, relates to the technical field of artificial intelligence, and comprises the following steps: acquiring a liver and gall surgery scheme training set; wherein the training set comprises multi-modal clinical indexes of a target patient, candidate treatment schemes and multi-level classification labels; by drawing a first surgery topology graph, combining the multi-level classification labels, performing deep learning on the first surgery topology graph and at least two layers of reconstructed topology networks, and constructing a liver and gall surgery scheme initial screening model; in a second surgery topology graph and the liver and gall surgery scheme initial screening model, a main branch with parameter freezing and a control branch with trainable parameters are introduced; by constructing a loss function, the main branch parameters are kept frozen, only the control branch is subjected to iterative gradient updating, and a liver and gall surgery scheme multi-level screening model is generated; and the application improves the accuracy and robustness of the screening model.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for constructing a multi-level screening model for hepatobiliary surgery based on artificial intelligence. Background Technology

[0002] Due to the complex vascular system and varied anatomical structures involved in hepatobiliary surgery, the formulation and multi-level screening of auxiliary decision-making schemes have always been a challenge and pain point in the field of clinical medical artificial intelligence. In recent years, with the popularization of graph neural network deep learning technology in the medical field, researchers have begun to try to use digital graph technology to perform topological modeling of patients' imaging features, biochemical indicators and surgical paths, in order to achieve automated and precise preoperative scheme screening and intraoperative dynamic control.

[0003] However, traditional multi-level screening techniques for hepatobiliary surgery still have the following shortcomings: Firstly, traditional models are often trained using single-modality data, such as relying solely on anatomical images or physiological and biochemical indicators, failing to fully integrate multi-dimensional clinical information such as anatomical sections, physiological and biochemical timelines, and past medical history, resulting in insufficient accuracy in determining borderline safety cases. Secondly, existing models often use flat classification labels, such as only distinguishing between resectable and unresectable, lacking grading standards such as whether the treatment plan violates safety constraints, anatomical procedures, and prognostic risks. When analyzing subsequent multi-modal clinical indicators, simple post-feature splicing or linear mapping is usually used, which loses the inherent semantic relationship between anatomical images and physiological and biochemical timelines. When using topological networks for feature extraction, if the network layer is too shallow, it cannot fit high-order multi-level classification clinical labels, resulting in fuzzy safety constraints and inaccurate feature extraction. If the network layer is blindly deepened, it is easy to trigger topological representation degradation, overfitting, and gradient vanishing problems when processing sparse medical atlas data. In this case, the accuracy and robustness of the constructed screening model are extremely low. On the other hand, traditional models mostly adopt an end-to-end training and inference mode, and the model parameters can be updated throughout the process. Although it can adapt to complex cases, it is easy for individual non-standard surgical cases to cause safety baseline drift, which cannot meet the needs of real-time decision-making during surgery and is difficult to adapt to PACS and EMR systems of different levels of hospitals, thus limiting its clinical application. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a method for constructing a multi-level screening model for hepatobiliary surgery based on artificial intelligence. By integrating multimodal clinical indicators, candidate treatment plans, and multi-level classification labels, and utilizing a multi-layer reconstructed topology network for deep and shallow cascade learning, an initial screening model with high generalization ability is constructed. By deploying a parameter-frozen main branch and a parameter-trainable control branch in the second surgical topology map and the initial screening model, and configuring an aligned loss function for cross-branch joint iterative training, lossless element-level matrix fusion of the two is achieved. This fundamentally improves the prediction accuracy and system robustness of the multi-level screening model in complex, high-concurrency clinical scenarios, solving the problems mentioned in the background technology.

[0006] (II) Technical Solution

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] This application provides a method for constructing a multi-level screening model for hepatobiliary surgical procedures based on artificial intelligence, the method comprising:

[0009] Obtain a training set of hepatobiliary surgery protocols; the training set includes multimodal clinical indicators, candidate treatment protocols, and multilevel classification labels for the target patients;

[0010] Based on the training set of hepatobiliary surgery protocols, a first surgical topology map was drawn. Combined with multi-level classification labels, deep learning was performed on the first surgical topology map and at least two layers of reconstructed topology network. The deep learning included setting up a preset shallow learning network and a deep learning network for iterative training to build an initial screening model for hepatobiliary surgery protocols.

[0011] In the second surgical topology map and the initial screening model for hepatobiliary surgery plans, a main branch with frozen execution parameters and a control branch with trainable parameters are introduced. By configuring the loss function, the parameters of the control branch and the main branch are iteratively trained to generate a well-trained multi-level screening model for hepatobiliary surgery plans.

[0012] Furthermore, a topological map of the first surgery is drawn, including:

[0013] Extracting sequences of hepatobiliary anatomical structures and physiological changes based on multimodal clinical indicators;

[0014] Extraction of surgical liver tangent lines based on candidate treatment options;

[0015] A multimodal feature matrix is ​​constructed, PCA is used to determine the number of levels of multi-level classification labels, and clustering algorithms are applied to determine the screening centers and safety constraints of each level of scheme. By using the hepatobiliary anatomical structure as nodes, a virtual three-dimensional dissection is performed along the surgical liver tangent to obtain the surgical anatomical configuration. The edges of the nodes are connected according to the vascular flow relationship to draw the first surgical topology map.

[0016] Furthermore, clustering algorithms are applied to determine the screening centers and safety constraints for each level of the scheme, including:

[0017] Select the same number of data points as the number of levels as the initial centroids, iteratively calculate the Euclidean distance between the data points and the initial centroids until the initial centroids converge. Determine the position of the converged centroids as the screening center of each level of scheme, and determine the boundary thresholds within the clusters corresponding to the screening centers of each level of schemes to form safe boundary constraints.

[0018] The response times of the screening centers at each level are extracted and determined simultaneously. The time delay correction description function of each modal response time is established according to the modal category to form the time disturbance description quantity. Using the time disturbance description quantity, the screening centers at each level are sorted and grouped, and the time period threshold is determined to form the time response constraint.

[0019] Furthermore, the two-layer reconstructed topology network process includes:

[0020] Random initial vectors are used to identify the weight vectors of each node in the first surgical topology graph. The first surgical topology graph is reconstructed according to the adjacency relationship to obtain the second surgical topology graph, and it is marked as performing a layer of topology reconstruction network.

[0021] A shallow learning network is used to extract features from the topological map of the first surgery to obtain a first feature representation. The first feature representation includes a first subset that satisfies the safety constraints, a second subset that does not satisfy the safety constraints, and a third subset that has fuzzy safety constraints. Simultaneously, a deep learning network is used to extract a second feature representation from the topological map of the second surgery through a two-layer reconstruction network. The second feature representation includes a fourth subset that satisfies the safety constraints and a fifth subset that does not satisfy the safety constraints.

[0022] The first, third, and fourth subsets are analyzed for determination, and the initial screening model parameters are iteratively updated in conjunction with the multi-level classification labels. The second, third, and fifth subsets are analyzed for determination, and the initial screening model parameters are iteratively updated in conjunction with the multi-level classification labels.

[0023] Furthermore, the safety constraints must be met, including:

[0024] Determine whether the first feature representation satisfies the safety boundary constraint, and if the safety boundary constraint is satisfied, determine whether the time response constraint is satisfied; if the first feature representation satisfies the time response constraint, output the first screening result and generate at least one of the first subset and the fourth subset; if the first feature representation does not satisfy the time response constraint, output the second screening result and generate at least one of the second subset and the fifth subset, and sort and group the candidate response times of each modality and perform adjustment processing; if the first feature representation does not satisfy the safety boundary constraint, terminate the operation.

[0025] Furthermore, the initial screening model parameters are defined as follows:

[0026] The cascaded temporal structure of the initial screening model is determined based on the screening centers at each level, the number of levels, and the security constraints.

[0027] Based on the subsets of the first and second feature representations and the generated candidate treatment fragments, an evaluation strategy is executed to determine the distribution function; the least squares method is used to process the cascaded temporal structure and distribution function to obtain the initial screening model parameters.

[0028] Furthermore, implement the assessment strategy, including:

[0029] Based on the time-series deviation of each subset of the first and second feature representations, the candidate treatment plans are sorted according to the magnitude of the deviation to determine their priority at the response time. Using the priority, feature screening is performed on each subset of the first and second feature representations, and the selected subset features are aggregated to generate a fusion representation quantity. The mean and variance of the fusion representation quantity in the hepatobiliary surgery plan training set are statistically analyzed, and the sum of the mean and the variance of multiples of k is set as the distribution function.

[0030] Furthermore, the control branch training process includes:

[0031] The control branch has M graph feature processing layers;

[0032] A composite feature representation is generated based on the first feature representation and the second feature representation, and a binary mask sequence corresponding to the composite feature representation is constructed; wherein, an all-zero mask is assigned to the first feature representation, and an all-one mask is assigned to the second feature representation.

[0033] The composite feature representation and the binary mask sequence are mapped to the first feature space. Feature concatenation is performed in the channel dimension. Using semantic analysis and multi-level classification labels, a topological conditional unit is assembled and input into the parameter-trainable control branch. Feature transformation is performed on the topological conditional unit to extract the first topological control matrix. The corresponding weights of the first topological control matrix are adjusted using the encoder built into the control branch.

[0034] Furthermore, the main branch training process includes:

[0035] The main branch consists of N cascaded graph feature processing layers, where N≥4 and M≤N. At the same time, the i-th graph feature processing layer in the control branch is set up, which corresponds one-to-one with the i-th graph feature processing layer in the main branch in terms of network position and feature dimension, where i is a positive integer greater than 0 and less than or equal to M.

[0036] By obtaining the first topology control matrix after adjusting the weights, the first topology control matrix after adjusting the weights is added element by element to the intermediate hidden state output by the main branch with the parameters frozen, and then injected into the graph feature processing layer of the main branch to output the second topology control matrix.

[0037] Furthermore, configure the loss function, including:

[0038] A pre-trained deep learning network is used to encode graph features of the second surgical topology map and extract target clinical semantic features; the second topology control matrix is ​​obtained; the second topology control matrix is ​​subjected to dimensional transformation and spatial mapping and projected into the feature space corresponding to the target clinical semantic features to obtain the corresponding generated features;

[0039] Extract each discrete feature map from the generated features and calculate the mean negative cosine similarity between it and the corresponding surgical anatomical configuration in the target clinical semantic features. Use this as the alignment loss function to jointly update the main branch and the control branch. By minimizing the overall training objective, gradient backpropagation is used to keep the parameters of the main branch frozen. Only the trainable parameters of the control branch are iteratively optimized to output the trained multi-level screening model.

[0040] (III) Beneficial Effects

[0041] This invention provides a method for constructing a multi-level screening model for hepatobiliary surgical procedures based on artificial intelligence, which has the following beneficial effects:

[0042] 1. This invention constructs a training set for hepatobiliary surgery plans by building multimodal clinical indicators, candidate treatment plans, and multi-level classification labels for target patients. This can comprehensively and three-dimensionally capture changes in the patient's physiological state and the surgical liver tangent. Based on this training set, a first surgical topology map is drawn. At least two layers of reconstructed topology networks are introduced for deep learning. By setting up a pre-set shallow learning network and a deep learning network for collaborative iterative training, it can not only accurately fit the geometric topology alignment relationship of the surgical topology map, but also deeply abstract the high-level surgical logic features. This effectively avoids feature degradation and overfitting during the process of deepening the network depth, and improves the generalization ability of the initial screening model for hepatobiliary surgery plans.

[0043] 2. This invention constructs a graph control network architecture by introducing a main branch with frozen execution parameters and a control branch with trainable parameters into the second surgical topology map and the initial screening model. By freezing the weight parameters of the main branch, the target clinical semantics in the historical training set can be strongly preserved. At the same time, combined with the iterative training of the parameters of the control branch, the cross-branch joint iterative training of the parameters of the control branch and the main branch is performed at the matrix level through one-to-one feature alignment and element addition, and a targeted loss function is configured. At the same time, time response constraints are superimposed to generate a well-trained multi-level screening model for hepatobiliary surgery plans. By combining multi-level classification labels, the high security of screening complex hepatobiliary surgery plans is improved, the computational efficiency of automatically optimizing surgical topology plan screening is increased, and the system's decision latency is reduced. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0045] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0046] Research concept: By integrating multimodal clinical indicators, candidate treatment options, and multi-level classification labels, and utilizing multi-layer reconstructed topology networks for deep and shallow cascade learning, an initial screening model with high generalization ability is constructed. By deploying parameter-frozen main branches and parameter-trainable control branches in the second surgical topology map and the initial screening model, and configuring an alignment loss function for cross-branch joint iterative training, lossless element-level matrix fusion of the two is achieved, thereby fundamentally improving the prediction accuracy and system robustness of the multi-level screening model in complex, high-concurrency clinical scenarios.

[0047] Example: This embodiment of the invention provides a method for constructing a multi-level screening model for hepatobiliary surgical procedures based on artificial intelligence; Figure 1 This is a flowchart illustrating the present invention; please refer to [link / reference]. Figure 1 The method includes the following steps:

[0048] S1: Obtain the training set for hepatobiliary surgery protocols; the training set includes multimodal clinical indicators, candidate treatment protocols, and multilevel classification labels for the target patients;

[0049] In this embodiment, a standard medical data interface is used to connect to the hospital's image archiving and PACS system, hospital surgical anesthesia system, and electronic medical record system. By setting a unique anonymized patient identification for the target patient, multimodal clinical indicators are statistically obtained across systems. This includes: batch importing the original DICOM sequence images of corresponding enhanced CT or MRI from the PACS system as anatomical image slices; extracting the text of continuous biochemical test reports from the target patient within 72 hours before surgery from the electronic medical record system as physiological change data; performing standardization and alignment on the generated data to generate multimodal clinical indicators; simultaneously, for the anatomical image slices, three-dimensional linear interpolation is used to resample anatomical image slices of different thicknesses into standard 1mm square voxels, and window width and window level clipping based on pixel grayscale is performed to extract the characteristics of the lesion entity and the relationship with vascular flow; for the physiological change data, a time series matrix is ​​constructed with a fixed step size of 12 hours, and the missing test report texts are filled in using a forward filling algorithm to generate physiological and biochemical sequences of uniform length.

[0050] Using named entity recognition technology in natural language processing, the historical surgical record text in the hospital's surgical anesthesia system and the medical record homepage is retrieved and parsed; based on the preset hepatobiliary surgery procedure dictionary, the structured surgical procedure name, resection range, vascular occlusion method and vascular reconstruction strategy that match the above multimodal clinical indicators are automatically extracted, and these structured surgical parameters are encapsulated into corresponding candidate treatment plans.

[0051] By retrieving the outpatient follow-up records, discharge summaries, and re-examination laboratory reports of target patients 30 days to 1 year after surgery, the actual clinical prognostic outcomes are obtained, such as: whether postoperative liver failure occurred, complication grade, and whether there is residual tumor. A multi-task rule mapping engine for prognosis is configured in the system to automatically derive multi-level classification labels based on the actual outcomes: Assuming the multi-level classification labels are 3 levels, the first level label indicates whether the treatment plan violates safety constraints, and its classification status corresponds to meeting safety constraints, fuzzy safety constraints, or not meeting safety constraints; the second level label indicates the anatomical procedure, used to accurately define the topological range of liver resection, and its classification status corresponds to irregular local excision, regular liver segment resection, standard hemihepatectomy, or extended resection combined with vascular reconstruction; the third level label indicates the prognostic risk layer, used to predict the outcome of postoperative follow-up complications, and its classification status corresponds to good residual liver compensation, transient dysfunction, or irreversible postoperative liver failure.

[0052] In addition, the system automatically backtracks and discretizes the data through a pre-defined rule mapping engine, then concatenates and combines them into a three-dimensional tensor, such as a [0, 2, 0] vector representing a safe treatment plan, right hemihepatectomy, and low risk. This tensor serves as a high-dimensional target clinical cue word in the subsequent screening model training, facilitating direct input into the trainable control branch for semantic analysis and assembly into topological conditional units. Finally, the spatiotemporal topological control features output by the control branch are injected and forcibly constrained in the feature evolution of the main branch with frozen parameters through unidirectional residual addition. This determines the direction of the main branch's winning surgical atlas generation at the underlying logic, ensuring that the plan is fully aligned with the expected safety constraints. Finally, the automatically generated discrete codes at all levels are combined into a tree-like hierarchical vector, which is then packaged and bound with multimodal clinical indicators and candidate treatment plans to form the final hepatobiliary surgery plan training set.

[0053] The following is an explanation of some of the terms involved:

[0054] Physiological and biochemical sequences: dynamic time-series test indicators used to characterize the liver reserve function and systemic tolerance status of target patients, including total bilirubin, alanine aminotransferase, albumin, prothrombin time, and indocyanine green 15-minute retention rate, etc.

[0055] Anatomical image slices: used to characterize tomographic spatial images with continuous slice thickness obtained by computed tomography or magnetic resonance imaging, in order to clarify the establishment of surgical topology;

[0056] Pathological characteristics: Multidimensional biological characterization of disease entities such as hepatobiliary malignancies requires clear pathological classification, such as primary hepatocellular carcinoma, cholangiocarcinoma, etc., staging, such as stage 0, stage A, stage B, stage C and stage D in the Barcelona BCLC staging, as well as microvascular invasion status.

[0057] Vascular flow relationship: used to characterize the anatomical morphology and spatial trend of the portal vein, hepatic vein and hepatic artery in the first and second hepatic hilum, which facilitates the subsequent establishment of surgical topology map;

[0058] The aforementioned data can be further enhanced by combining web scraping technology to obtain more relevant real-world hepatobiliary surgery case data from publicly available medical literature networks or international multicenter clinical databases. After screening, outlier removal, and cleaning, the data dimensions of high-dimensional features are enriched, further improving the breadth and depth of data collection. Simultaneously, by combining the attending physician's surgical actions for various complex procedures, the accuracy of manually delineating the anatomical boundaries of lesions can be judged. Through this combination, the statistically derived multi-level classification labels will be more accurate, thus providing a more professional and standardized high-quality hepatobiliary surgery training set for model training. This system improves the resource utilization rate of clinical sample data and the efficiency of automatic training and iteration of the hepatobiliary surgery treatment plan screening model by optimizing the model's acquisition methods and the accuracy of its requirements for multimodal clinical indicators and candidate treatment plan elements.

[0059] S2: Based on the training set of hepatobiliary surgery plans, draw the first surgical topology map, combine multi-level classification labels, and perform deep learning on the first surgical topology map and at least two layers of reconstructed topology network. The deep learning includes setting a preset shallow learning network and deep learning network for iterative training to build an initial screening model for hepatobiliary surgery plans.

[0060] The first surgical topology map was constructed, including: extracting hepatobiliary anatomical structures and physiological change sequences based on multimodal clinical indicators; analyzing candidate treatment plans, determining the NER and relation extraction of candidate treatment plans, extracting all structured surgical procedure names, resection ranges, vascular occlusion methods, and vascular reconstruction strategies, and mapping these texts into constraint vectors in three-dimensional space; establishing a relative coordinate system of the surgical path with respect to the hepatobiliary anatomical structure by using key anatomical landmarks, such as the root of the hepatic vein, as anchor points, and forming the surgical liver tangent by collecting the trajectory of the surgical path in the three-dimensional space; combining the hepatobiliary anatomical structure and physiological change sequences, and aligning them with timestamps, cascading and superimposing the hepatobiliary anatomical structure, physiological change sequences, and surgical liver tangent in the channel dimension to form a multimodal feature matrix;

[0061] PCA is used to determine the number of levels of multi-level classification labels, and clustering algorithms are applied to determine the screening centers and safety constraints for each level of scheme. The specific steps include: standardizing the multimodal feature matrix to obtain the covariance matrix, which measures the linear correlation between different features. A value greater than 0 indicates a positive correlation (e.g., larger remaining liver volume corresponds to a lower risk of postoperative liver failure); a value less than 0 indicates a negative correlation (e.g., larger tumor diameter corresponds to a higher risk of insufficient resection margins); and a value equal to 0 indicates no linear correlation. The covariance matrix is ​​then eigenvalued to extract several principal component eigenvectors. The larger the corresponding eigenvalue, the more fluctuations in the original data that the principal component can explain, and the richer the effective decision-making information it contains. For example, after eigenvalue decomposition, 27 principal components are obtained. The first principal component corresponds to the first-level label and has the largest eigenvalue, such as core indicators like remaining liver volume, resection margin distance, and Child-Pugh grade. The second principal component corresponds to the second-level label, such as tumor location, vascular variation, and liver segment dominance. The third principal component corresponds to the third-level label, such as ICG-R15, degree of cirrhosis, and albumin level. The number of levels is less than or equal to 3.

[0062] Based on the proportion of each feature value to the total sum of feature values, the individual variance contribution rate of the principal components at each level is calculated sequentially. The contribution rates are accumulated in descending order. Accumulation stops when the cumulative contribution rate of the principal components first exceeds or equals a preset threshold, indicating that too many redundant features have been introduced. In this embodiment, the preset threshold is set to 85%, at which point accumulation stops, and the absolute number of principal components participating in the accumulation at this point is used as the initial label level. In addition, a single-level contribution rate threshold of 10% is set to filter out principal components with an individual variance contribution rate lower than 10%, thus determining the final number of multi-level classification labels.

[0063] For example: Suppose that the individual variance contribution rates of the first three principal components are 41.5%, 25.2%, and 11.5% respectively, with a cumulative contribution rate of 78.2%. Since 78.2% < 85%, the preset quantity threshold has not been reached, so the accumulation continues. Now, four principal components are added, and the cumulative rate is 86% > 85%. However, the fourth principal component is 7.8% < 10%, so it is still determined as a three-level classification label: the first principal component corresponds to the first-level label, the second principal component corresponds to the second-level label, and the third principal component corresponds to the third-level label, adaptively determining the number of levels for multi-level classification labels. If the preset quantity threshold is reached only when the fourth principal component is added, and the fourth principal component is greater than 10%, since the threshold for the number of levels is set, the semantics of the fourth principal component are merged into the third-level label.

[0064] After determining the number of levels, based on the multimodal feature matrix, data points with the same number of levels are selected as the initial centroids of the clustering algorithm. The Euclidean distance between the data points and the initial centroids is calculated iteratively. The clustering algorithm is used to repeatedly divide the data points alternately until the initial centroids meet the preset maximum convergence threshold. Several centroid positions that finally converge are determined and used as the screening centers for each level of the scheme. At the same time, the boundary thresholds within the clusters corresponding to the screening centers of each level of the scheme are determined as safety boundary constraints. For example, taking the screening centers of each level of the scheme as the center, the geometric outer envelope boundary threshold of the corresponding clusters at a 95% confidence level is extracted and used as the safety boundary constraint for the hepatobiliary surgery scheme.

[0065] During convergence, a high-precision timer is attached to each screening center to record the relative timestamp from the input to the output of that screening center, and this timestamp is determined as the response time of each screening center. Since the update frequency and transmission link of different modal data in PACS systems, electronic medical record systems, etc. are different, the system establishes a time delay correction description function for each original timestamp according to modal categories, such as: anatomical hepatobiliary imaging modality, physiological and biochemical time series modality, and surgical liver tangential modality. For each screening center, the output values ​​of the time delay correction description function of the modal categories involved are summed and the calculation constant of this level is added to obtain the theoretical baseline time delay.

[0066] For the j-th level scheme selection center Cj, determine the set of modes Pj it calls. Assume C1 calls the biochemical time series mode, C2 calls the hepatobiliary imaging mode, and C3 calls the surgical liver tangential mode. Look up the table to obtain the output value of the delay correction description function F(x) corresponding to each mode in Pj, sum the F(x) to obtain the theoretical reference delay of the center on the mode transmission and decoding side. At the same time, add the constant of this level of calculation on this basis. This constant is pre-calibrated by Cj and its value range is usually greater than 0. In addition, the summation aggregation of each mode is a basic calibration method. In the mode parallel decoding deployment environment, the implementer can replace it with max aggregation, which will not be elaborated here.

[0067] For the anatomical hepatobiliary imaging modality, a first time delay correction description function is constructed with the number of image voxels as the variable: F1(x)=a1×(x1 / 10) 6 In the formula, x1 represents the number of image voxels, which is an integer variable greater than 0; a1 represents the transmission and decoding time per million voxels, which is greater than 0 and reflects the link bandwidth and decoder performance when the PACS system pushes anatomical liver and gallbladder image slices; 10 6For constant-level throughput normalization operators, F1(x) is used to smoothly map the number of image voxels to the million-level industrial scale. F1(x) represents the reference latency required for the image modality to enter the screening center alone. It should be noted that during image loading, the amount of data is positively correlated with the transmission time. The larger the number of image voxels, the larger the amount of data and the larger the output value of the first latency correction description function.

[0068] For the physiological and biochemical time series modality, a second time delay correction description function is constructed with the number of biochemical indicators as the variable: F2(x)=a2×x2+γ×(T-1); where x2 represents the number of test indicators to be processed in a single physiological and biochemical time series modality, which is an integer greater than 0; a2 represents the time consumption of structured parsing and vectorized embedding of a single biochemical indicator, which is greater than 0; T represents the length of the physiological and biochemical time series obtained by the target patient in the data collection interval, T=1 indicates only a single preoperative blood sampling; γ represents the time consumption of time series alignment between two adjacent samplings, which is activated when T>1 and used for time axis interpolation alignment, the specific steps of which are not described in detail here; F2(x) characterizes the baseline time delay of the biochemical time series modality from the electronic medical record interface to the screening center; it should be noted that the more biochemical indicators, the more frequent the sampling, the more complex the data preprocessing, the higher the delay in entering the screening center, and the larger the output value of the second time delay correction description function.

[0069] For the surgical liver tangent modality, a third time-delay correction description function is constructed with the number of nodes in the surgical topology map as the variable: F3(x) = a3 × x3; where x3 represents the number of nodes in the surgical topology map, a3 represents the feature initialization time of a single node (its value is greater than 0), and F3(x) represents the baseline time of the surgical topology map being retrieved from local storage / intraoperative reconstruction and called by the screening center. It should be noted that the more nodes the surgical topology map has, the longer the loading and initialization time, the greater the overhead of memory reading and feature assignment, and the larger the output value of the third time-delay correction description function. For example, a map with 20 nodes takes longer to load than a map with 8 nodes.

[0070] The difference between the response time and the theoretical baseline delay is calculated and defined as the time disturbance description of a certain level of scheme selection center. The time disturbance descriptions of all selection centers are aggregated into a time disturbance description vector. The aggregation method is vector concatenation. The response time of a certain level of scheme selection center is subtracted from the product of the preset smoothing factor and the time disturbance description. The difference is the calibration response time. The preset smoothing factor is greater than 0. It should be noted that this calibration response time eliminates external disturbances such as PACS system transmission and interface waiting, and only retains the computation time characteristics of the model itself.

[0071] The candidate response times for each modality are sorted and grouped according to their temporal proximity in the calibration response time. Specifically, the calibration response times are arranged in ascending order, and the time difference between adjacent screening centers is calculated. A time period threshold is preset. If the adjacent difference is less than or equal to the time period threshold, the two are considered to have a temporal proximity relationship and are assigned to the same response window, allowing parallel decision-making within the group. If the difference is greater than the time period threshold, they belong to different response windows, and the groups must follow a strict temporal sequence relationship, forming a time response constraint. In subsequent analysis, based on the above sorting and grouping results, within the same response window, the control branch can be fine-tuned based on the output of the main branch. If the response time lag of the control branch is greater than the time period threshold, it is determined that its input data has deviated from the current response window, triggering the control branch failure mechanism and forcibly reverting to the output result of the main branch. In this way, the decision-making safety and timeliness of the multi-level screening model in the complex data flow environment of hepatobiliary surgery are effectively guaranteed.

[0072] The preset time period threshold is obtained based on statistical analysis. By collecting the calibration response time differences of screening centers at all levels within the same sample of no less than 200 target patients in history, the mean and standard deviation are calculated. Let the time period threshold = mean + κ × standard deviation, where the value of κ ranges from [0.5, 2.0], and can be taken as 1.0 to 1.5 in practice. It should be noted that this method is suitable for long-term iteration after model deployment, so that the time period threshold can adapt to the PACS link and equipment performance of specific hospitals.

[0073] By using hepatobiliary anatomical structures as nodes, including but not limited to liver segment regions, portal vein trunk and branches, hepatic artery trunk and branches, hepatic vein trunk and branches, biliary system, and tumor lesions, a virtual three-dimensional dissection is performed along the surgical liver tangent to obtain the surgical anatomical configuration. The adjacency relationship of each node in three-dimensional space is determined, and the connection edges between nodes are established according to the blood flow relationship. The direction of the edges follows the blood flow direction or bile flow direction, and the weight of the edges can be set according to the blood vessel diameter or blood flow rate. The first surgical topology map is drawn, providing structured input features for subsequent deep learning networks.

[0074] The two-layer reconstruction of the topology network includes: generating a random initial vector by uniformly selecting values ​​in the interval [0, 1]; allocating and identifying the weight vectors of each node in the first surgical topology graph using the random initial vectors; extracting the weight vector of the current node and the weight vectors of its associated nodes based on adjacency relationships; integrating the weight vector of the current node and the weight vectors of its associated nodes by performing aggregation and nonlinear transformation to obtain the reconstruction weight vector of the current node; reconstructing the first surgical topology graph based on the reconstruction weight vector to obtain the second surgical topology graph, and marking it as performing one layer of topology network reconstruction; simultaneously, the reconstruction steps involve at least two layers, resulting in at least two layers of reconstructed topology networks;

[0075] A shallow learning network is used to extract features from the topological map of the first surgery to obtain a first feature representation. This first feature representation includes a first subset that satisfies safety constraints, a second subset that does not satisfy safety constraints, and a third subset with fuzzy safety constraints. The first subset that satisfies safety constraints involves: determining whether the first feature representation satisfies safety boundary constraints; if so, determining whether it satisfies time response constraints; if so, outputting a first filtering result and generating at least one of the first and fourth subsets; if not, outputting a second filtering result and generating at least one of the second and fifth subsets, and sorting and grouping the candidate response times of each modality and performing adjustment processing; and terminating the operation if the first feature representation does not satisfy safety boundary constraints.

[0076] It should be noted that the input to the shallow learning network is the first surgical topology map and the hepatobiliary surgery protocol training set. Its specific structure includes a first graph convolutional layer, a second graph convolutional layer, a first fully connected mapping layer, and an output layer. The first graph convolutional layer initially has 64 hidden units and uses ReLU activation, used to perform coarse aggregation of neighborhood node features within a 1-hop range on the input first surgical topology map. The second graph convolutional layer initially has 32 hidden units and uses ReLU activation, used to perform local spatial semantic fusion within a 2-hop range on the initial graph features. The first fully connected mapping layer... The number of hidden units is 16, which is used to map the topological features of the fused local graph into a one-dimensional space. Finally, it passes through the output layer, which has a built-in Sigmoid activation function. The high confidence threshold is initially configured to be 0.85 and the low confidence threshold is initially configured to be 0.15. Features with calculated probability values ​​greater than or equal to 0.85 are divided into the first subset that meets the safety constraints. Features with calculated probability values ​​less than or equal to 0.15 are divided into the second subset that does not meet the safety constraints. Complex overlapping features with probability values ​​between 0.15 and 0.85 are divided into the third subset that meets the fuzzy safety constraints.

[0077] In this embodiment, satisfying the safety constraints means that the current data meets the safety boundary constraints and the time response constraints, corresponding to the sample sets in the first and fourth subsets; not meeting the safety constraints means that the current data does not meet at least one of the safety boundary constraints or the time response constraints, corresponding to the sample sets in the second and fifth subsets; fuzzy safety constraints mean that it is impossible to clearly determine whether the current data meets the safety boundary constraints or the time response constraints, possibly due to incomplete feature extraction caused by the latency of multimodal data transmission, corresponding to the sample set in the third subset, which is the key object of model training and optimization.

[0078] By reconstructing the topology network in two layers, a second feature representation of the second surgical topology map is extracted using a deep learning network; wherein, the second feature representation includes a fourth subset that satisfies the safety constraints and a fifth subset that does not satisfy the safety constraints;

[0079] It should be noted that the deep learning network consists of a graph attention layer, residual blocks, normalization layers, a second fully connected layer, and an output layer. The graph attention layer cascades four graph attention network layers vertically, with eight parallel attention heads per layer. Each attention head initially has 16 hidden units, resulting in a total feature dimension of 128 after multi-head parallel concatenation. The activation function is LeakyReLU, used to extract associated features from the input second surgical topology graph. The residual blocks perform element-wise addition of the input and output tensors of the graph attention layer and suppress feature flattening caused by the deep network through layer normalization. The slippage defect is addressed; the second fully connected layer has 64 hidden units and uses ReLU activation function to perform nonlinear fusion and multidimensional dimensionality reduction compression on the high-order global graph features after eliminating oversmoothing; finally, the output layer, which has a built-in Sigmoid activation function and an initial confidence threshold of 0.50, normalizes the results to the 0 to 1 interval and performs absolute binary classification. If the calculated probability value is greater than or equal to 0.50, the fourth subset that meets the safety constraints is directly output; if the calculated probability value is less than 0.50, the fifth subset that does not meet the safety constraints is output, thus achieving fine division of the data corresponding to the previous fuzzy safety constraints.

[0080] The analysis of the first, third, and fourth subsets, combined with multi-level classification labels, involves parameter updates: First, a set difference operation is performed on the datasets of the first and fourth subsets: data belonging to the third subset is removed from the first subset, resulting in the sixth subset. Samples in this subset have clear characteristics and meet the safety constraints. Second, a set difference operation is performed: data belonging to the first subset is removed from the third subset, resulting in the seventh subset. Although samples in this subset may have fuzzy safety constraints, preliminary screening suggests they do not meet the safety constraints. Subsequently, the sixth and seventh subsets, along with the multi-level classification labels, are used as input information to train the initial screening model.

[0081] The second, third, and fifth subsets are analyzed, and parameters are updated in conjunction with multi-level classification labels: First, a set difference operation is performed: data not present in the second subset is obtained from the third subset, resulting in the eighth subset. After analysis, these samples are determined to meet the safety constraints. Second, a set difference operation is performed: data not present in the third subset is obtained from the second subset, resulting in the ninth data set. After analysis, these samples are determined to not meet the safety constraints. Subsequently, the eighth and ninth subsets, along with the multi-level classification labels, are used as input information to train the initial screening model.

[0082] The initial screening model parameters are defined as follows: The cascaded temporal structure of the initial screening model is determined based on the screening centers at each level, the number of levels, and safety constraints; An evaluation strategy is executed based on each subset of the first and second feature representations and the generated candidate treatment fragments to determine the distribution function. Specific steps include: The temporal deviation of each subset of the first and second feature representations is considered; the temporal deviation represents the temporal stability of the decision result corresponding to that subset; the smaller the deviation value, the closer the response time of that subset is to the theoretical baseline delay, and the higher the reliability of its output; Candidate treatment protocols are sorted according to the magnitude of the deviation to determine their priority at the response time, i.e., the subset with the smallest temporal deviation has the highest priority, and the subset with the largest temporal deviation has the lowest priority.

[0083] Using priority order, feature filtering is performed on each subset of the first feature representation and the second feature representation. High-priority subsets retain all or most feature dimensions, while low-priority subsets are subjected to dimensionality reduction or only feature dimensions that are complementary to high-priority subsets are retained. In subsequent channel dimensions, the inverse of the priority order is used as a weighting factor to perform spatial weighted aggregation representation on the effective subset features retained after filtering. The mapping and compression generate a fusion representation quantity to represent the spatiotemporal representation of the current candidate treatment plan.

[0084] For example: Suppose the five time-series deviations calculated above are sorted in ascending order as follows: third subset < first subset < fourth subset < second subset < fifth subset. This sorting relationship determines the priority order of each subset at the current response time. Based on this priority order, a feature filter is configured. The filter is set to determine that if the time-series deviation of a subset is greater than a preset time-series threshold, the feature is deemed invalid. In this example, if the fifth subset is greater than the preset time-series threshold, it is directly truncated and removed by the filter, and its weight factor is set to 0, thus achieving feature filtering. For the first, third, fourth, and second subset features retained after filtering, the system uses a linear mapping layer to align their dimensions and performs weighted concatenation using the reciprocal of the priority order as the weight factor, fusing them into a unified high-dimensional one-dimensional vector, i.e., the fused representation.

[0085] The mean and variance of the fusion representation in the training set of hepatobiliary surgery protocols are statistically analyzed. The sum of the mean and a multiple of the variance (k times) is set as the distribution function. Here, k is a preset scaling factor, and the value of k ranges from [0.5, 3.0], typically between 1.0 and 2.0. It should be noted that the distribution function, by using the statistical mean of the training set and adding a multiple of the variance (k times), is used to measure the degree of matching between the fusion representation of the new candidate treatment protocol and the ideal distribution. The least squares method is used to jointly process the cascaded temporal structure and the distribution function. The training set data is input into the cascaded temporal structure to obtain the predicted output value of the model. The residual between the predicted output value and the ideal score given by the distribution function is calculated. The network weight parameters in the cascaded temporal structure are iteratively adjusted using the backpropagation algorithm to minimize the sum of squared residuals. When the loss function converges to a preset first convergence threshold, the obtained network weight parameters are the initial screening model parameters. Here, the first convergence threshold is typically set to 0.1.

[0086] By constructing a training set for hepatobiliary surgery protocols using multimodal clinical indicators, candidate treatment plans, and multi-level classification labels for target patients, a comprehensive and three-dimensional understanding of changes in patients' physiological states can be captured. Based on this training set, a first surgical topology map is drawn, and at least two layers of reconstructed topology networks are introduced for deep learning. By setting up pre-defined shallow and deep learning networks for collaborative iterative training, not only can the geometric topological alignment relationship of the surgical topology map be accurately fitted, but also the high-level surgical logical features can be deeply abstracted. This effectively avoids feature degradation and overfitting during the process of deepening the network, and improves the generalization ability of the initial screening model for hepatobiliary surgery protocols.

[0087] S3: In the second surgical topology map and the initial screening model of hepatobiliary surgery plan, a main branch with frozen execution parameters and a control branch with trainable parameters are introduced. By configuring the loss function, the parameters of the control branch and the main branch are iteratively trained to generate a trained multi-level screening model of hepatobiliary surgery plan.

[0088] The control branch training process includes: a graph feature processing layer with M main branches in the control branch configuration;

[0089] In the graph feature processing layer, the first feature representation and the second feature representation of the aforementioned output are obtained. The first feature representation and the second feature representation are spatiotemporally concatenated to generate a composite feature representation, ensuring the compatibility of subsequent feature fusion, and constructing a binary mask sequence corresponding to the composite feature representation. Among them, the first feature representation is assigned an all-zero mask, which indicates that its confidence is high and no significant adjustment of the control branch is required. The second feature representation is assigned an all-one mask, which indicates that it is an accurate judgment from the deep network and has high confidence, and is used to guide the correction of the data corresponding to the fuzzy safety constraints.

[0090] The composite feature representation and the binarized mask sequence are concatenated and superimposed along the channel dimension, then input into the first fully connected mapping layer and mapped to the first feature space to obtain the features of the fused mask information. Using semantic analysis and multi-level classification labels, the features of the fused mask information are semantically weighted, assembled into topological conditional units, and input into a parameter-trainable control branch. Inside the control branch, the topological conditional unit performs layer-by-layer feature transformation through M cascaded graph feature processing layers, each consisting of a graph convolution operator with 128 hidden units, extracting the first topological control matrix with a channel dimension of 128. The encoder built into the control branch is then used for tuning... The first topology control matrix is ​​assigned corresponding weights to prevent the introduction of harmful control noise in the early stages of training: nodes with a mask of 1 are processed first, and special attention heads are assigned to adjust weights for different hepatobiliary anatomical structures, and finally the first topology control matrix is ​​extracted; this matrix has the same dimension as the intermediate hidden state of the corresponding layer of the main branch, and the element values ​​range from [0, 2], which is used for subsequent gain or suppression adjustment of the main branch features. When the element value is greater than 1, the feature of the corresponding node of the main branch is enhanced; when the element value is less than 1, the feature of the corresponding node of the main branch is suppressed; when the element value is equal to 1, the feature of the corresponding node of the main branch remains unchanged, and the adjustment process conforms to the time response constraint. If the control branch output times out, it will directly fail.

[0091] A layer-by-layer feature transformation is performed on the topological conditional unit to extract the first topological control matrix. The main branch is derived from the pre-reconstruction modules of the shallow and deep learning networks of the initial screening model. All parameters are frozen throughout the process and do not participate in gradient updates. During the training of the main branch, the following steps are taken: the main branch consists of N cascaded graph feature processing layers, where N≥4 and M≤N; at the same time, the i-th graph feature processing layer in the control branch is set to correspond one-to-one with the i-th graph feature processing layer in the main branch in terms of network position and feature dimension; where i is a positive integer greater than 0 and less than or equal to M.

[0092] During the forward propagation training of the main branch, all parameters of the main branch are locked, i.e., they do not receive gradient backpropagation updates. Based on the topology control matrix output by the i-th graph feature processing layer in the control branch, and judging whether the topology control matrix output by the i-th layer of the control branch meets the time response constraint, if the difference between it and the output response time of the i-th layer of the main branch is greater than the preset time threshold, the topology control matrix is ​​set to a matrix of all 1s; otherwise, the original topology control matrix is ​​retained.

[0093] By obtaining the first topology control matrix after adjusting the weights, and since this output topology control matrix has the same feature dimension as the intermediate hidden state output by the i-th graph feature processing layer in the main branch, the first topology control matrix after adjusting the weights is added element-wise to the intermediate hidden state output by the main branch with the parameters frozen, and then injected into the i-th graph feature processing layer of the main branch to output the second topology control matrix: Second topology control matrix = Intermediate hidden state of the main branch + μ⊙First topology control matrix after adjusting the weights; where μ is the fusion coefficient, and its value ranges from 0 to 1, and ⊙ means element-wise multiplication;

[0094] It should be noted that ⊙ is calculated based on certain activation functions, such as sigmoid or tanh, with values ​​between 0 and 1. The injection step does not change the internal parameters of the main branch graph feature processing layer; it only replaces the output of that layer. All parameters of the main branch remain frozen throughout the inference process. The first topology control matrix, after adjusting the weights, is multiplied element-wise and then added element-wise to the hidden states in the middle of the main branch. This matrix is ​​then injected into the i-th layer of the main branch, and the second topology control matrix is ​​output. In this approach, the criteria of the shallow learning network mainly focus on the third subset of fuzzy safety constraints, while the fine-grained criteria of the deep learning network raise the upper limit. A fusion function is set to prevent the deep learning network from excessively affecting the main branch, thus ensuring the model's accuracy and facilitating subsequent training and optimization.

[0095] The loss function is configured as follows: The aforementioned deep learning network used for extracting the second feature representation is used as a frozen teacher network. This deep learning network has undergone pre-training on a large number of hepatobiliary surgical cases, has built-in mature clinical semantic knowledge, and its parameters are frozen throughout the process. The teacher network is used to encode the second surgical topology map using graph features to extract target clinical semantic features. Each dimension of this feature corresponds to a clear semantic meaning of hepatobiliary anatomy or surgical path, serving as a benchmark for subsequent alignment. It should be noted that during the pre-training phase, the teacher network, i.e., the deep learning network, uses multi-level classification labels as supervision signals. The first-level labels force the features to learn safety constraint semantics such as remaining liver volume, resection margin distance, and Child-Pugh classification. The second-level labels force the features to learn surgical semantics such as tumor-vascular spatial relationships and the extent of liver segment dominance. The third-level labels force the features to learn predictive semantics such as predicting postoperative follow-up complications.

[0096] The second topology control matrix output from the preceding steps is obtained. This matrix undergoes dimensional transformation and spatial mapping, projecting it into the feature space corresponding to the target clinical semantic features to obtain the corresponding generated features. This establishes a semantic coordinate system between the control branch features and the teacher network features. Each discrete feature map in the generated features is extracted, and its cosine similarity to the corresponding surgical anatomical configuration in the target clinical semantic features is calculated. The mean of all similarities is taken, and a negative value is used as the alignment loss function for jointly updating the main branch and the control branch. This alignment loss function measures the degree of deviation between the control branch output and the target clinical semantics. When there is a semantic conflict between the control branch output and the target clinical semantics, the similarity value is low, the negative similarity value is high, the deviation is greater, and the loss value is higher. Conversely, when there is semantic consistency between the control branch output and the target clinical semantics, the similarity value is high, the negative similarity value is low, the deviation is smaller, and the loss is smaller.

[0097] By minimizing the overall training objective, specifically by calculating the weighted sum of the cross-entropy and alignment loss of the multi-level classification labels, gradient backpropagation is required. Since the parameters of the main branch and the teacher network are frozen, the gradient only flows to the trainable parameters of the control branch, i.e., the feature processing layers of each graph in the control branch, thus achieving iterative optimization only on the control branch. When the overall training objective converges to the preset second convergence threshold, the trained multi-level screening model is output. At this point, the control branch has the same target clinical semantic understanding ability as the teacher network. The second convergence threshold is less than or equal to the first convergence threshold, and its value is usually 0.05.

[0098] By introducing a main branch with frozen execution parameters and a trainable control branch into the second surgical topology map and the initial screening model, a graph control network architecture was constructed. By freezing the weight parameters of the main branch, the target clinical semantics in the historical training set can be strongly preserved. At the same time, combined with the iterative training of the parameters of the control branch, cross-branch joint iterative training of the parameters of the control branch and the main branch is performed at the matrix level through one-to-one feature alignment and element addition, and a targeted loss function is configured to generate a well-trained multi-level screening model for hepatobiliary surgery plans. By combining multi-level classification labels, millisecond-level adaptive circuit breaking interception can be performed on cross-modal temporal mismatch, outdated or high-risk candidate treatment plan fragments. At the underlying data source level, the potential risks of clinical misdiagnosis and plan deviation are locked out. This not only ensures the high security of complex hepatobiliary surgery decision-making under high-concurrency multimodal requests, but also significantly improves the computational efficiency of automatic optimization of surgical topology plan screening and reduces the system's decision latency.

[0099] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from the most recent real-world situation by collecting a large amount of data and conducting software simulations. The formulas are set by those skilled in the art according to the actual situation.

[0100] 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.

[0101] 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.

[0102] 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 constructing a multi-level screening model for hepatobiliary surgical procedures based on artificial intelligence, characterized in that, The method includes: Obtain a training set of hepatobiliary surgery protocols; The training set includes multimodal clinical indicators of the target patients, candidate treatment options, and multi-level classification labels. Based on the training set of hepatobiliary surgery protocols, a first surgical topology map was drawn. Combined with multi-level classification labels, deep learning was performed on the first surgical topology map and at least two layers of reconstructed topology network. The deep learning included setting up a preset shallow learning network and a deep learning network for iterative training to build an initial screening model for hepatobiliary surgery protocols. In the second surgical topology map and the initial screening model for hepatobiliary surgery plans, a main branch with frozen execution parameters and a control branch with trainable parameters are introduced. By configuring the loss function, the parameters of the control branch and the main branch are iteratively trained to generate a well-trained multi-level screening model for hepatobiliary surgery plans.

2. The method for constructing a multi-level screening model for hepatobiliary surgery based on artificial intelligence according to claim 1, characterized in that, Drawing the topological map of the first surgery, including: Extracting sequences of hepatobiliary anatomical structures and physiological changes based on multimodal clinical indicators; Extraction of surgical liver tangent lines based on candidate treatment options; A multimodal feature matrix is ​​constructed, PCA is used to determine the number of levels of multi-level classification labels, and clustering algorithms are applied to determine the screening centers and safety constraints of each level of scheme. By using the hepatobiliary anatomical structure as nodes, a virtual three-dimensional dissection is performed along the surgical liver tangent to obtain the surgical anatomical configuration. The edges of the nodes are connected according to the vascular flow relationship to draw the first surgical topology map.

3. The method for constructing a multi-level screening model for hepatobiliary surgery based on artificial intelligence according to claim 2, characterized in that, Clustering algorithms are used to determine the selection centers and safety constraints for each level of the scheme, including: Select the same number of data points as the number of levels as the initial centroids, iteratively calculate the Euclidean distance between the data points and the initial centroids until the initial centroids converge. Determine the position of the converged centroids as the screening center of each level of scheme, and determine the boundary thresholds within the clusters corresponding to the screening centers of each level of schemes to form safe boundary constraints. The response times of the screening centers at each level are extracted and determined simultaneously. The time delay correction description function of each modal response time is established according to the modal category to form the time disturbance description quantity. Using the time disturbance description quantity, the screening centers at each level are sorted and grouped, and the time period threshold is determined to form the time response constraint.

4. The method for constructing a multi-level screening model for hepatobiliary surgery based on artificial intelligence according to claim 1, characterized in that, The process of reconstructing the topology network in two layers includes: Random initial vectors are used to identify the weight vectors of each node in the first surgical topology graph. The first surgical topology graph is reconstructed according to the adjacency relationship to obtain the second surgical topology graph, and it is marked as performing a layer of topology reconstruction network. A shallow learning network is used to extract features from the topological map of the first surgery to obtain a first feature representation. The first feature representation includes a first subset that satisfies the safety constraints, a second subset that does not satisfy the safety constraints, and a third subset that has fuzzy safety constraints. Simultaneously, a deep learning network is used to extract a second feature representation from the topological map of the second surgery through a two-layer reconstruction network. The second feature representation includes a fourth subset that satisfies the safety constraints and a fifth subset that does not satisfy the safety constraints. The first, third, and fourth subsets are analyzed for determination, and the initial screening model parameters are iteratively updated in conjunction with the multi-level classification labels. The second, third, and fifth subsets are analyzed for determination, and the initial screening model parameters are iteratively updated in conjunction with the multi-level classification labels.

5. The method for constructing a multi-level screening model for hepatobiliary surgical procedures based on artificial intelligence according to claim 3 or 4, characterized in that, The safety constraints must be met, including: Determine whether the first feature representation satisfies the safety boundary constraint, and if the safety boundary constraint is satisfied, determine whether the time response constraint is satisfied; if the first feature representation satisfies the time response constraint, output the first screening result and generate at least one of the first subset and the fourth subset; if the first feature representation does not satisfy the time response constraint, output the second screening result and generate at least one of the second subset and the fifth subset, and sort and group the candidate response times of each modality and perform adjustment processing; if the first feature representation does not satisfy the safety boundary constraint, terminate the operation.

6. The method for constructing a multi-level screening model for hepatobiliary surgery based on artificial intelligence according to claim 4, characterized in that, The initial screening model parameters are defined as follows: The cascaded temporal structure of the initial screening model is determined based on the screening centers at each level, the number of levels, and the security constraints. Based on the subsets of the first and second feature representations and the generated candidate treatment fragments, an evaluation strategy is executed to determine the distribution function; the least squares method is used to process the cascaded temporal structure and distribution function to obtain the initial screening model parameters.

7. The method for constructing a multi-level screening model for hepatobiliary surgery based on artificial intelligence according to claim 6, characterized in that, Implement evaluation strategies, including: Based on the time-series deviation of each subset of the first and second feature representations, the candidate treatment plans are sorted according to the magnitude of the deviation to determine their priority at the response time. Using the priority, feature screening is performed on each subset of the first and second feature representations, and the selected subset features are aggregated to generate a fusion representation quantity. The mean and variance of the fusion representation quantity in the hepatobiliary surgery plan training set are statistically analyzed, and the sum of the mean and the variance of multiples of k is set as the distribution function.

8. The method for constructing a multi-level screening model for hepatobiliary surgery based on artificial intelligence according to claim 6, characterized in that, The control branch training process includes: The control branch has M graph feature processing layers; A composite feature representation is generated based on the first feature representation and the second feature representation, and a binary mask sequence corresponding to the composite feature representation is constructed; wherein, an all-zero mask is assigned to the first feature representation, and an all-one mask is assigned to the second feature representation. The composite feature representation and the binary mask sequence are mapped to the first feature space. Feature concatenation is performed in the channel dimension. Using semantic analysis and multi-level classification labels, a topological conditional unit is assembled and input into the parameter-trainable control branch. Feature transformation is performed on the topological conditional unit to extract the first topological control matrix. The corresponding weights of the first topological control matrix are adjusted using the encoder built into the control branch.

9. The method for constructing a multi-level screening model for hepatobiliary surgery based on artificial intelligence according to claim 8, characterized in that, The main branch training process includes: The main branch consists of N cascaded graph feature processing layers, where N≥4 and M≤N. At the same time, the i-th graph feature processing layer in the control branch is set up, which corresponds one-to-one with the i-th graph feature processing layer in the main branch in terms of network position and feature dimension, where i is a positive integer greater than 0 and less than or equal to M. By obtaining the first topology control matrix after adjusting the weights, the first topology control matrix after adjusting the weights is added element by element to the intermediate hidden state output by the main branch with the parameters frozen, and then injected into the graph feature processing layer of the main branch to output the second topology control matrix.

10. The method for constructing a multi-level screening model for hepatobiliary surgery based on artificial intelligence according to claim 9, characterized in that, Configure the loss function, including: A pre-trained deep learning network is used to encode graph features of the second surgical topology map and extract target clinical semantic features; the second topology control matrix is ​​obtained; the second topology control matrix is ​​subjected to dimensional transformation and spatial mapping and projected into the feature space corresponding to the target clinical semantic features to obtain the corresponding generated features; Extract each discrete feature map from the generated features and calculate the mean negative cosine similarity between it and the corresponding surgical anatomical configuration in the target clinical semantic features. Use this as the alignment loss function to jointly update the main branch and the control branch. By minimizing the overall training objective, gradient backpropagation is used to keep the parameters of the main branch frozen. Only the trainable parameters of the control branch are iteratively optimized to output the trained multi-level screening model.

Images