Intelligent postoperative ileus diagnosis and intervention decision system for gastrointestinal tract and method thereof

Abstract

This invention discloses an intelligent diagnostic and intervention decision-making system and method for postoperative gastrointestinal obstruction, comprising the following steps: S1, collecting and preprocessing multi-source data from postoperative patients to construct an input dataset; S2, constructing a multi-scale visual Transformer model and extracting multi-scale image feature representations; S3, constructing a multi-modal state representation vector; S4, inputting the multi-modal state representation vector into an intervention strategy network to generate intervention decisions; S5, collecting feedback data, calculating reward values, and optimizing the strategy network; S6, processing the multi-modal state vectors into the optimized strategy network and outputting new intervention decisions; S7, constructing an image feature response evaluation function and supervising the optimization of the multi-scale visual Transformer. This invention achieves intelligent diagnosis and personalized intervention for postoperative intestinal obstruction by combining a multi-scale visual Transformer with trust region strategy optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image analysis and intelligent decision support, in particular to an intelligent gastrointestinal postoperative intestinal obstruction diagnosis and intervention decision system and method thereof. BACKGROUND

[0002] At present, gastrointestinal postoperative intestinal obstruction is one of the common and serious complications with high disability rate and mortality in digestive system surgery. With the popularization of minimally invasive surgery and fast-track surgery concepts, the number of gastrointestinal postoperative patients is increasing year by year, and higher requirements are put forward for early diagnosis and precise intervention of intestinal obstruction in clinic. Traditional intestinal obstruction diagnosis mainly relies on manual reading and interpretation of abdominal images (such as abdominal X-ray and CT) by clinicians, combined with comprehensive analysis of patient's vital signs, symptoms and behavior records. However, manual reading is highly subjective and depends on the experience of doctors, and it is difficult to deeply integrate massive and multi-modal data, especially in the early or atypical intestinal obstruction, which may lead to missed diagnosis, misdiagnosis and other problems, affecting the prognosis of patients.

[0003] With the rapid development of artificial intelligence and deep learning technology, some studies attempt to use convolutional neural networks and other methods to automatically extract features from abdominal images to assist in diagnosis, but most existing algorithms are only for single modality or single scale images, lack comprehensive modeling of multi-scale features of abdominal structures, and are difficult to handle complex and variable images and pathological manifestations in real clinical scenarios. At the same time, existing intelligent decision systems mostly use fixed strategy optimization algorithms, which are difficult to balance safety and individual differences. When facing dynamic feedback and abnormal high-risk events in the patient intervention process, the strategy adjustment is not flexible enough, which may lead to distorted intervention suggestions or insufficient safety.

[0004] In addition, existing systems generally lack consideration of reward mechanisms and rare high-risk events, lack dynamic supervision and multi-objective collaborative optimization mechanisms, and cannot realize deep integration of individualized feedback, complex risk states and multi-source data, so the accuracy and clinical usability of intelligent intervention decisions need to be improved.

[0005] Therefore, how to provide an intelligent gastrointestinal postoperative intestinal obstruction diagnosis and intervention decision system and method thereof is a problem that needs to be solved by those skilled in the art. SUMMARY

[0006] One purpose of the present application is to propose an intelligent postoperative intestinal obstruction diagnosis and intervention decision system and method for gastrointestinal tract, which fully integrates multi-scale feature extraction of abdominal images, multi-modal data fusion analysis of vital signs and postoperative behaviors, and introduces an individualized intervention decision network optimized by a reliance domain strategy. The whole process from multi-source data acquisition, preprocessing, feature fusion, intelligent intervention decision to dynamic feedback optimization is described in detail, and the present application has the advantages of high diagnostic accuracy, high intelligent intervention level, strong safety, strong response ability to rare high-risk events, and outstanding continuous self-adaptive optimization ability.

[0007] The intelligent postoperative intestinal obstruction diagnosis and intervention decision method for gastrointestinal tract according to the embodiment of the present application comprises the following steps:

[0008] S1, collecting multi-source data of a postoperative patient of gastrointestinal tract, preprocessing the multi-source data, and constructing an input data set;

[0009] S2, constructing a multi-scale visual Transformer model, extracting abdominal image data from the input data set, inputting the data into the multi-scale visual Transformer model, and extracting multi-scale image feature representation;

[0010] S3, fusing the multi-scale image feature representation with vital sign data and postoperative behavior record data in the input data set, and constructing a patient multi-modal state representation vector;

[0011] S4, inputting the patient multi-modal state representation vector into an intervention strategy network based on a reliance domain strategy optimization, setting a maximum KL divergence limit threshold, and generating an intervention behavior decision;

[0012] S5, implementing corresponding intervention measures on the patient according to the intervention behavior decision, collecting feedback data after the intervention is executed, calculating a strategy reward value between the intervention behavior and the result, using a reliance domain strategy optimization algorithm to iteratively update parameters of the intervention strategy network within a preset KL divergence constraint range, and obtaining an optimized intervention strategy network;

[0013] S6, processing the patient multi-modal state representation vector by using the optimized intervention strategy network, continuously outputting a new intervention behavior decision, and realizing a dynamic update and application cycle of the intervention strategy;

[0014] S7, constructing an image feature response evaluation function according to the parameters of the intervention strategy network and the feedback data, inputting the function as a supervision signal into the multi-scale visual Transformer model, and optimizing the multi-scale visual Transformer model.

[0015] Optionally, the multi-source data specifically includes heterogeneous information such as endoscopic images, abdominal imaging data, vital sign data, postoperative behavioral record data, physiological parameters, and electronic medical record texts of postoperative gastrointestinal patients.

[0016] Optionally, the preprocessing of multi-source data specifically includes image normalization, text segmentation and encoding, physiological parameter standardization, and data time alignment.

[0017] Optionally, S2 specifically includes:

[0018] S21. Extract abdominal image data of the abdominal region from the input dataset and construct the original image tensor. ,in, Indicates the image height. Indicates the image width. R represents the number of image channels, where R is the set of real numbers;

[0019] S22. The obtained original image tensor A multi-scale pyramid strategy was employed for downsampling to obtain multiple sets of abdominal image tensors with different spatial resolutions. ,in, Let represent the abdominal image tensor at the s-th spatial resolution. This represents the total number of scales of the ventral image tensor in a multi-scale pyramid structure.

[0020] S23. Tensors of abdominal images at each scale obtained Divide the window into sections of fixed size to obtain a set of non-overlapping windows. And within each window, a local self-attention mechanism is used to calculate the self-attention features within the window, where, This represents the tensor of the abdominal image at the s-th spatial resolution. The image sub-block obtained after dividing the i-th window region, Represents the lower abdominal image tensor at the s-th scale. The total number of windows that have been divided;

[0021] S24. For all window regions obtained at each scale, extract the feature representation of each window using a local self-attention mechanism, and concatenate the features of all windows according to the original spatial order to obtain the encoded feature map at each scale. ;

[0022] S25. Combine the encoded feature maps at all scales by concatenating channels and weighting the sums to obtain the fused multi-scale image feature representation.

[0023] S26. Represent the features of the fused multi-scale image. Input dual-branch collaborative architecture, respectively input to the structural decoding branch And auxiliary discrimination branch The structural decoding branch extracts abdominal structure feature information, and the auxiliary discrimination branch extracts intestinal obstruction risk discrimination related features;

[0024] S27, in the decoder of the structural decoding branch The hierarchical deconvolution and jump connection mechanism are used to perform multi-level up-sampling and spatial feature fusion on the abdominal structure feature information in the order of deep to shallow levels, and the abdominal structure prediction map is reconstructed;

[0025] S28, in the auxiliary discrimination branch The intestinal obstruction risk discrimination related features are processed by using full connection layer and activation function to output auxiliary risk discrimination results;

[0026] S29, further introduce sliding window local attention mechanism on multi-scale image feature representation Through sliding window covering each region of the feature map, local self-attention is calculated in each window;

[0027] S210, in the training process of the multi-scale visual Transformer model, the structural loss function and the functional loss function are combined by weighting, and the multi-scale visual Transformer model parameters are optimized by minimizing the combined loss function, the structural loss function is constructed by using the pixel-level cross entropy loss function, the functional loss function is constructed by using the binary classification cross entropy loss function, and finally the structural loss function and the functional loss function are linearly combined by setting the weighting coefficient to form the combined loss function;

[0028] S211, in the training process, the high-resolution feature branch is used to perform knowledge distillation on the low-resolution feature branch, and the KL divergence of the output distribution of the high and low resolution branches is minimized;

[0029] S212, after the training of the multi-scale visual Transformer model is completed, all parameter configurations of the multi-scale visual Transformer model are saved, the newly collected abdominal image data is input into the trained multi-scale visual Transformer model, and the final output multi-scale image feature representation is obtained.

[0030] Optionally, the S3 specifically comprises:

[0031] S31, extracting vital sign data and postoperative behavior record data at the same time point as the abdominal image data from the input data set, the vital sign data including time series monitoring parameters of body temperature, heart rate, respiratory rate, blood pressure and blood oxygen saturation, and the postoperative behavior record data including key behavior time series of patient exhaust, defecation, eating and activity;

[0032] S32, normalizing and time series aligning the extracted vital sign data and postoperative behavior record data to construct a multi-modal structured feature tensor , wherein, represents the number of time steps, represents the number of clinical feature dimensions, is a real set;

[0033] S33, performing time series index alignment on the obtained multi-scale image feature representation , which corresponds to the image feature and the clinical data at the same time step;

[0034] S34, performing feature-level fusion on the aligned multi-scale image feature representation and the obtained multi-modal structured feature tensor to obtain a patient multi-modal state representation vector in a splicing manner ;

[0035] S35, batch arranging the patient multi-modal state representation vectors at all time steps to form patient multi-modal state representation vectors for intervention strategy network input.

[0036] Optionally, the S4 specifically comprises:

[0037] S41, inputting the obtained patient multi-modal state representation vector to the intervention strategy network based on the reliance domain strategy optimization as the current input state;

[0038] S42, setting an intervention behavior set in the intervention strategy network , wherein, represents the th intervention behavior, is the total number of intervention behaviors;

[0039] S43, initializing the intervention strategy network parameters , inputting the patient multi-modal state representation vector into the intervention strategy network through a deep neural network, and obtaining a probability distribution result of each intervention behavior through forward calculation of the intervention strategy network;

[0040] S44, setting a reliance domain constraint to limit the maximum KL divergence threshold of each strategy distribution update ;

[0041] S45, in the process of policy optimization, embedding medical prior constraints, integrating clinical medical prior knowledge such as medical guidelines, surgical contraindications and expert experience into the policy optimization process in the form of constraint terms, and reducing the weight in the probability distribution of the intervention strategy network for intervention behaviors that do not meet medical standards or have significant medical risks;

[0042] S46, using a dynamic self-adaptive mechanism of trust region, according to the signal function dynamic adjustment of KL divergence threshold :

[0043] ;

[0044] S47, introducing a multi-objective hierarchical trust region optimization mechanism, refining the intervention target into multiple hierarchical targets in the process of intervention strategy network optimization, including maximizing efficacy, minimizing safety risk and optimizing medical resources, establishing independent trust region constraints for each target, and balancing the strategy output under different targets through hierarchical aggregation mechanism;

[0045] S48, in the training and optimization process of the intervention strategy network, defining the event samples with historical occurrence probability lower than the preset threshold and the corresponding risk score higher than the preset threshold as rare high-risk event samples, setting an experience memory pool for rare high-risk event samples, and increasing the sampling weight of rare high-risk event samples in the training samples, and applying more strict trust region constraints to the training batches involving rare high-risk events in the process of policy update to strengthen the convergence constraints;

[0046] S49, collecting feedback after the patient intervention is executed, and constructing a reward function :

[0047] ;

[0048] wherein, the clinical feedback reward mapping function is used to dynamically adjust the reward weight according to the individual response of the patient, and the reward signal is divided into different levels of short-term effect, long-term outcome and risk punishment; at the same time, a feedback traceability explanation mechanism is introduced in the reward signal generation process, the key contribution factors of the reward signal are labeled and traceability analysis is performed, the individuality, dynamics and explainability of the reward feedback are improved, and more rich feedback information support is provided for the continuous optimization and precise adjustment of the intervention strategy network;

[0049] S410, outputting the intervention behavior decision generated by the optimized intervention strategy network, applying the intervention behavior decision to the current actual intervention of the patient, and collecting closed-loop effect feedback and dynamically updating the strategy.

[0050] Optionally, the S5 specifically comprises:

[0051] S51, according to the intervention behavior decision output by the intervention strategy network, implementing corresponding medical intervention measures on the patient, the medical intervention measures including conservative observation, diet control, drug treatment, gastrointestinal decompression, image review or surgical treatment;

[0052] S52, collecting feedback data after each intervention measure is implemented, the feedback data including symptom performance, vital sign parameters, laboratory examination indexes and adverse event information of the patient, forming a feedback data set ;

[0053] S53, preprocessing the feedback data set to obtain an intervention effect change feature vector , and constructing a comprehensive reward function according to the intervention effect change feature vector :

[0054] ;

[0055] Wherein, is a clinical feedback reward mapping function;

[0056] S54, inputting the comprehensive reward function and a rare high-risk event indication function as feedback signals into the intervention strategy network, the rare high-risk event indication function being used to determine whether the patient has an event after the intervention, the event having a historical occurrence probability lower than a preset threshold and a risk score higher than a preset threshold, when the feedback data includes such an event, the rare high-risk event indication function takes a value of 1, otherwise, it takes a value of 0, and the strategy is optimized in combination with the current state-action pair , wherein, represents a multi-modal state representation vector of the patient at the t-th time step, represents an intervention behavior decision selected and executed at the t-th time step;

[0057] S55, parameter updating is performed by using a trust region strategy optimization algorithm, and the optimization target is ;

[0058] S56, continuously collecting and archiving each intervention and its feedback data, and dynamically adjusting the reward function parameters and the rare event weight parameters;

[0059] S57, the above steps S51 to S56 are repeatedly executed to realize continuous iterative updating of the intervention strategy network based on the reward mechanism and dynamic strategy optimization.

[0060] Optionally, the S7 specifically comprises:

[0061] S71, constructing an image feature response evaluation function according to the parameters of the latest intervention strategy network and the feedback data after the patient intervention is performed, in combination with the current fused multi-scale image feature representation

[0062] S72, inputting the image feature response evaluation function as a supervision signal into the multi-scale visual Transformer model as the basis for adjusting the attention distribution weight, and optimizing the attention weight parameters of the multi-scale visual Transformer model at different spatial scales;

[0063] S73, in the optimization process of the multi-scale visual Transformer model, iteratively updating the attention distribution weight of each scale Transformer encoding layer, and the multi-scale visual Transformer model can adaptively strengthen the feature extraction ability of the key region highly related to the feedback when processing new input abdominal image data;

[0064] S74, saving the optimized multi-scale visual Transformer model parameters to form a multi-scale visual Transformer model that can be continuously and adaptively updated, and continuously using the image feature response evaluation function and the attention guiding mechanism in the abdominal image data analysis process to realize dynamic supervision optimization and individualized feature extraction ability improvement of the multi-scale visual Transformer model.

[0065] The intelligent gastrointestinal postoperative intestinal obstruction diagnosis and intervention decision system according to the embodiment of the application comprises the following modules:

[0066] A data acquisition and preprocessing module is configured to acquire multi-source data of a patient after gastrointestinal surgery and perform preprocessing to construct an input data set;

[0067] A multi-scale visual Transformer feature extraction module is configured to extract a multi-scale image feature representation from the input data set;

[0068] A multi-modal state fusion module is configured to fuse the multi-scale image feature representation with the input data set to generate a multi-modal state representation vector;

[0069] An intervention strategy network module is configured to output a patient intervention behavior decision based on a trust region policy optimization algorithm in combination with the multi-modal state representation vector;

[0070] A reward and rare event discrimination module is configured to construct a reward function according to feedback data after intervention is performed, and to discriminate whether a rare high-risk event occurs to generate an intervention strategy network for policy optimization;

[0071] ​​The model supervision optimization module is configured to construct an image feature response evaluation function according to the intervention strategy network parameters and the feedback data, and use the function as a supervision signal to guide dynamic optimization of the multi-scale visual Transformer model.

[0072] The present application has the following advantages:

[0073] The present application realizes intelligent diagnosis and individualized intervention decision for postoperative intestinal obstruction of the gastrointestinal tract by deeply integrating the multi-scale visual Transformer model with the intervention strategy network optimized by the trust region strategy. Compared with the prior art, the present application can extract multi-scale and multi-level structural features from abdominal image data, and combine patient vital signs, behavior records and other multi-modal clinical information to improve the recognition ability of the model for early, complex and atypical intestinal obstruction cases. By introducing the trust region strategy optimization algorithm, the present application can not only output safe and controllable, dynamic and adaptive individualized intervention suggestions, but also adjust the strategy parameters in real time according to the clinical feedback of the patient, thereby ensuring the flexibility and efficiency of the decision-making process. Further, the present application constructs a comprehensive reward mechanism and a rare high-risk event weighting mechanism, so that the intervention strategy network can respond and optimize to extreme risk states, thereby significantly improving the processing ability for abnormal and rare clinical events.

[0074] The system of the present application has the characteristics of continuous adaptive optimization, multi-objective collaborative management and traceable decision-making process, which not only improves the accuracy and scientificity of diagnosis and intervention, but also enhances the safety, individualization and stability of the intelligent medical system in actual clinical application, thereby providing more intelligent, precise and efficient diagnosis and treatment services for postoperative patients of the gastrointestinal tract, and having significant clinical application promotion value and social and economic benefits. BRIEF DESCRIPTION OF DRAWINGS

[0075] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application and explain the principles of the present application, and do not constitute a limitation of the present application. In the drawings:

[0076] Figure 1 A flowchart of the intelligent postoperative intestinal obstruction diagnosis and intervention decision method of the gastrointestinal tract according to the present application;

[0077] Figure 2 A structural schematic diagram of the intelligent postoperative intestinal obstruction diagnosis and intervention decision system of the gastrointestinal tract according to the present application. DETAILED DESCRIPTION

[0078] The present application will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0079] ReferenceFigure 1 An intelligent post-gastrointestinal surgery intestinal obstruction diagnosis and intervention decision-making method, comprising the following steps:

[0080] S1, collecting multi-source data of post-gastrointestinal surgery patients, preprocessing the multi-source data, and constructing an input data set;

[0081] S2, constructing a multi-scale visual Transformer model, extracting abdominal image data from the input data set, inputting the data into the multi-scale visual Transformer model, and extracting multi-scale image feature representation;

[0082] S3, fusing the multi-scale image feature representation with the vital sign data and postoperative behavior record data in the input data set to construct a patient multi-modal state representation vector;

[0083] S4, inputting the patient's multi-modal state representation vector into an intervention strategy network based on a trust region policy optimization, setting a maximum KL divergence limit threshold, and generating an intervention behavior decision;

[0084] S5, implementing corresponding intervention measures on the patient according to the intervention behavior decision, collecting feedback data after the intervention is executed, calculating the strategy reward value between the intervention behavior and the result, and using the trust region policy optimization algorithm to iteratively update the parameters of the intervention strategy network within the preset KL divergence constraint range, to obtain an optimized intervention strategy network;

[0085] S6, processing the patient's multi-modal state representation vector with the optimized intervention strategy network, continuously outputting new intervention behavior decisions, and realizing the dynamic update and application cycle of the intervention strategy;

[0086] S7, constructing an image feature response evaluation function according to the parameters of the intervention strategy network and the feedback data, as a supervision signal input into the multi-scale visual Transformer model, and optimizing the multi-scale visual Transformer model.

[0087] This invention achieves intelligent and refined management of the entire process of diagnosis and intervention for postoperative intestinal obstruction after gastrointestinal surgery. It employs multi-source data acquisition and preprocessing to ensure the comprehensiveness and quality of patient information; utilizes a multi-scale visual Transformer model to extract multi-level features from abdominal images, effectively improving the recognition ability of complex structural regions; and accurately constructs a multimodal representation reflecting the patient's overall condition by fusing image features with vital signs and behavioral data. An intervention strategy network based on trust region optimization enables safe and controllable individualized intervention decisions, and dynamic KL divergence constraints ensure the stability and reliability of the optimization process. The system continuously collects feedback data after intervention execution, utilizing reward mechanisms and adaptive strategy updates to achieve closed-loop decision-making and continuous optimization. Furthermore, feedback-driven image feature response evaluation and Transformer model optimization mechanisms further enhance the model's sensitivity to key clinical information and feature extraction capabilities. Overall, this invention significantly improves the accuracy of intestinal obstruction diagnosis and the level of individualized intervention, enhances the system's response to abnormal events and adaptive optimization capabilities, and has significant clinical application value and promising prospects for wider application.

[0088] In this embodiment, the multi-source data specifically includes heterogeneous information such as endoscopic images, abdominal imaging data, vital sign data, postoperative behavioral record data, physiological parameters, and electronic medical record text of postoperative gastrointestinal patients.

[0089] In this embodiment, the preprocessing of multi-source data specifically includes image normalization, text segmentation and encoding, physiological parameter standardization, and data time alignment. The image normalization, text segmentation and encoding, physiological parameter standardization, and data time alignment refer to normalizing the pixel value distribution of the image, performing word segmentation and vector encoding on the text to extract semantic features, standardizing the physiological parameters to eliminate dimensional differences, and aligning various types of data according to timestamps.

[0090] In this embodiment, S2 specifically includes:

[0091] S21. Extract abdominal image data of the abdominal region from the input dataset and construct the original image tensor. ,in, Indicates the image height. Indicates the image width. R represents the number of image channels, where R is the set of real numbers.

[0092] S22. The obtained original image tensor A multi-scale pyramid strategy was employed for downsampling to obtain multiple sets of abdominal image tensors with different spatial resolutions. Specifically, abdominal image data is downsampled according to different spatial resolutions to obtain multiple abdominal image tensors with different spatial resolutions. Each image tensor at each resolution retains the original structural information of the abdomen. The abdominal image tensors at all scales together form a multi-scale pyramid structure. Let represent the abdominal image tensor at the s-th spatial resolution. This represents the total number of scales of the ventral image tensor in a multi-scale pyramid structure.

[0093] S23. Tensors of abdominal images at each scale obtained Divide the window into sections of fixed size to obtain a set of non-overlapping windows. And within each window, a local self-attention mechanism is used to calculate the self-attention features within the window, where, This represents the tensor of the abdominal image at the s-th spatial resolution. The image sub-block obtained after dividing the i-th window region, Represents the lower abdominal image tensor at the s-th scale. The total number of windows that have been divided;

[0094] S24. For all window regions obtained at each scale, extract the feature representation of each window using a local self-attention mechanism, and concatenate the features of all windows according to the original spatial order to obtain the encoded feature map at each scale. The encoded feature map fully reflects the spatial structure and fine-grained information of the abdominal image at scale;

[0095] S25. The encoded feature maps at all scales are fused by concatenating channels and weighting the sums to obtain a fused multi-scale image feature representation. The multi-scale image feature representation comprehensively reflects the structural feature information of the abdominal image at different spatial resolutions.

[0096] S26. Represent the features of the fused multi-scale image. Input the dual-branch collaborative architecture, and input them into the structure decoding branch respectively. and auxiliary decision branches The structural decoding branch extracts abdominal structural feature information, and the auxiliary discrimination branch extracts relevant features for intestinal obstruction risk discrimination.

[0097] S27, in the structural decoding branch In the decoder, hierarchical deconvolution and skip connection mechanism are adopted to perform multi-level up-sampling and spatial feature fusion on the abdominal structure feature information in the order from deep to shallow, and an abdominal structure prediction map is obtained by reconstruction, wherein the hierarchical deconvolution and skip connection mechanism refers to gradually up-sampling and fusing spatial information of abdominal structure features at different levels through hierarchical deconvolution and skip connection mechanism, so as to realize high-precision reconstruction of abdominal structure.

[0098] In the auxiliary discrimination branch S28, in the auxiliary discrimination branch

[0099] S29, in the multi-scale image feature representation Further introduce a sliding window local attention mechanism, calculate the local self-attention in each window by covering each region of the feature map through the sliding window.

[0100] S210, in the multi-scale visual Transformer model training process, a structure loss function and a function loss function are combined by weighting, and the multi-scale visual Transformer model parameters are optimized by minimizing the combined joint loss function, the structure loss function is constructed by using the pixel-level cross entropy loss function, the function loss function is constructed by using the binary classification cross entropy loss function, and finally the structure loss function and the function loss function are linearly combined by setting the weighting coefficient to form the joint loss function.

[0101] S211, in the training process, the high-resolution feature branch is used to perform knowledge distillation on the low-resolution feature branch, and the KL divergence of the output distribution of the high and low resolution branches is minimized.

[0102] S212, after the multi-scale visual Transformer model is trained, save all parameter configurations of the multi-scale visual Transformer model, input the newly collected abdominal image data into the trained multi-scale visual Transformer model, and obtain the final output multi-scale image feature representation.

[0103] The application realizes multi-scale and multi-level structured feature extraction and discriminant analysis of abdominal images through the defined multi-scale visual Transformer model construction and optimization process. Through multi-scale pyramid downsampling, local self-attention mechanism and channel fusion operation, global and fine-grained structural features of abdominal images at different spatial resolutions can be accurately captured, greatly improving the recognition ability of the model for complex abdominal structures and small lesions. The double-branch collaborative architecture enables synchronous and efficient structure decoding and risk discrimination, ensuring the accuracy of structure reconstruction and enhancing the sensitivity to high-risk areas of intestinal obstruction. The use of sliding window local attention mechanism and knowledge distillation effectively improves the attention of the model to key areas and the information fusion efficiency between high and low resolution features. The weighted joint optimization of structure loss and function loss ensures the comprehensive performance of the model in structure reconstruction and risk discrimination tasks. Overall, the application significantly improves the accuracy, robustness and model generalization ability of intelligent analysis of abdominal images, providing a solid feature foundation and reliable support for subsequent intelligent diagnosis and intervention.

[0104] In the embodiment, the S3 specifically includes:

[0105] S31, extracting vital sign data and postoperative behavior record data at the same time point as the abdominal image data from the input data set, the vital sign data including time sequence monitoring parameters of body temperature, heart rate, respiratory rate, blood pressure and blood oxygen saturation, and the postoperative behavior record data including key behavior time sequences of patient exhaust, defecation, eating and activity;

[0106] S32, normalizing and time sequence aligning the extracted vital sign data and postoperative behavior record data to construct a multi-modal structured feature tensor , wherein, represents the number of time steps, represents the number of clinical feature dimensions, is a real set, and the normalizing and time sequence aligning refers to standardizing clinical data of different dimensions to a unified numerical range and sorting according to time stamps, so that each type of data corresponds synchronously at the same time step;

[0107] S33, performing time sequence index alignment on the obtained multi-scale image feature representation , for the image feature and the clinical data corresponding at the same time step, the time sequence index alignment refers to one-to-one matching the multi-scale image feature with the clinical data at the corresponding time step according to the time stamp, to ensure that the multi-modal information at the same time point is aligned in the feature sequence;

[0108] S34, performing feature-level fusion on the aligned multi-scale image feature representation and the obtained multi-modal structured feature tensor to obtain a patient multi-modal state representation vector in a splicing manner ;

[0109] S35, batch collate the patient multi-modal state representation vectors of all time steps to form the patient multi-modal state representation vectors for intervention strategy network input, which refers to uniformly encoding the multi-modal feature vectors of patient physiology, behavior and environment generated at each time step in chronological order and combining them into a two-dimensional batch tensor.

[0110] The present application realizes precise modeling of the overall and dynamic state of postoperative patients in the gastrointestinal tract through the extraction and fusion process of multi-modal time sequence features. Through the vital signs and postoperative behavior data extracted at the same time point as the abdominal image data, the physiological changes and behavior characteristics of the patient at different periods after the operation can be fully reflected. Normalization and time sequence alignment operations ensure the synchronicity and comparability of multi-modal data, significantly reducing the information loss caused by data heterogeneity. By fusing multi-scale image features and structured clinical data, the patient multi-modal state representation is constructed using feature splicing, which not only comprehensively understands the information at the anatomical, functional and behavioral levels, but also enhances the model's ability to capture individual differences and dynamic changes. Finally, the batch collated patient multi-modal state representation sequence provides a high-quality and time-sequenced comprehensive representation basis for the input of the subsequent intelligent intervention strategy network, greatly improving the model's representation ability of the patient's complex state and the scientific nature of intelligent decision-making, and laying a solid data foundation for individualized precise intervention and risk prediction.

[0111] In the embodiment, the S4 specifically comprises:

[0112] S41, collate the obtained patient multi-modal state representation vectors as the current input state into the intervention strategy network based on the trust domain policy optimization;

[0113] S42, set an intervention behavior set in the intervention strategy network , wherein, represents the th intervention behavior, is the total number of intervention behaviors;

[0114] S43, initialize the intervention strategy network parameters , input the patient multi-modal state representation vector into the intervention strategy network using a deep neural network, and obtain the probability distribution result of each intervention behavior through forward calculation of the intervention strategy network;

[0115] S44, set a trust domain constraint to limit the maximum KL divergence threshold of each strategy distribution update ;

[0116] S45, in the process of policy optimization, embedding medical prior constraints, integrating clinical medical prior knowledge such as medical guidelines, surgical contraindications and expert experience into the policy optimization process in the form of constraint terms, and reducing the weight in the probability distribution of intervention strategy network for the intervention behavior that does not meet the medical specification or has major medical risk;

[0117] S46, adopting a dynamic self-adaptive mechanism of confidence domain, according to the signal function dynamically adjusting the KL divergence threshold :

[0118] ;

[0119] S47, introducing a multi-objective hierarchical confidence domain optimization mechanism, refining the intervention target into multiple hierarchical targets in the process of intervention strategy network optimization, including maximizing efficacy, minimizing safety risk and optimizing medical resources, establishing independent confidence domain constraints for each target respectively, and balancing the strategy output under different targets through hierarchical aggregation mechanism, wherein the maximizing efficacy, minimizing safety risk and optimizing medical resources specifically refer to maximizing efficacy by quantifying the improvement degree of key indicators after intervention, controlling safety risk by evaluating the probability and severity of potential side effects, and optimizing resource allocation efficiency by combining treatment cost and resource consumption rate;

[0120] S48, in the training and optimization process of intervention strategy network, defining the event samples with historical occurrence probability lower than a preset threshold and corresponding risk score higher than a preset threshold as rare high-risk event samples, setting an experience memory pool for rare high-risk event samples, and increasing the sampling weight of rare high-risk event samples in the training samples, applying more strict confidence domain constraints to the training batches involving rare high-risk events in the process of policy update to strengthen the convergence constraints;

[0121] S49, collecting feedback after patient intervention execution, and constructing a reward function :

[0122] ;

[0123] wherein, represents a clinical feedback reward mapping function, dynamically adjusts the reward weight according to the individual response of the patient, and divides the reward signal into different levels of short-term effect, long-term outcome and risk punishment; at the same time, a feedback traceability explanation mechanism is introduced in the reward signal generation process, the key contribution factors of the reward signal are labeled and traceability analysis is performed, the individuality, dynamics and explainability of the reward feedback are improved, and more rich feedback information support is provided for the continuous optimization and precise adjustment of the intervention strategy network.

[0124] S410, output the intervention behavior decision made by the optimized intervention strategy network, apply the intervention behavior decision to the current actual intervention of the patient, and collect closed-loop effect feedback and dynamically update the strategy.

[0125] The present application significantly improves the scientificity and individualization level of postoperative intestinal obstruction intervention decision of gastrointestinal tract by constructing and optimizing the intervention strategy network based on the trust region strategy optimization. By introducing multi-modal state representation as the input of the strategy network, the dynamic changes of the patient can be fully reflected, providing rich state basis for the intervention strategy. The intervention behavior set and the deep neural network output behavior probability distribution are set to realize the intelligent exploration of the high-dimensional and variable medical decision space. The trust region constraint and the dynamic adaptive mechanism ensure the safety and stability of the strategy optimization process, preventing the model from outputting extreme intervention suggestions. The fusion of medical prior constraints and multi-objective hierarchical trust region mechanism enables the decision network to balance risk prevention and resource allocation while pursuing efficacy, achieving multi-dimensional target collaborative optimization. The experience memory and reinforcement constraint of rare high-risk event samples improve the model's response and safety protection capability to abnormal clinical conditions. Through the multi-level reward function design and feedback tracing mechanism, the intervention strategy network can continuously adapt and optimize, significantly improving the individualization, interpretability and closed-loop intelligent optimization capability of the intervention suggestions. Overall, the present application can greatly improve the intelligent, precise and clinical usability of postoperative intestinal obstruction management.

[0126] In the present embodiment, the S5 specifically comprises:

[0127] S51, according to the intervention behavior decision output by the intervention strategy network, implementing corresponding medical intervention measures on the patient, the medical intervention measures including conservative observation, diet control, drug treatment, gastrointestinal decompression, image review or surgical treatment;

[0128] S52, collecting feedback data after each intervention measure is implemented, the feedback data including the patient's symptom performance, vital sign parameters, laboratory examination indexes and adverse event information, forming a feedback data set ;

[0129] S53, preprocessing the feedback data set to obtain an intervention effect change feature vector , and constructing a comprehensive reward function according to the intervention effect change feature vector , the preprocessing of the feedback data set refers to normalizing, time aligning and denoising the post-intervention patient physiological parameters, image indexes and subjective score data collected in the feedback data set:

[0130] ;

[0131] wherein, is a clinical feedback reward mapping function;

[0132] S54, the comprehensive reward function and a rare high-risk event indication function are input into the intervention strategy network as feedback signals, the rare high-risk event indication function is used to determine whether an event with a historical occurrence probability lower than a preset threshold and a risk score higher than a preset threshold occurs after the intervention, when such an event occurs in the feedback data, the rare high-risk event indication function takes a value of 1, otherwise it takes a value of 0, and the current state-action pair is optimized, wherein, represents a multi-modal state representation vector of the patient at the t-th time step, represents the intervention behavior decision selected and executed at the t-th time step;

[0133] S55, a trust region strategy optimization algorithm is used for parameter updating, and the optimization target is :

[0134] ;

[0135] and a dynamic KL divergence constraint is satisfied:

[0136] ;

[0137] wherein, represents the current intervention strategy network parameter , for the multi-modal state representation vector of the patient, the probability of selecting the intervention behavior decision , represents the probability of the same state-action pair before the last update of the intervention strategy network parameter , is a reward signal weighting coefficient, is a rare event reward weighting coefficient, is an expectation operator, represents the corresponding parameter when the objective function takes the maximum value, is a KL divergence threshold, represents the KL divergence;

[0138] In the present application, the optimization target The formula is essentially used to guide the intervention strategy network to continuously improve the scientificity and safety of clinical decision-making. The practical significance is that after receiving new patient feedback and reward signals each time, the model will comprehensively consider the intervention effect and the response of rare high-risk events as the measurement standard, automatically adjust the intervention strategy, and make the subsequent output decision-making scheme more in line with the individual needs of patients and the requirements of clinical risk prevention and control. In the optimization process, the model will also automatically strictly constrain the change amplitude between the new and old strategy distributions, ensuring that each adjustment of the intervention strategy will not have drastic fluctuations or uncontrollable phenomena, thereby ensuring the stability and clinical safety of medical decision-making. Through the above mechanism, the model not only can continuously accumulate experience, strengthen the learning and protection of rare high-risk events, but also can dynamically adapt to the feedback of different patients, realize individualization, precision and high safety of intelligent intervention. This optimization method provides a solid guarantee for realizing the clinical practicability and wide application of the intelligent intervention system for postoperative intestinal obstruction of the gastrointestinal tract.

[0139] S56, continuously collecting and archiving each intervention and its feedback data, dynamically adjusting the reward function parameters and the rare event weight parameters;

[0140] S57, the above steps S51 to S56 are repeatedly executed to realize the continuous iterative update of the intervention strategy network based on the reward mechanism and dynamic strategy optimization.

[0141] The intervention decision and continuous optimization process provided by the present application realizes a scientific closed loop of intelligent intervention management for postoperative intestinal obstruction of the gastrointestinal tract. By applying the decision output by the intervention strategy network to the patient and collecting multi-dimensional feedback data including symptom performance, vital signs and adverse events, the actual effect of the intervention measures can be fully reflected. The intervention effect is quantified by using a comprehensive reward function, and combined with a rare high-risk event indicator function, the model ensures that rare and high-risk clinical events are responded to and learned intensively. The network parameter is updated safely and iteratively by using the dynamic KL divergence constraint trust region strategy optimization algorithm, which effectively avoids the risk of extreme strategy. The feedback after each intervention is continuously input into the model, which not only dynamically adjusts the reward function parameters and the rare event weight, but also realizes the individual evolution and risk adaptive adjustment of the decision system. Overall, the present application significantly improves the accuracy, robustness and risk prevention and control ability of intelligent intervention decision-making, ensures the individualization, scientificity and clinical safety of the patient intervention process, and provides strong support for intelligent medical closed loop management and precision treatment.

[0142] In the embodiment, the S6 specifically comprises:

[0143] The optimized intervention strategy network is used to process the current multi-modal state representation vector of the patient, combined with multi-dimensional information such as patient condition, vital signs, postoperative recovery indicators, etc. Real-time inference and generation of personalized intervention behavior decisions. These decisions include but are not limited to drug administration recommendations, dietary intake control, exercise intervention plans, and clinical review reminders. With the continuous change of the patient's state, the system dynamically inputs the updated state information, continuously iterates and optimizes the intervention strategy, forming a "perception-decision-feedback-re-decision" cyclic closed-loop control mechanism, so as to realize the precise intervention and continuous management of postoperative intestinal obstruction risk.

[0144] In the embodiment, the S7 specifically includes:

[0145] S71, according to the latest intervention strategy network parameters and the feedback data after the patient intervention is executed, combined with the current fused multi-scale image feature representation , construct an image feature response evaluation function ;

[0146] S72, input the image feature response evaluation function into the multi-scale visual Transformer model as a supervision signal, as the basis for adjusting the attention distribution weight, and optimize the attention weight parameters of the multi-scale visual Transformer model at different spatial scales;

[0147] S73, in the optimization process of the multi-scale visual Transformer model, iteratively update the attention distribution weight of each scale Transformer encoding layer, and the multi-scale visual Transformer model can adaptively strengthen the feature extraction ability of the key region highly related to the feedback when processing new input abdominal image data;

[0148] S74, save the optimized multi-scale visual Transformer model parameters to form a multi-scale visual Transformer model that can be continuously and adaptively updated, and continuously use the image feature response evaluation function and the attention guiding mechanism in the abdominal image data analysis process, realize the dynamic supervision optimization of the multi-scale visual Transformer model and the improvement of the individualized feature extraction ability.

[0149] The application significantly improves the sensitivity and adaptive ability of the model to the key areas of abdominal images and clinical feedback through the dynamic supervised optimization process of the defined multi-scale visual Transformer model. By combining the latest intervention strategy network parameters and patient feedback data, an image feature response evaluation function is constructed to realize effective closed-loop between model feature extraction and actual intervention effect. The evaluation value is used as a supervision signal to guide the attention distribution weight adjustment, so that the multi-scale visual Transformer model can focus on the areas most related to the clinical results at different spatial scales, improving the recognition ability of abnormal structures and early lesions. Through iterative optimization of attention distribution, the model realizes adaptive adjustment to new input images and continuous enhancement of feature extraction ability. Finally, the model parameter saving and updating mechanism ensures that the multi-scale visual Transformer model has long-term sustainable evolution and individualized feature learning ability. Overall, the application improves the intelligent, precise and dynamic response ability of abdominal image analysis, providing a solid technical foundation and clinical support for early identification, risk stratification and intervention effect evaluation of intestinal obstruction.

[0150] Reference Figure 2 , an intelligent gastrointestinal postoperative intestinal obstruction diagnosis and intervention decision system, comprising the following modules:

[0151] A data acquisition and preprocessing module is used to acquire multi-source data of patients after gastrointestinal surgery and perform preprocessing to construct an input data set.

[0152] A multi-scale visual Transformer feature extraction module is used to extract multi-scale image feature representations from the input data set.

[0153] A multi-modal state fusion module is used to fuse the multi-scale image feature representations with the input data set to generate a multi-modal state representation vector.

[0154] An intervention strategy network module is used to output patient intervention behavior decisions based on a trust region policy optimization algorithm and the multi-modal state representation vector.

[0155] A reward and rare event discrimination module is used to construct a reward function according to feedback data after intervention execution, and to discriminate whether a rare high-risk event has occurred to generate an intervention strategy network for policy optimization.

[0156] A model supervised optimization module is used to construct an image feature response evaluation function according to the intervention strategy network parameters and feedback data, and to use it as a supervision signal to guide the dynamic optimization of the multi-scale visual Transformer model.

[0157] Embodiment 1:

[0158] To verify the feasibility of the present application in implementation, the present application is applied to a general surgery department of a large three-level first-class hospital. From May 2024 to May 2025, the hospital fully deployed and actually applied the intelligent gastrointestinal postoperative intestinal obstruction diagnosis and intervention decision system of the present application, and the system runs through the entire postoperative hospitalization management process of gastrointestinal surgery patients. The hospital has more than 2000 gastrointestinal surgeries per year, with a large patient base and complex cases, making it an advanced application scenario in this field.

[0159] Taking a 65-year-old male patient with colorectal cancer after surgery as an example, the patient developed mild abdominal distension and reduced food intake on the third day after surgery. In the past, doctors relied on manual judgment, which required repeated physical examinations, review of paper or scattered electronic records, and waiting for image consultations. The average diagnosis and intervention decision-making cycle was 8-12 hours. Today, the system automatically collects the patient's abdominal CT, B-ultrasound, body temperature, heart rate, respiration, blood pressure, blood oxygen, diet, and activity data. The data preprocessing module realizes automatic standardization and time alignment. The multi-scale visual Transformer feature extraction module models the multi-level information of the abdominal image, and combines the multi-modal state fusion module to generate individual state representation. The intervention strategy network assesses the patient's intestinal obstruction risk in real time, and the system suggests that the doctor "restricts diet and closely monitors" and continuously optimizes the intervention plan according to feedback.

[0160] When the patient's abdominal distension worsens significantly, body temperature rises, and abdominal imaging shows intestinal expansion and gas-liquid plane on the fifth day after surgery, the system automatically identifies it as a high-risk state and immediately pushes the "gastrointestinal decompression and emergency intravenous fluid infusion" suggestion. After the intervention is executed, the patient's symptoms improve significantly, and no intestinal perforation or emergency surgery occurs. Compared with before the system was enabled, the doctor said, "The condition is discovered earlier, the treatment is faster, and the decision is based on evidence, which greatly reduces the on-duty pressure and medical risk." During the system operation, the hospital statistically analyzed the early identification rate of intestinal obstruction, the rate of severe cases, the average length of stay, and the incidence of adverse events of gastrointestinal postoperative patients in 2023 (before the system was applied) and 2024 (after the system was applied).

[0161] Table 1 Comparison of key indicators of gastrointestinal postoperative intestinal obstruction patients before and after the application of the intelligent system

[0162]

[0163] According to the data statistics results of Table 1, it can be seen that the intelligent gastrointestinal postoperative intestinal obstruction diagnosis and intervention system of the application has achieved remarkable clinical results after practical application. In 2023 (system not applied) and 2024 (system applied), the number of patient cases is comparable, both close to two thousand cases, so the data has strong comparability. After the application of the system, the early missed diagnosis rate of intestinal obstruction decreased from 9.0% to 2.1%, effectively avoiding the delay of the disease due to the untimely early identification. The rate of severe cases also decreased from 4.6% to 1.2%, indicating that through intelligent risk assessment and dynamic intervention, the risk of patients turning into severe cases can be significantly reduced.

[0164] In terms of average hospitalization days, it was shortened from 13.1 days to 9.5 days, fully embodying the optimization effect of the system on intervention opportunity and individualized management, helping patients recover earlier and saving medical resources. The incidence of adverse events after intervention decreased from 2.3% to 0.7%, further proving that intelligent intervention recommendations are safer and can reduce complications and secondary damage. In addition, the proportion of doctors adopting system recommendations is as high as 92%, showing that the system recommendations are scientific and have very high clinical recognition. It is particularly worth noting that the early warning rate of rare high-risk events increased from 35% to 100%, greatly improving the protection ability for critically ill patients. Patient satisfaction also increased from 88% to 97%, indicating that the system has optimized the medical experience and treatment confidence.

[0165] Overall, these data clearly prove that the application can improve the diagnosis and treatment quality and safety of gastrointestinal postoperative patients, not only reducing missed diagnosis, severe outcome and complications, but also optimizing the hospitalization process, obtaining the high recognition of doctors and patients, and showing strong clinical application value and promotion prospects.

[0166] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, any person skilled in the art can make equivalent replacement or change according to the technical scheme and inventive concept of the application within the technical range disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. An intelligent postoperative ileus diagnosis and intervention decision system for gastrointestinal tract, characterized in that, Comprise the following modules: Data acquisition and preprocessing module, for collecting postoperative gastrointestinal tract of patients with multiple source data and preprocessing, build input data set; Multi-scale visual Transformer feature extraction module, for extracting multi-scale image feature representation from the input data set; Multi-modal state fusion module, for fusing multi-scale image feature representation and input data set, generating multi-modal state representation vector; Intervention strategy network module, for combining multi-modal state representation vector based on trust domain policy optimization algorithm, output patient intervention behavior decision; Reward and rare event discrimination module, for constructing reward function according to feedback data after intervention execution, and discriminating whether rare high-risk event occurs, generating intervention strategy network for policy optimization; Model supervision optimization module, for constructing image feature response evaluation function according to intervention strategy network parameters and feedback data, and as a supervision signal to guide the dynamic optimization of multi-scale visual Transformer model; The intelligent gastrointestinal postoperative intestinal obstruction diagnosis and intervention decision system running process specifically comprises the following steps: S1, collecting multiple source data of postoperative gastrointestinal tract of patients, preprocessing multiple source data, and building input data set; S2, constructing multi-scale visual Transformer model, extracting abdominal image data from input data set, inputting into multi-scale visual Transformer model, and extracting multi-scale image feature representation; S3, fuse multi-scale image feature representation with vital sign data and postoperative behavior record data in input data set, and construct patient multi-modal state representation vector; S4, input the multi-modal state representation vector of the patient into the intervention strategy network based on the trust domain policy optimization, set the maximum KL divergence limit threshold, and generate intervention behavior decision; S5, according to the intervention behavior decision, implement the corresponding intervention measures to the patient, collect the feedback data after the intervention execution, calculate the strategy reward value between the intervention behavior and the result, adopt the trust domain policy optimization algorithm, and update the parameters of the intervention strategy network in the preset KL divergence constraint range, obtain the optimized intervention strategy network; S6, the optimized intervention strategy network processes the multi-modal state representation vector of the patient, continuously outputs new intervention behavior decision, realizes the dynamic update and application cycle of intervention strategy; S7, according to the parameters of intervention strategy network and feedback data, construct image feature response evaluation function, as a supervision signal input into multi-scale visual Transformer model, optimize multi-scale visual Transformer model; S2 specifically comprises: S21, extracting the abdominal image data of the abdominal region from the input data set, constructing an original image tensor wherein, denotes the image height, denotes the image width, denotes the image channel number, R is the real number set; S22, obtaining the original image tensor Down-sampling is performed using a multi-scale pyramid strategy to obtain a set of abdominal image tensors with different spatial resolutions , wherein, represents an abdominal image tensor at the s-th spatial resolution, represents the total number of scales of the abdominal image tensors in the multi-scale pyramid structure; S23, obtaining an abdominal image tensor of each scale , dividing according to a fixed size window to obtain a non-overlapping window set , and calculating the self-attention features within the window by using a local self-attention mechanism within each window, wherein, represents the image sub-block of the i-th window region obtained after dividing the abdominal image tensor at the s-th spatial resolution, represents the abdominal image tensor at the s-th scale, the total number of windows after dividing. S24, for each scale, the feature representation of each window is extracted by a local self-attention mechanism, and the features of all windows are spliced in the original spatial order to obtain the encoding feature map at the scale ; S25, the encoding feature maps of all scales are fused by channel splicing and weighted summation, and the encoding feature maps of each scale are fused to obtain the fused multi-scale image feature representation; S26, the fused multi-scale image feature representation An input dual-branch collaborative architecture is input to the structural decoding branch and the auxiliary discrimination branch respectively. The structural decoding branch extracts abdominal structure feature information, and the auxiliary discrimination branch extracts intestinal obstruction risk discrimination related features.​ S27、in the structure decoding branch In the decoder of the hierarchical deconvolution and skip connection mechanism, the abdominal structure feature information is reconstructed to obtain the abdominal structure prediction map by multi-level up-sampling and spatial feature fusion in the hierarchical order from deep to shallow. S28, in the auxiliary discrimination branch The features related to the risk of intestinal obstruction are processed by using a full connection layer and an activation function, and an auxiliary risk discrimination result is output. S29, in the multi-scale image feature representation Further introduce sliding window local attention mechanism, cover each area of the feature map by sliding window, calculate local self-attention in each window S210, in the multi-scale visual Transformer model training process, a structure loss function and a function loss function are combined by weighting, the multi-scale visual Transformer model parameters are optimized by minimizing the combined joint loss function, the structure loss function is constructed by using a pixel-level cross-entropy loss function, the function loss function is constructed by using a binary classification cross-entropy loss function, and finally the structure loss function and the function loss function are linearly combined by setting a weighting coefficient to form a joint loss function; S211, in the training process, knowledge distillation is performed on the low-resolution feature branch by the high-resolution feature branch, and the KL divergence of the output distribution of the high and low resolution branches is minimized; S212, after the multi-scale visual Transformer model is trained, all parameter configurations of the multi-scale visual Transformer model are saved, the newly collected abdominal image data is input into the trained multi-scale visual Transformer model, and the final output multi-scale image feature representation is obtained; The S3 specifically comprises: S31, extract vital sign data and postoperative behavior record data at the same time point as the abdominal image data from the input data set, the vital sign data includes time sequence monitoring parameters such as body temperature, heart rate, respiratory rate, blood pressure and blood oxygen saturation, and the postoperative behavior record data includes key behavior time sequence of patient exhaust, defecation, eating and activity; S32, normalize and time align the extracted vital sign data and postoperative behavior record data, and construct a multi-modal structured feature tensor wherein, denotes the number of time steps, denotes the number of clinical feature dimensions, is a real set; S33, obtaining the multi-scale image feature representation Temporal index alignment is performed, so that the image features and the clinical data correspond to the same time step. S34, aligning the multi-scale image feature representation with the obtained multi-modal structured feature tensor performing feature-level fusion to obtain a multi-modal state representation vector of the patient in a concatenation manner ; S35, batch arranging the patient multi-modal state representation vectors at all time steps to form patient multi-modal state representation vectors for intervention strategy network input; The S4 specifically comprises: S41, obtaining the patient multi-modal state representation vector as the current input state into the intervention policy network based on the trust region policy optimization; S42, set the intervention behavior set in the intervention strategy network wherein, represents the first intervention behavior, is the total number of intervention behaviors; S43, initialize the intervention strategy network parameters , using a deep neural network to represent the patient's multi-modal state vector input the intervention strategy network, and obtain the probability distribution result of each intervention behavior through the forward calculation of the intervention strategy network; S44, set the trust domain constraint, limit the maximum KL divergence threshold of each policy distribution update ; S45, in the strategy optimization process, medical prior constraints are embedded, and clinical medical prior knowledge such as medical guidelines, surgical contraindications and expert experience is integrated into the strategy optimization process in the form of constraint terms, and the weight in the probability distribution of the intervention strategy network is reduced for intervention behaviors that do not meet medical standards or have major medical risks; S46, a dynamic self-adaptive mechanism is adopted according to the signal function Dynamic adjustment of KL divergence threshold : ； S47, a multi-objective hierarchical trust region optimization mechanism is introduced, intervention target is refined into multiple hierarchical targets including efficacy maximization, safety risk minimization and medical resource optimization in the intervention strategy network optimization process, independent trust region constraints are established for each target, and strategy output under different targets is balanced through hierarchical aggregation mechanism; S48. During the training and optimization of the intervention strategy network, historical occurrence probabilities below a preset threshold are used. And the corresponding risk score is higher than the preset threshold. The event samples are defined as rare high-risk event samples, and an experience memory pool is set up for rare high-risk event samples. Furthermore, the sampling weight of rare high-risk event samples in the training samples is increased, and more stringent trust region constraints are applied to training batches involving rare high-risk events during policy updates. Strengthen convergence constraints; S49, collect feedback after patient intervention execution, build reward function : ； wherein, denotes the clinical feedback reward mapping function, dynamically adjusts the reward weight according to the individual response of the patient, divides the reward signal into different levels of short-term effect, long-term outcome and risk punishment; at the same time, a feedback traceability explanation mechanism is introduced in the reward signal generation process, the key contribution factors of the reward signal are labeled and traceability analysis is performed, the individualization, dynamics and explainability of the reward feedback are improved, and more rich feedback information support is provided for the continuous optimization and accurate adjustment of the intervention strategy network; S410, output the intervention behavior decision generated by the optimized intervention strategy network, apply the intervention behavior decision to the current actual intervention of the patient, and collect and dynamically update the closed-loop effect feedback; The S5 specifically comprises: S51, according to the intervention behavior decision output by the intervention strategy network, implement corresponding medical intervention measures on the patient, the medical intervention measures include conservative observation, diet control, drug treatment, gastrointestinal decompression, image review or surgical treatment; S52, collect feedback data after each intervention measure is implemented, the feedback data including symptom performance of the patient, vital sign parameters, laboratory examination indexes, and adverse event information, to form a feedback data set ; S53、to the feedback dataset preprocessing to obtain an intervention effect change feature vector , and constructing a comprehensive reward function according to the intervention effect change feature vector : ； wherein, is a clinical feedback reward mapping function; S54, inputting the comprehensive reward function and a rare high-risk event indicator function as feedback signals into the intervention policy network, the rare high-risk event indicator function being used to determine whether an event with a historical occurrence probability lower than a preset threshold and a risk score higher than a preset threshold occurs to the patient after the intervention, the rare high-risk event indicator function taking a value of 1 when such an event occurs in the feedback data, and otherwise taking a value of 0, in combination with the current state-action pair to perform policy optimization, wherein, denotes a multi-modal state representation vector of the patient at the t-th time step, denotes an intervention behavior decision selected and executed at the t-th time step; S55, parameter updating is performed by using a trust region strategy optimization algorithm, and the optimization target is ; ； And meet the dynamic KL divergence constraint: ； wherein, denotes the current intervention policy network parameters Next, a multi-modal state representation vector for the patient selects an intervention action decision the probability, denotes the intervention policy network parameters before the last update Next, the probability for the same state-action pair, is a reward signal weighting coefficient, is a rare event reward weighting coefficient, is an expectation operator, denotes the parameters corresponding to the maximum of the objective function , is a KL divergence threshold, denotes the KL divergence; S56, continuously collect and archive each intervention and its feedback data, dynamically adjust the reward function parameters and rare event weight parameters; S57, the steps S51-S56 are cyclically executed to realize continuous iterative updating of the intervention strategy network based on the reward mechanism and dynamic strategy optimization; The S7 specifically comprises: S71, according to the latest intervention strategy network parameters and patient intervention after the feedback data, combined with the current fusion of multi-scale image feature representation , construct image feature response evaluation function ; S72, input the image feature response evaluation function as a supervisory signal into the multi-scale visual Transformer model as the basis for adjusting the attention distribution weight, and optimize the attention weight parameters of the multi-scale visual Transformer model at different spatial scales; S73, in the optimization process of the multi-scale visual Transformer model, iteratively update the attention distribution weight of each scale Transformer encoding layer, and the multi-scale visual Transformer model can adaptively strengthen the feature extraction capability for the key region highly related to the feedback when processing new input abdominal image data; S74, save the optimized multi-scale visual Transformer model parameters to form a multi-scale visual Transformer model capable of sustainable adaptive updating, and continuously use the image feature response evaluation function and the attention guiding mechanism in the abdominal image data analysis process to realize dynamic supervision optimization and individualized feature extraction capability improvement of the multi-scale visual Transformer model. 2.The intelligent gastrointestinal postoperative ileus diagnosis and intervention decision system according to claim 1, characterized in that, The multi-source data specifically includes endoscopic images, abdominal image data, vital sign data, postoperative behavior record data, physiological parameters and heterogeneous information of electronic medical record texts of postoperative gastrointestinal patients. 3.The intelligent gastrointestinal postoperative ileus diagnosis and intervention decision system according to claim 1, characterized in that, The pre-processing of the multi-source data specifically includes image normalization, text segmentation and coding, physiological parameter standardization and data time alignment.

Images