Esophageal cancer operation dynamic risk assessment method based on topological data analysis

Abstract

The invention discloses an esophageal cancer operation dynamic risk assessment method based on topological data analysis. The method comprises the following steps: S1, constructing a multi-modal medical data set; s2, constructing a topological model, and extracting topological features of the multi-modal medical data set; s3, processing the extracted topological features by using a bidirectional generative adversarial network; s4, dynamically updating the topological model and extracting current topological characteristics in the operation process; s5, evaluating the risk of the current operation in real time, and generating a corresponding risk evaluation result and an operation path adjustment suggestion; s6, in the risk assessment process, when the system detects a potential risk, suggestions for adjusting the operation path are provided for a doctor according to an assessment result; and S7, storing all data generated in the operation process in a database for risk management and rehabilitation plans in the recovery stage of the postoperative patient. According to the invention, the possibility of operation failure is effectively reduced, and the success rate of the operation and the postoperative recovery quality of a patient are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of esophageal cancer, and in particular to a dynamic risk assessment method for esophageal cancer surgery based on topological data analysis. Background Art

[0002] With the advancement of technology in the medical field, real-time dynamic risk assessment in surgical operations has gradually become a research hotspot. In complex esophageal cancer operations, risk assessment is crucial to the success rate of the operation and the patient's postoperative recovery. Surgical risk assessment in existing technologies mostly relies on static data before the operation, including the patient's preoperative medical history, imaging data, and preoperative physiological parameters. Static methods evaluate potential risks by analyzing preoperative data, usually using statistical models or experience-based rules for judgment. However, risk factors during surgery will change dynamically with time and changes in operations. Traditional static assessment methods cannot be updated and adjusted in real time during the operation, and therefore lack the ability to respond to real-time changes during surgery.

[0003] Existing risk assessment methods, such as prediction systems based on statistical models, can provide predictions of overall risks before surgery, but they cannot handle real-time complex conditions that occur during surgery. In addition, physiological data and image information during surgery are usually nonlinear and high-dimensional, and traditional linear statistical models are difficult to accurately capture these complex change patterns. During surgery, risk factors such as blood pressure, heart rate, and oxygen saturation physiological parameters will undergo dynamic changes, and these changes often have a significant impact on surgical outcomes. Existing technologies cannot effectively integrate these multimodal data, resulting in delayed risk warnings or misjudgments. In addition, most existing surgical assistance systems lack the ability to provide intelligent decision-making support for risk changes during surgery, and usually rely only on preoperative prediction results and manual judgment during surgery.

[0004] There are several main problems with existing technologies when facing complex esophageal cancer surgeries: First, existing risk assessment methods rely on static data, lack the ability of dynamic real-time assessment, and cannot respond to emergencies during the operation in a timely manner; second, existing technologies have difficulty processing multi-dimensional and complex physiological and pathological data during the operation, and cannot fully integrate and analyze these data to accurately capture risk changes; finally, existing surgical assistance systems lack intelligent real-time risk warning and path adjustment functions, making it difficult to provide surgeons with effective surgical decision support. Technical limitations make existing surgical risk assessment methods perform less than ideal when faced with complex and changeable surgical scenarios. Summary of the invention

[0005] One purpose of the present invention is to propose a dynamic risk assessment method for esophageal cancer surgery based on topological data analysis, which effectively reduces the possibility of surgical failure and significantly improves the success rate of surgery and the quality of postoperative recovery of patients.

[0006] A method for dynamic risk assessment of esophageal cancer surgery based on topological data analysis according to an embodiment of the present invention comprises the following steps:

[0007] S1. Obtain the patient's medical imaging data, physiological parameter data and pathological data, construct a multimodal medical data set, and preprocess the multimodal medical data set to perform denoising, standardization and feature extraction;

[0008] S2. Building a topological model based on the preprocessed multimodal medical data set, extracting the topological features of the multimodal medical data set, and using the topological features to reveal the nonlinear change pattern of the patient's status during surgery;

[0009] S3. The extracted topological features are processed using a bidirectional generative adversarial network, where the bidirectional generator generates potential surgical risk patterns and the bidirectional discriminator distinguishes the potential surgical risk patterns from the risk status during the actual surgery;

[0010] S4. During the operation, the patient's multimodal medical data set is acquired in real time, the topological model is dynamically updated, and the current topological features are extracted;

[0011] S5. Input the dynamically updated topological features into the trained bidirectional generative adversarial network model, bidirectionally evaluate the risk of the current surgery in real time based on the output of the discriminator, and generate corresponding risk assessment results and surgical path adjustment suggestions;

[0012] S6. During the risk assessment process, when the system detects potential risks, the bidirectional generator predicts the future risk change trend, the bidirectional discriminator provides risk warning information, and provides doctors with suggestions for adjusting the surgical path based on the assessment results;

[0013] S7. All data generated during the operation are stored in the database for risk management and rehabilitation planning during the postoperative patient recovery phase.

[0014] Optionally, the S1 specifically includes:

[0015] S11. Obtaining the patient's medical imaging data I m , physiological parameter data P s and pathological data P d ;

[0016] S12. Construction of multimodal medical dataset:

[0017] D m = {Im ,P s ,P d};

[0018] Among them, medical imaging data I m Including CT images and MRI images, physiological parameter data P s Including heart rate, blood pressure and blood oxygen saturation, pathological data P d Including tumor tissue pathological characteristics and genetic testing results;

[0019] S13. The multimodal medical dataset D m Perform preprocessing to obtain the preprocessed multimodal medical dataset D m,processed .

[0020] Optionally, the S2 specifically includes:

[0021] S21. Constructing a topological model T(D) based on the preprocessed multimodal medical dataset m ), when constructing the topological model, the persistent homology algorithm is used to process the multimodal medical data set, extract the topological features of each dimension in the multimodal medical data set, and calculate the topological feature matrix T(D m ), topological models reveal the hidden structures in multimodal medical data by maintaining the topological relationships between high-dimensional data in a parsimonious manner;

[0022] S22. Extract the topological feature matrix H p Reveal the nonlinear change pattern of the patient's state during surgery. The nonlinear change pattern reflects the fluctuation of the patient's physiological state by detecting the change trend of the topological feature matrix at different time points.

[0023] S23. The constructed topological model T(D m ) is dynamically adjusted to enable the topological model to adaptively handle data changes during the operation.

[0024] Optionally, the step S21 specifically includes:

[0025] S211. Construct a topological model based on the preprocessed multimodal medical data set, and use a persistent homology algorithm to process the data to extract topological features of each dimension;

[0026] S212. The persistent homology algorithm captures the topological information of different dimensions in the multimodal medical data set by constructing simplified complexes of different scales. The simplified complex K is defined as:

[0027]

[0028] Among them, σ i is an i-dimensional simplex consisting of k+1 vertices v iComposition, ω i is the weight function, ω i It changes according to the different dimensions of medical data extracted during the operation, reflecting the difference in weights of physiological, imaging and pathological data in dynamic surgical evaluation;

[0029] S213. Computing persistent homology H based on the simplified complex K p (K):

[0030]

[0031] Among them, Φ p (K(t)) represents the p-persistent homology characteristic of K at time t, λ p (t) is the life cycle length of each p-dimensional feature, f(λ p (t)) is a nonlinear function that adjusts the weight of the survival period, T represents the duration of surgery;

[0032] S214. Through persistent homology H p (K) Extracting the topological feature matrix H from multimodal medical datasets p :

[0033]

[0034] Among them, α i is a simplex i The weight coefficient in feature extraction, ψ(σ i ) represents the simplex σ i The topological property function of Γ(p) is the nonlinear transformation matrix in multidimensional space, which is used to reveal the potential nonlinear change pattern during surgery.

[0035] Optionally, the step S3 specifically includes:

[0036] S31. Based on the extracted topological feature matrix H p A bidirectional generative adversarial network model is constructed. The bidirectional generative adversarial network model includes a generator G, an encoder E, and a discriminator D. The generator G is used to generate potential surgical risk patterns from the noise space, the encoder E is used to generate potential codes from the actual surgical data, and the discriminator D is used to judge the authenticity of the generated potential pattern and the actual surgical risk data and the consistency of the latent space;

[0037] S32. Generator G receives the topological feature matrix H p and random noise vector z to generate potential surgical risk patterns

[0038] S33. Encoder E receives actual surgical risk status data R(H p), mapping the actual surgical risk status data back to the latent space and generating the corresponding latent code z';

[0039] S34. The discriminator D simultaneously receives the potential risk pattern generated by the generator and the potential code z' generated by the encoder, and the actual surgical risk data R(H p ) and the noise vector z input by the generator, distinguishing the difference between the generated pattern and the real data, the generated noise and the potential code, the loss function of the discriminator is defined as:

[0040]

[0041] Among them, D(R(H p ),z ′ ) is the probability of the discriminator judging the true surgical risk state and the corresponding potential code, The probability of the discriminator judging the generated pattern and the generated noise;

[0042] S35. The generator G and encoder E are trained by optimizing their respective loss functions. The loss function of the generator is L G and the encoder loss function L E Combined with the topological feature matrix L E And the distribution characteristics of the noise space z are designed, the loss function L of the generator G It consists of the following parts:

[0043]

[0044] in, represents the generated potential risk pattern, represents the probability of the discriminator judging the relationship between the generated risk pattern and the noise space, λ G is the regularization coefficient, is the gradient of the generator for the i-th feature in the topological feature matrix, is the variation of the topological feature at different time points during the operation, represents the discriminability of the potential risk patterns generated by the generator and the real surgical risk status;

[0045] The encoder loss function L E It consists of the following parts:

[0046]

[0047] Among them, z ′ =E(R(H p )) represents the potential encoding of the encoder for the true surgical risk status, α E is the adjustment factor for encoder loss, represents the distance between the latent code z' and the noise vector z, γ i is the weight of the topological feature, represents the nonlinear mapping of the ith topological feature in the real surgical risk state, Represents the mapping of the generator to the same features;

[0048] S36. Through the adversarial training of the bidirectional generative adversarial network, the generator G and the encoder E are optimized. The potential risk pattern generated by the generator and the potential representation generated by the encoder are close to the actual risk state. The discriminator D improves the ability to distinguish between the generated potential pattern and the actual surgical risk state. The adversarial loss function L Bi-GAN The comprehensive loss of the generator, encoder and discriminator is expressed as:

[0049]

[0050] in, represents the probability of the discriminator's judgment on the actual surgical risk state and the potential coding, represents the probability of the discriminator's judgment on the generated mode and noise, β is the weight coefficient of the gradient penalty term, which is used to control the responsiveness of the generator and encoder to changes in topological features. is the loss gradient of the generator, encoder and discriminator under the i-th topological feature, is the Frobenius norm.

[0051] Optionally, the step S6 specifically includes:

[0052] S61. Obtain multimodal medical data of patients in real time during risk assessment m,real-time , and update the topological model T(D m )'s topological characteristic matrix H p (t);

[0053] S62. The bidirectional generator G updates the topological feature matrix H in real time by inputting p (t) Predict future risk trends

[0054]

[0055] Where Δt represents the predicted future time interval, z is the random noise vector used by the generator, The potential risk change trend in the future;

[0056] S63. The bidirectional discriminator D receives the potential risk pattern generated during the actual surgery. and real-time risk data R(H p(t)), and provide risk warning information based on the identification results:

[0057]

[0058] Among them, D(R(H p (t))) represents the judgment of the discriminator on the current actual risk data, Represents the discriminator's judgment on the generated future risk change trend;

[0059] S64. Based on the output result of the discriminator, when the system detects a potential risk, it generates corresponding risk warning information, predicts the degree and location of the risk change, and sends the risk warning information to the surgeon through the system;

[0060] S65. Based on the real-time evaluation of the risk warning information system, the surgeon is provided with suggestions for adjusting the surgical path according to the risk level. The adjustment suggestions include the adjustment direction, adjustment range and expected risk reduction effect of the surgical path:

[0061]

[0062] Among them, w i is the weight adjusted for different paths, ΔP i represents the expected effect of each path adjustment on risk reduction.

[0063] Optionally, the step S64 specifically includes:

[0064] S641. When the discriminator D compares the real-time risk data R(H p (t)) and the resulting future risk change trend When a potential risk is detected, a risk warning message is generated, which includes the specific location where the risk occurs. i , Risk severity S i and the expected risk trend R i ;

[0065] S642. The system sends the risk warning information to the surgeon in real time via wireless means, and the 3D visualization model of the surgical area marks the location where the risk occurs in combination with the risk warning information. i , and color-coded to show the severity of the risk S i , the darker the color, the greater the risk;

[0066] S643. The system uses the time factor T in risk assessment f Predict the time t+Δt at which the future risk will occur, and write the timestamp T of the warning information f (t+Δt) Synchronize to the doctor interface:

[0067]

[0068] Among them, α is the adjustment coefficient of risk severity, S i The larger it is, the shorter Δt is, indicating a more urgent risk.

[0069] The beneficial effects of the present invention are:

[0070] (1) The present invention acquires the patient's multimodal medical data in real time during the operation, and uses topological data analysis technology to construct a topological model to effectively extract and capture the complex change patterns in multidimensional data. Different from the traditional risk assessment method that relies on static data, the present invention uses a persistent coherence algorithm to extract the nonlinear topological features of the patient's physiological, imaging and pathological data during the operation, and dynamically updates the topological model, so that the system can timely capture changes in the patient's status and predict potential surgical risks in real time, thereby achieving more accurate dynamic risk assessment and greatly improving surgical safety.

[0071] (2) The present invention significantly enhances the ability to predict surgical risks by introducing a bidirectional generative adversarial network model. The bidirectional generator generates potential surgical risk patterns based on the topological feature matrix extracted in real time, and the bidirectional discriminator determines the risk status based on the difference between the generated risk pattern and the actual data. Through adversarial training, the generator continuously optimizes the generated risk pattern, and the discriminator continuously improves its ability to distinguish between potential risks and actual surgical risks, ultimately forming an accurate risk prediction system.

[0072] (3) When detecting potential surgical risks, the system of the present invention can combine the prediction results of the bidirectional generative adversarial network to provide surgeons with intelligent surgical path adjustment suggestions. The system automatically generates the optimal path adjustment plan based on the topological characteristics and risk change trends, and combines the risk warning information to display the degree and location of the risk change and the surgical path adjustment suggestions to the surgeon in real time. During the operation, it not only improves the timeliness and accuracy of risk identification, but also effectively reduces the possibility of surgical failure, significantly improving the success rate of the operation and the quality of postoperative recovery of the patient. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0074] Figure 1 A flowchart of a dynamic risk assessment method for esophageal cancer surgery based on topological data analysis proposed by the present invention;

[0075] Figure 2This is a schematic diagram of the structure of a bidirectional generative adversarial network in a dynamic risk assessment method for esophageal cancer surgery based on topological data analysis proposed in the present invention. DETAILED DESCRIPTION

[0076] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0077] refer to Figure 1-2 , a dynamic risk assessment method for esophageal cancer surgery based on topological data analysis, comprising the following steps:

[0078] S1. Obtain the patient's medical imaging data, physiological parameter data and pathological data, construct a multimodal medical data set, and preprocess the multimodal medical data set to perform denoising, standardization and feature extraction;

[0079] S2. Building a topological model based on the preprocessed multimodal medical data set, extracting the topological features of the multimodal medical data set, and using the topological features to reveal the nonlinear change pattern of the patient's status during surgery;

[0080] S3. The extracted topological features are processed using a bidirectional generative adversarial network, where the bidirectional generator generates potential surgical risk patterns and the bidirectional discriminator distinguishes the potential surgical risk patterns from the risk status during the actual surgery;

[0081] S4. During the operation, the patient's multimodal medical data set is acquired in real time, the topological model is dynamically updated, and the current topological features are extracted;

[0082] S5. Input the dynamically updated topological features into the trained bidirectional generative adversarial network model, bidirectionally evaluate the risk of the current surgery in real time based on the output of the discriminator, and generate corresponding risk assessment results and surgical path adjustment suggestions;

[0083] S6. During the risk assessment process, when the system detects potential risks, the bidirectional generator predicts the future risk change trend, the bidirectional discriminator provides risk warning information, and provides doctors with suggestions for adjusting the surgical path based on the assessment results;

[0084] S7. All data generated during the operation are stored in the database for risk management and rehabilitation planning during the postoperative patient recovery phase.

[0085] In this implementation, S1 specifically includes:

[0086] S11. Obtaining the patient's medical imaging data I m , physiological parameter data P s and pathological data P d;

[0087] S12. Construction of multimodal medical dataset:

[0088] D m = {I m ,P s ,P d};

[0089] Among them, medical imaging data I m Including CT images and MRI images, physiological parameter data P s Including heart rate, blood pressure and blood oxygen saturation, pathological data P d Including tumor tissue pathological characteristics and genetic testing results;

[0090] S13. Multimodal medical dataset D m Perform preprocessing to obtain the preprocessed multimodal medical dataset D m,processed .

[0091] In this implementation, S2 specifically includes:

[0092] S21. Constructing a topological model T(D) based on the preprocessed multimodal medical dataset m ), when constructing the topological model, the persistent homology algorithm is used to process the multimodal medical data set, extract the topological features of each dimension in the multimodal medical data set, and calculate the topological feature matrix T(D m ), topological models reveal the hidden structures in multimodal medical data by maintaining the topological relationships between high-dimensional data in a parsimonious manner;

[0093] S22. Extract the topological feature matrix H p Reveal the nonlinear change pattern of the patient's state during surgery. The nonlinear change pattern reflects the fluctuation of the patient's physiological state by detecting the change trend of the topological feature matrix at different time points.

[0094] S23. The constructed topological model T(D m ) is dynamically adjusted to enable the topological model to adaptively handle data changes during the operation.

[0095] In this implementation, step S21 specifically includes:

[0096] S211. Construct a topological model based on the preprocessed multimodal medical data set, and use a persistent homology algorithm to process the data to extract topological features of each dimension;

[0097] S212. The persistent homology algorithm captures the topological information of different dimensions in the multimodal medical data set by constructing simplified complexes of different scales. The simplified complex K is defined as:

[0098]

[0099] Among them, σ i is an i-dimensional simplex consisting of k+1 vertices v i Composition, ω i is the weight function, ω i It changes according to the different dimensions of medical data extracted during the operation, reflecting the difference in weights of physiological, imaging and pathological data in dynamic surgical evaluation;

[0100] S213. Computing persistent homology H based on the simplified complex K p (K):

[0101]

[0102] Among them, Φ p (K(t)) represents the p-persistent homology characteristic of K at time t, λ p (t) is the life cycle length of each p-dimensional feature, f(λ p (t)) is a nonlinear function that adjusts the weight of the survival period, T represents the duration of surgery;

[0103] S214. Through persistent homology H p (K) Extracting the topological feature matrix H from multimodal medical datasets p :

[0104]

[0105] Among them, α i is a simplex i The weight coefficient in feature extraction, ψ(σ i ) represents the simplex σ i The topological property function of Γ(p) is the nonlinear transformation matrix in multidimensional space, which is used to reveal the potential nonlinear change pattern during surgery.

[0106] In this implementation, step S3 specifically includes:

[0107] S31. Based on the extracted topological feature matrix H p A bidirectional generative adversarial network model is constructed. The bidirectional generative adversarial network model includes a generator G, an encoder E, and a discriminator D. The generator G is used to generate potential surgical risk patterns from the noise space, the encoder E is used to generate potential codes from the actual surgical data, and the discriminator D is used to judge the authenticity of the generated potential pattern and the actual surgical risk data and the consistency of the latent space;

[0108] S32. Generator G receives the topological feature matrix H pand random noise vector z to generate potential surgical risk patterns

[0109] S33. Encoder E receives actual surgical risk status data R(H p ), mapping the actual surgical risk status data back to the latent space and generating the corresponding latent code z';

[0110] S34. The discriminator D simultaneously receives the potential risk pattern generated by the generator and the potential code z' generated by the encoder, and the actual surgical risk data R(H p ) and the noise vector z input by the generator, distinguishing the difference between the generated pattern and the real data, the generated noise and the potential code, the loss function of the discriminator is defined as:

[0111]

[0112] Among them, D(R(H p ),z ′ ) is the probability of the discriminator judging the true surgical risk state and the corresponding potential code, The probability of the discriminator judging the generated pattern and the generated noise;

[0113] S35. The generator G and encoder E are trained by optimizing their respective loss functions. The loss function of the generator is L G and the encoder loss function L E Combined with the topological feature matrix L E And the distribution characteristics of the noise space z are designed, the loss function L of the generator G It consists of the following parts:

[0114]

[0115] in, represents the generated potential risk pattern, represents the probability of the discriminator judging the relationship between the generated risk pattern and the noise space, λ G is the regularization coefficient, is the gradient of the generator for the i-th feature in the topological feature matrix, is the variation of the topological feature at different time points during the operation, represents the discriminability of the potential risk patterns generated by the generator and the real surgical risk status;

[0116] The encoder loss function L E It consists of the following parts:

[0117]

[0118] Among them, z ′ =E(R(H p )) represents the potential encoding of the encoder for the true surgical risk status, α E is the adjustment factor for encoder loss, represents the distance between the latent code z' and the noise vector z, γ i is the weight of the topological feature, represents the nonlinear mapping of the ith topological feature in the true surgical risk state, Represents the mapping of the generator to the same features;

[0119] S36. Through the adversarial training of the bidirectional generative adversarial network, the generator G and the encoder E are optimized. The potential risk pattern generated by the generator and the potential representation generated by the encoder are close to the actual risk state. The discriminator D improves the ability to distinguish between the generated potential pattern and the actual surgical risk state. The adversarial loss function L Bi-GAN The comprehensive loss of the generator, encoder and discriminator is expressed as:

[0120]

[0121] in, represents the probability of the discriminator's judgment on the actual surgical risk state and the potential coding, represents the probability of the discriminator's judgment on the generated mode and noise, β is the weight coefficient of the gradient penalty term, which is used to control the responsiveness of the generator and encoder to changes in topological features. is the loss gradient of the generator, encoder and discriminator under the i-th topological feature, is the Frobenius norm.

[0122] In this implementation, step S6 specifically includes:

[0123] S61. Obtain multimodal medical data of patients in real time during risk assessment m,real-time , and update the topological model T(D m )’s topological characteristic matrix H p (t);

[0124] S62. The bidirectional generator G updates the topological feature matrix H in real time by inputting p (t) Predict future risk trends

[0125]

[0126] Where Δt represents the predicted future time interval, z is the random noise vector used by the generator, The potential risk change trend in the future;

[0127] S63. The bidirectional discriminator D receives the potential risk pattern generated during the actual surgery. and real-time risk data R(H p (t)), and provide risk warning information based on the identification results:

[0128]

[0129] Among them, D(R(H p (t))) represents the judgment of the discriminator on the current actual risk data, Represents the discriminator's judgment on the generated future risk change trend;

[0130] S64. Based on the output result of the discriminator, when the system detects a potential risk, it generates corresponding risk warning information, predicts the degree and location of the risk change, and sends the risk warning information to the surgeon through the system;

[0131] S65. Based on the real-time evaluation of the risk warning information system, the surgeon is provided with suggestions for adjusting the surgical path according to the risk level. The adjustment suggestions include the adjustment direction, adjustment range and expected risk reduction effect of the surgical path:

[0132]

[0133] Among them, w i is the weight adjusted for different paths, ΔP i represents the expected effect of each path adjustment on risk reduction.

[0134] In this implementation, step S64 specifically includes:

[0135] S641. When the discriminator D compares the real-time risk data R(H p (t)) and the resulting future risk change trend When a potential risk is detected, a risk warning message is generated, which includes the specific location where the risk occurs. i , Risk severity S i and the expected risk trend R i ;

[0136] S642. The system sends the risk warning information to the surgeon in real time via wireless means, and the 3D visualization model of the surgical area marks the location where the risk occurs in combination with the risk warning information. i , and color-coded to show the severity of the risk S i , the darker the color, the greater the risk;

[0137] S643. The system uses the time factor T in risk assessment f Predict the time t+Δt at which the future risk will occur, and write the timestamp T of the warning information f (t+Δt) Synchronize to the doctor interface:

[0138]

[0139] Among them, α is the adjustment coefficient of risk severity, S i The larger it is, the shorter Δt is, indicating a more urgent risk.

[0140] Embodiment 1:

[0141] In order to verify the feasibility of the present invention in implementation, the implementers conducted experiments and simulations on the dynamic risk assessment method in esophageal cancer surgery. The experiment used two sets of publicly available surgical and multimodal medical datasets: the MIMIC-IV dataset and the SurgAI dataset. MIMIC-IV is a public medical dataset that contains physiological parameters and medical imaging data from real ICU patients, covering a large number of different types of surgical scenarios, including complex esophageal cancer surgery-related data. The SurgAI dataset focuses on real-time medical imaging and physiological data acquisition during surgical operations, which is used to train and evaluate algorithm models of surgical assistance systems. In order to be closer to the real surgical scene, the implementers selected 80 esophageal cancer surgery case data in MIMIC-IV, covering the period from 2016 to 2021, and conducted risk assessment experiments in combination with the real-time imaging data of esophageal cancer surgery in SurgAI.

[0142] The implementers processed and divided the two data sets. The implementers divided the MIMIC-IV and SurgAI data sets into training sets and test sets in a ratio of 7:3, ensuring that typical cases of complex esophageal cancer surgery were retained in the test set. For multimodal data processing, physiological data included patients' real-time blood pressure, heart rate, and blood oxygen saturation, while imaging data included intraoperative CT images and endoscopic images. The implementers used ResNet50 to extract features from CT and endoscopic images, and the image embedding dimension was 2048. For physiological data, the implementers used recurrent neural networks to extract time series features, and the embedding dimension of physiological data was 512. In order to unify the data dimension, the implementers introduced a multi-layer perceptron to convert imaging and physiological features into the same dimension, and formed the final multimodal feature matrix through splicing.

[0143] The model of the present invention adopts a persistent homology algorithm to extract topological features in multimodal data, and combines a bidirectional generative adversarial network to predict potential risks during surgery. Specifically, the generator generates potential surgical risk patterns based on the multimodal feature matrix collected in real time during surgery, and the discriminator continuously optimizes the generator's prediction ability by comparing the generated risk patterns with the actual risk states. The implementer uses the Adam optimizer to train the model, with the learning rate set to 0.001, the batch size to 32, and the total number of training epochs to 150. In terms of loss function, the generator uses a combination of adversarial loss and regularization loss of topological feature changes, and the discriminator uses binary cross entropy loss.

[0144] In order to verify the effectiveness of the present invention in complex surgical scenarios, the implementers compared it with the traditional static risk assessment method. The traditional method uses preoperative data and statistical models for risk prediction, which cannot process real-time changing data during surgery. The main evaluation indicators of the implementers include the accuracy of real-time risk prediction during surgery, abnormal event detection time, and surgical success rate.

[0145] The experimental results on the MIMIC-IV dataset are shown in Table 1:

[0146]

[0147] The experimental results on the SurgAI dataset are shown in Table 2:

[0148]

[0149] It can be seen from the experimental results that the dynamic risk assessment method of the present invention is significantly superior to the traditional static assessment method on both the MIMIC-IV and SurgAI datasets. In the MIMIC-IV dataset, the accuracy of the method of the present invention reached 89.2%, which is 17.7 percentage points higher than the 71.5% of the traditional method. At the same time, in the detection of key abnormal events, the model of the present invention can issue risk warnings within an average of 12 seconds, while the traditional method requires 45 seconds, with a significant gap. On the SurgAI dataset, the dynamic risk assessment method of the present invention also performed well, with an accuracy of 91.4%, which is 21.6 percentage points higher than the 69.8% of the traditional method. Especially in the real-time risk prediction and emergency response time during the operation, the present invention shows higher efficiency, and it only takes 10 seconds to detect potential risks and issue warnings, while the traditional method takes an average of 52 seconds.

[0150] To further verify the effectiveness of the model, the implementers selected a high-risk esophageal cancer surgery in the MIMIC-IV dataset for detailed analysis. The patient experienced a sudden drop in blood pressure and heart rate fluctuations during the operation. Traditional methods failed to provide timely warnings, resulting in the doctor interrupting the operation when the problem was discovered, and ultimately prolonging the patient's recovery time. Using the dynamic risk assessment method of the present invention, the system detected the risk within 8 seconds after the heart rate fluctuation occurred, and issued a potential warning of cardiac arrest within 12 seconds. The doctor immediately adjusted the surgical path according to the system prompts, avoiding more serious risks and shortening the patient's postoperative recovery time by 20%.

[0151] The above experiments have verified that the dynamic risk assessment method for esophageal cancer surgery based on topological data analysis and bidirectional generative adversarial networks of the present invention performs excellently in complex surgical scenarios. Compared with traditional static risk assessment methods, the present invention can capture the changes in physiological and imaging data during the operation in real time, issue risk warnings in a timely manner, and provide doctors with intelligent surgical path adjustment suggestions, thereby significantly improving the safety and success rate of the operation.

[0152] The present invention obtains the patient's multimodal medical data in real time during the operation, and uses topological data analysis technology to build a topological model to effectively extract and capture complex change patterns in multidimensional data. Different from traditional risk assessment methods that rely on static data, the present invention uses a persistent coherence algorithm to extract the nonlinear topological features of the patient's physiological, imaging and pathological data during the operation, and dynamically updates the topological model, so that the system can timely capture changes in the patient's status and predict potential surgical risks in real time, achieving more accurate dynamic risk assessment and greatly improving surgical safety.

[0153] The present invention significantly enhances the ability of surgical risk prediction by introducing a bidirectional generative adversarial network model. The bidirectional generator generates potential surgical risk patterns based on the topological feature matrix extracted in real time, and the bidirectional discriminator judges the risk status based on the difference between the generated risk pattern and the actual data. Through adversarial training, the generator continuously optimizes the generated risk pattern, and the discriminator continuously improves the ability to distinguish between potential risks and actual surgical risks, ultimately forming an accurate risk prediction system.

[0154] When detecting potential surgical risks, the system of the present invention can combine the prediction results of the bidirectional generative adversarial network to provide surgeons with intelligent surgical path adjustment suggestions. The system automatically generates the optimal path adjustment plan based on the topological characteristics and risk change trends, and combines the risk warning information to display the degree and location of the risk change and the surgical path adjustment suggestions to the surgeon in real time. During the operation, it not only improves the timeliness and accuracy of risk identification, but also effectively reduces the possibility of surgical failure, significantly improving the success rate of the operation and the patient's postoperative recovery quality.

[0155] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A dynamic risk assessment method for esophageal cancer surgery based on topological data analysis, characterized in that: The steps include: S1. Obtain the patient's medical imaging data, physiological parameter data and pathological data, construct a multimodal medical data set, and preprocess the multimodal medical data set to perform denoising, standardization and feature extraction; S2. Building a topological model based on the preprocessed multimodal medical data set, extracting the topological features of the multimodal medical data set, and using the topological features to reveal the nonlinear change pattern of the patient's status during surgery; S3. The extracted topological features are processed using a bidirectional generative adversarial network, where the bidirectional generator generates potential surgical risk patterns and the bidirectional discriminator distinguishes the potential surgical risk patterns from the risk status during the actual surgery; S4. During the operation, the patient's multimodal medical data set is acquired in real time, the topological model is dynamically updated, and the current topological features are extracted; S5. Input the dynamically updated topological features into the trained bidirectional generative adversarial network model, bidirectionally evaluate the risk of the current surgery in real time based on the output of the discriminator, and generate corresponding risk assessment results and surgical path adjustment suggestions; S6. During the risk assessment process, when the system detects potential risks, the bidirectional generator predicts the future risk change trend, the bidirectional discriminator provides risk warning information, and provides doctors with suggestions for adjusting the surgical path based on the assessment results; S7. All data generated during the operation are stored in the database for risk management and rehabilitation planning during the postoperative patient recovery phase.

2. According to claim 1, a dynamic risk assessment method for esophageal cancer surgery based on topological data analysis is characterized in that: The S1 specifically includes: S11. Obtaining the patient's medical imaging data I m , physiological parameter data P s and pathological data P d ; S12. Construction of multimodal medical dataset: D m ={I m ,P s ,P d }; Among them, medical imaging data I m Including CT images and MRI images, physiological parameter data P s Including heart rate, blood pressure and blood oxygen saturation, pathological data P d Including tumor tissue pathological characteristics and genetic testing results; S13. The multimodal medical dataset D m Perform preprocessing to obtain the preprocessed multimodal medical dataset D m,processed .

3. The method for dynamic risk assessment of esophageal cancer surgery based on topological data analysis according to claim 1, characterized in that: The S2 specifically includes: S21. Constructing a topological model T(D) based on the preprocessed multimodal medical dataset m ), when constructing the topological model, the persistent homology algorithm is used to process the multimodal medical data set, extract the topological features of each dimension in the multimodal medical data set, and calculate the topological feature matrix T(D m ), topological models reveal the hidden structures in multimodal medical data by maintaining the topological relationships between high-dimensional data in a parsimonious manner; S22. Extract the topological feature matrix H p Reveal the nonlinear change pattern of the patient's state during surgery. The nonlinear change pattern reflects the fluctuation of the patient's physiological state by detecting the change trend of the topological feature matrix at different time points. S23. The constructed topological model T(D m ) is dynamically adjusted to enable the topological model to adaptively handle data changes during the operation.

4. The method for dynamic risk assessment of esophageal cancer surgery based on topological data analysis according to claim 1, characterized in that: The step S21 specifically includes: S211. Construct a topological model based on the preprocessed multimodal medical data set, and use a persistent homology algorithm to process the data to extract topological features of each dimension; S212. The persistent homology algorithm captures the topological information of different dimensions in the multimodal medical data set by constructing simplified complexes of different scales. The simplified complex K is defined as: Among them, σ i is an i-dimensional simplex consisting of k+1 vertices v i Composition, ω i is the weight function, ω i It changes according to the different dimensions of medical data extracted during the operation, reflecting the difference in weights of physiological, imaging and pathological data in dynamic surgical evaluation; S213. Computing persistent homology H based on the simplified complex K p (K): Among them, Φ p (K(t)) represents the p-persistent homology feature of K at time t, λ p (t) is the life cycle length of each p-dimensional feature, f(λ p (t)) is a nonlinear function that adjusts the weight of the survival period, T represents the duration of surgery; S214. Through persistent homology H p (K) Extracting the topological feature matrix H from multimodal medical datasets p : Among them, α i is a simplex i The weight coefficient in feature extraction, ψ(σ i ) represents the simplex σ i The topological property function of Γ(p) is the nonlinear transformation matrix in multidimensional space, which is used to reveal the potential nonlinear change pattern during surgery.

5. The method for dynamic risk assessment of esophageal cancer surgery based on topological data analysis according to claim 1, characterized in that: The step S3 specifically includes: S31. Based on the extracted topological feature matrix H p A bidirectional generative adversarial network model is constructed. The bidirectional generative adversarial network model includes a generator G, an encoder E, and a discriminator D. The generator G is used to generate potential surgical risk patterns from the noise space, the encoder E is used to generate potential codes from the actual surgical data, and the discriminator D is used to judge the authenticity of the generated potential pattern and the actual surgical risk data and the consistency of the latent space; S32. Generator G receives topological feature matrix H p and random noise vector z to generate potential surgical risk patterns S33. Encoder E receives actual surgical risk status data R(H p ), mapping the actual surgical risk status data back to the latent space and generating the corresponding latent code z'; S34. The discriminator D simultaneously receives the potential risk pattern generated by the generator and the potential code z' generated by the encoder, and the actual surgical risk data R(H p ) and the noise vector z input by the generator, distinguishing the difference between the generated pattern and the real data, the generated noise and the potential code, the loss function of the discriminator is defined as: Among them, D(R(H p ),z ′ ) is the probability of the discriminator judging the true surgical risk state and the corresponding potential code, The probability of the discriminator judging the generated pattern and the generated noise; S35. The generator G and encoder E are trained by optimizing their respective loss functions. The loss function of the generator is L G and the encoder loss function L E Combined with the topological feature matrix L E And the distribution characteristics of the noise space z are designed, the loss function L of the generator G It consists of the following parts: in, represents the generated potential risk pattern, represents the probability of the discriminator judging the relationship between the generated risk pattern and the noise space, λ G is the regularization coefficient, is the gradient of the generator for the i-th feature in the topological feature matrix, is the variation of the topological feature at different time points during the operation, represents the discriminability of the potential risk patterns generated by the generator and the real surgical risk status; The encoder loss function L E It consists of the following parts: Among them, z ′ =E(R(H p )) represents the potential encoding of the encoder for the true surgical risk status, α E is the adjustment factor for encoder loss, represents the distance between the latent code z' and the noise vector z, γ i is the weight of the topological feature, represents the nonlinear mapping of the ith topological feature in the true surgical risk state, Represents the mapping of the generator to the same features; S36. Through the adversarial training of the bidirectional generative adversarial network, the generator G and the encoder E are optimized. The potential risk pattern generated by the generator and the potential representation generated by the encoder are close to the actual risk state. The discriminator D improves the ability to distinguish between the generated potential pattern and the actual surgical risk state. The adversarial loss function L Bi-GAN The comprehensive loss of the generator, encoder and discriminator is expressed as: in, represents the probability of the discriminator's judgment on the actual surgical risk state and the potential coding, represents the probability of the discriminator's judgment on the generated mode and noise, β is the weight coefficient of the gradient penalty term, which is used to control the responsiveness of the generator and encoder to changes in topological features. is the loss gradient of the generator, encoder and discriminator under the i-th topological feature, is the Frobenius norm.

6. The method for dynamic risk assessment of esophageal cancer surgery based on topological data analysis according to claim 1, characterized in that: The step S6 specifically includes: S61. Obtain multimodal medical data of patients in real time during risk assessment m,real-time , and update the topological model T(D m )’s topological characteristic matrix H p (t); S62. The bidirectional generator G updates the topological feature matrix H in real time by inputting p (t) Predict future risk trends Where Δt represents the predicted future time interval, z is the random noise vector used by the generator, The potential risk change trend in the future; S63. The bidirectional discriminator D receives the potential risk pattern generated during the actual surgery. and real-time risk data R(H p (t)), and provide risk warning information based on the identification results: Among them, D(R(H p (t))) represents the judgment of the discriminator on the current actual risk data, Represents the discriminator's judgment on the generated future risk change trend; S64. Based on the output result of the discriminator, when the system detects a potential risk, it generates corresponding risk warning information, predicts the degree and location of the risk change, and sends the risk warning information to the surgeon through the system; S65. Based on the real-time evaluation of the risk warning information system, the surgeon is provided with suggestions for adjusting the surgical path according to the risk level. The adjustment suggestions include the adjustment direction, adjustment range and expected risk reduction effect of the surgical path: Among them, w i is the weight adjusted for different paths, ΔP i represents the expected effect of each path adjustment on risk reduction.

7. The method for dynamic risk assessment of esophageal cancer surgery based on topological data analysis according to claim 1, characterized in that: The step S64 specifically includes: S641. When the discriminator D compares the real-time risk data R(H p (t)) and the resulting future risk change trend When a potential risk is detected, a risk warning message is generated, which includes the specific location where the risk occurs. i , Risk severity S i and the expected risk trend R i ; S642. The system sends the risk warning information to the surgeon in real time via wireless means, and the 3D visualization model of the surgical area marks the location where the risk occurs in combination with the risk warning information. i , and color-coded to show the severity of the risk S i , the darker the color, the greater the risk; S643. The system uses the time factor T in risk assessment f Predict the time t+Δt at which the future risk will occur, and write the timestamp T of the warning information f (t+Δt) Synchronize to the doctor interface: Among them, α is the adjustment coefficient of risk severity, S i The larger it is, the shorter Δt is, indicating a more urgent risk.

Images