Intelligent gastrointestinal tract postoperative intestinal obstruction diagnosis and intervention decision-making system and method thereof

Abstract

The invention discloses an intelligent gastrointestinal postoperative intestinal obstruction diagnosis and intervention decision-making system and method, and the method comprises the following steps: S1, collecting and preprocessing the multi-source data of a postoperative patient, and constructing an input data set; s2, a multi-scale visual Transform model is constructed, and multi-scale image feature representation is extracted; s3, constructing a multi-modal state representation vector; s4, inputting the multi-modal state representation vector into an intervention strategy network to generate an intervention decision; s5, collecting feedback data, calculating a reward value, and optimizing the strategy network; s6, processing the multi-modal state vector by the optimized strategy network, and outputting a new intervention decision; and S7, constructing an image feature response evaluation function, and supervising and optimizing a multi-scale vision Transform. According to the method, the intelligent diagnosis and personalized intervention of the postoperative intestinal obstruction are realized by combining the multi-scale visual Transform with the trust region strategy optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis and intelligent decision support, and particularly to an intelligent diagnosis and intervention decision-making system and method for postoperative intestinal obstruction of the gastrointestinal tract. Background Art

[0002] At present, postoperative intestinal obstruction of the gastrointestinal tract is one of the common and serious complications in digestive surgery with a relatively high disability rate and fatality rate. With the popularization of concepts such as minimally invasive surgery and enhanced recovery after surgery, the number of postoperative gastrointestinal patients has been increasing year by year, and higher requirements for the early diagnosis and precise intervention of intestinal obstruction have been put forward clinically. Traditional diagnosis of intestinal obstruction mainly relies on clinicians to manually read and interpret abdominal images (such as abdominal X-rays, CT), and comprehensively analyze the patient's vital signs, symptom manifestations, and behavior records. However, manual image reading is highly subjective, relying on doctor experience, and it is difficult to deeply integrate a large amount of multi-modal data. Especially in the case of early or atypical manifestations of intestinal obstruction, problems such as missed diagnosis and misdiagnosis are likely to occur, affecting the prognosis of patients.

[0003] With the rapid development of artificial intelligence and deep learning technologies, some studies have attempted to use methods such as convolutional neural networks to automatically extract features from abdominal images for auxiliary diagnosis. However, most of the existing algorithms only target single-modal or single-scale images, lacking a comprehensive modeling of multi-scale features of abdominal structures, and it is difficult to handle the complex and variable images and pathological manifestations in real clinical scenarios. At the same time, most of the existing intelligent decision-making systems use fixed strategy optimization algorithms, which are difficult to balance safety and individual differences. When facing the dynamic feedback and abnormally high-risk events during the patient intervention process, the strategy adjustment is not flexible enough, easily leading to distorted intervention suggestions or insufficient safety.

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

[0005] Therefore, how to provide an intelligent diagnosis and intervention decision-making system and method for postoperative intestinal obstruction of the gastrointestinal tract is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] An object of the present invention is to propose an intelligent diagnosis and intervention decision-making system and method for postoperative intestinal obstruction of the gastrointestinal tract. The present invention 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-making network optimized by a trust region strategy, and details the whole process from multi-source data collection, preprocessing, feature fusion, intelligent intervention decision-making to dynamic feedback optimization, and has the advantages of high diagnostic accuracy, high degree of intervention intelligence, strong safety, strong response ability to rare high-risk events and outstanding continuous adaptive optimization ability.

[0007] The intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract according to an embodiment of the present invention includes the following steps: S1. Collect multi-source data of patients after gastrointestinal surgery, preprocess the multi-source data, and construct an input data set; S2. Construct a multi-scale Vision Transformer model, extract abdominal image data from the input data set, input it into the multi-scale Vision Transformer model, and extract multi-scale image feature representations; S3. Integrate the multi-scale image feature representations with the vital sign data and postoperative behavior record data in the input data set to construct a patient multi-modal state representation vector; S4. Input the patient's multi-modal state representation vector into an intervention strategy network optimized by a trust region strategy, set a maximum KL divergence limit threshold, and generate an intervention behavior decision; S5. Implement corresponding intervention measures on the patient according to the intervention behavior decision, collect feedback data after the intervention is executed, calculate the policy reward value between the intervention behavior and the result, and use the trust region strategy optimization algorithm to iteratively update the parameters of the intervention strategy network within a preset KL divergence constraint range to obtain an optimized intervention strategy network; S6. Process the patient's multi-modal state representation vector with the optimized intervention strategy network, continuously output new intervention behavior decisions, and realize the dynamic update and application cycle of the intervention strategy; S7. According to the parameters of the intervention strategy network and the feedback data, construct an image feature response evaluation function, use it as a supervision signal to input into the multi-scale Vision Transformer model, and optimize the multi-scale Vision Transformer model.

[0008] Optionally, the multi-source data specifically includes heterogeneous information such as endoscopic images, abdominal image data, vital sign data, postoperative behavior record data, physiological parameters, and electronic medical record texts of patients after gastrointestinal surgery.

[0009] Optionally, the preprocessing of the multi-source data specifically includes image normalization, text tokenization and encoding, physiological parameter standardization, and data time alignment.

[0010] Optionally, S2 specifically includes: S21. Extract abdominal image data of the abdominal region from the input dataset to construct an original image tensor , where represents the image height, represents the image width, represents the number of image channels, and R is the set of real numbers; S22. Perform downsampling on the obtained original image tensor using a multi-scale pyramid strategy to obtain a set of abdominal image tensors with multiple different spatial resolutions , where represents the 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. Divide each scale of the obtained abdominal image tensor into non-overlapping window sets according to a window of a fixed size , and calculate the self-attention features within each window using a local self-attention mechanism, where 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 total number of windows into which the abdominal image tensor is divided at the s-th scale; S24. For all window regions obtained at each scale, extract the feature representations of each window through a local self-attention mechanism, and splice the features of all windows in the original spatial order to obtain the encoded feature map at the scale ; S25. Fuse the encoded feature maps at all scales by splicing them according to channels and performing weighted summation to obtain a fused multi-scale image feature representation; S26. Input the fused multi-scale image feature representation into a dual-branch collaborative architecture and input it into 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 features related to the discrimination of intestinal obstruction risk; S27. In the decoder of the structural decoding branch , adopt a hierarchical deconvolution and skip connection mechanism to perform multi-level upsampling and spatial feature fusion on the abdominal structure feature information in the hierarchical order from deep to shallow to reconstruct an abdominal structure prediction map; S28. In the auxiliary discrimination branch the features related to intestinal obstruction risk discrimination are processed using a fully connected layer and an activation function to output an auxiliary risk discrimination result; S29. Further introduce a sliding window local attention mechanism in the multi-scale image feature representation to cover each region of the feature map through a sliding window and calculate local self-attention within each window; S210. During the training process of the multi-scale Vision Transformer model, a structural loss function and a functional loss function are combined with weights. The parameters of the multi-scale Vision Transformer model are optimized by minimizing the combined joint loss function. The structural loss function is constructed using a pixel-level cross-entropy loss function, and the functional loss function is constructed using a binary cross-entropy loss function. Finally, by setting a weighting coefficient, the structural loss function and the functional loss function are linearly weighted and combined to form a joint loss function; S211. During the training process, the high-resolution feature branch is used to perform knowledge distillation on the low-resolution feature branch by minimizing the KL divergence of the output distributions of the high and low-resolution branches; S212. After the training of the multi-scale Vision Transformer model is completed, all parameter configurations of the multi-scale Vision Transformer model are saved, and the newly acquired abdominal imaging data is input into the trained multi-scale Vision Transformer model to obtain the finally output multi-scale image feature representation.

[0011] Optionally, the specific content of S3 includes: S31. Extract the vital sign data and postoperative behavior record data at the same time point as the abdominal imaging data from the input dataset. The vital sign data includes the time series monitoring parameters of body temperature, heart rate, respiratory rate, blood pressure, and blood oxygen saturation. The postoperative behavior record data includes the key behavior time series of the patient's exhaust, defecation, eating, and activity; S32. Normalize and time-align the extracted vital sign data and postoperative behavior record data to construct a multi-modal structured feature tensor , where represents the number of time steps, represents the number of clinical feature dimensions, is the set of real numbers; S33. Align the obtained multi-scale image feature representation in time series index so that the image features and clinical data correspond at the same time step; S34. Combine the aligned multi-scale image feature representation with the obtained multi-modal structured feature tensor Perform feature-level fusion and obtain the patient's multi-modal state representation vector through concatenation. ; S35. Batch process the patient's multi-modal state representation vectors at all time steps to form the patient's multi-modal state representation vector for input to the intervention policy network.

[0012] Optionally, S4 specifically includes: S41. Use the obtained patient's multi-modal state representation vector as the current input state and input it into the intervention policy network based on trust region policy optimization; S42. Set the intervention behavior set in the intervention policy network, where represents the th intervention behavior, and is the total number of intervention behaviors; S43. Initialize the parameters of the intervention policy network. Use a deep neural network to input the patient's multi-modal state representation vector into the intervention policy network, and obtain the probability distribution result of each intervention behavior through the forward calculation of the intervention policy network; S44. Set the trust region constraint to limit the maximum KL divergence threshold for each policy distribution update; S45. During the policy optimization process, embed medical prior constraints, and incorporate the clinical medical prior knowledge of medical guidelines, surgical taboos, and expert experience into the policy optimization process in the form of constraint terms. For intervention behaviors that do not conform to medical norms or have significant medical risks, reduce the weight in the probability distribution of the intervention policy network; S46. Adopt a trust region dynamic adaptive mechanism to dynamically adjust the KL divergence threshold according to the signal function : ; S47. Introduce a multi-objective hierarchical trust region optimization mechanism. During the optimization process of the intervention policy network, refine the intervention objectives into multiple hierarchical objectives, including maximizing efficacy, minimizing safety risks, and optimizing medical resources. Establish independent trust region constraints for each objective, and balance the policy outputs under different objectives through a hierarchical aggregation mechanism; S48. During the training and optimization process of the intervention policy network, define event samples with a historical occurrence probability lower than the preset threshold and a corresponding risk score higher than the preset threshold as rare high-risk event samples. For rare high-risk event samples, set up an experience memory pool And increase the sampling weight of rare high-risk event samples in the training samples, and apply more stringent trust region constraints to the training batches involving rare high-risk events during the policy update process , strengthening the convergence constraint; S49. Collect the feedback after the patient's intervention is executed, and construct a reward function : ; Among them, represents the 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 effects, long-term outcomes and risk penalties; at the same time, a feedback traceability explanation mechanism is introduced during the generation of the reward signal to label and trace the key contributing factors of the reward signal, improving the personalization, dynamics and interpretability of the reward feedback, and providing richer feedback information support for the continuous optimization and precise 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 patient's current actual intervention, and collect the closed-loop effect feedback and perform dynamic policy update.

[0013] Optionally, the S5 specifically includes: S51. According to the intervention behavior decision output by the intervention strategy network, implement corresponding medical intervention measures on the patient, and the medical intervention measures include conservative observation, diet control, drug treatment, gastrointestinal decompression, imaging review or surgical treatment; S52. Collect the feedback data after each intervention measure is implemented, and the feedback data includes the patient's symptom manifestations, vital sign parameters, laboratory test indicators and adverse event information, and form a feedback data set ; S53. Preprocess the feedback data set to obtain the intervention effect change feature vector , and construct a comprehensive reward function according to the intervention effect change feature vector : ; Among them, is the clinical feedback reward mapping function; S54. Input the comprehensive reward function and the rare high-risk event indicator function as feedback signals into the intervention strategy network. The rare high-risk event indicator 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 patient's intervention. When such an event appears in the feedback data, the rare high-risk event indicator function takes a value of 1, otherwise it takes a value of 0, and jointly optimize the policy with the current state-action pair , where, represents the 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; S55. Adopt the trust region strategy optimization algorithm for parameter update, and the optimization objective is ; S56. Continuously collect and archive each intervention and its feedback data, and dynamically adjust the reward function parameters and rare event weight parameters; S57. Loop through the above steps S51 to S56 to achieve continuous iterative update of the intervention strategy network based on the reward mechanism and dynamic strategy optimization.

[0014] Optionally, the specific steps of S7 include: S71. According to the parameters of the latest intervention strategy network and the feedback data after the patient's intervention execution, combined with the current fused multi-scale image feature representation , construct an image feature response evaluation function ; S72. Take the image feature response evaluation function as a supervision signal and input it into the multi-scale vision Transformer model as the basis for adjusting the attention distribution weights, and optimize the attention weight parameters of the multi-scale vision Transformer model at different spatial scales; S73. During the optimization process of the multi-scale vision Transformer model, iteratively update the attention distribution weights of each scale Transformer encoding layer. When the multi-scale vision Transformer model processes new input abdominal imaging data, it can adaptively strengthen the feature extraction ability for key regions highly relevant to the feedback; S74. Save the parameters of the optimized multi-scale vision Transformer model to form a multi-scale vision Transformer model that can be continuously and adaptively updated. During the analysis of abdominal imaging data, continuously utilize the image feature response evaluation function and the attention guidance mechanism to achieve dynamic supervision optimization of the multi-scale vision Transformer model and improvement of the individual feature extraction ability.

[0015] The intelligent diagnosis and intervention decision-making system for postoperative intestinal obstruction of the gastrointestinal tract according to the embodiment of the present invention includes the following modules: A data acquisition and preprocessing module for collecting multi-source data of postoperative gastrointestinal patients and performing preprocessing to construct an input data set; A multi-scale vision Transformer feature extraction module for extracting multi-scale image feature representations from the input data set; A multi-modal state fusion module, which is used to fuse multi-scale image feature representations with an input data set to generate a multi-modal state representation vector; An intervention strategy network module, which is used to output a patient intervention behavior decision based on the trust region policy optimization algorithm and in combination with the multi-modal state representation vector; A reward and rare event discrimination module, which is used to construct a reward function according to the feedback data after intervention execution, and discriminate whether a rare high-risk event occurs, and generate an intervention strategy network for policy optimization; A model supervision and optimization module, which is used to construct an image feature response evaluation function according to the intervention strategy network parameters and feedback data, and use it as a supervision signal to guide the dynamic optimization of the multi-scale vision Transformer model.

[0016] The beneficial effects of the present invention are: By deeply integrating the multi-scale vision Transformer model with the intervention strategy network of trust region policy optimization, the present invention realizes the intelligent diagnosis and individualized intervention decision-making of postoperative intestinal obstruction of the gastrointestinal tract. Compared with the prior art, the present invention can extract multi-scale and multi-level structural features from abdominal imaging data, and combine multi-modal clinical information such as the patient's vital signs and behavior records, improving the model's recognition ability for early, complex and atypical intestinal obstruction cases. By introducing the trust region policy optimization algorithm, the present invention can not only output safe, controllable and dynamically adaptive individualized intervention suggestions, but also adjust the policy parameters in real time according to the patient's clinical feedback, ensuring the flexibility and efficiency of the decision-making process. Further, the present invention constructs a comprehensive reward mechanism and a rare high-risk event weighting mechanism, enabling the intervention strategy network to make key responses and optimizations to extreme risk states, significantly improving the ability to handle abnormal and rare clinical events.

[0017] The system of the present invention has characteristics such as continuous adaptive optimization, multi-objective collaborative management and traceable decision-making process. It 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 applications, providing more intelligent, precise and efficient diagnosis and treatment services for postoperative gastrointestinal patients, and having significant clinical application promotion value and social and economic benefits. Description of the Drawings

[0018] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. In the drawings: Figure 1 It is a flowchart of the intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract proposed by the present invention; Figure 2This is a schematic structural diagram of the intelligent diagnosis and intervention decision-making system for postoperative intestinal obstruction of the gastrointestinal tract proposed by the present invention. Detailed implementation manners

[0019] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0020] Refer to Figure 1 , the intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract includes the following steps: S1. Collect multi-source data of patients after gastrointestinal tract surgery, preprocess the multi-source data, and construct an input data set; S2. Construct a multi-scale vision Transformer model, extract abdominal image data from the input data set, input it into the multi-scale vision Transformer model, and extract multi-scale image feature representations; S3. Fuse the multi-scale image feature representations with the vital sign data and postoperative behavior record data in the input data set to construct a patient multi-modal state representation vector; S4. Input the patient's multi-modal state representation vector into the intervention strategy network based on trust region policy optimization, set the maximum KL divergence limit threshold, and generate an intervention behavior decision; S5. Implement corresponding intervention measures on the patient according to the intervention behavior decision, collect the feedback data after the intervention is executed, calculate the policy reward value between the intervention behavior and the result, and use 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; S6. Process the patient's multi-modal state representation vector with the optimized intervention strategy network, continuously output new intervention behavior decisions, and realize the dynamic update and application cycle of the intervention strategy; S7. According to the parameters of the intervention strategy network and the feedback data, construct an image feature response evaluation function, use it as a supervision signal to input into the multi-scale vision Transformer model, and optimize the multi-scale vision Transformer model.

[0021] The present invention realizes the intelligent and refined management of the whole process of diagnosis and intervention for postoperative intestinal obstruction of the gastrointestinal tract. By adopting multi-source data acquisition and preprocessing, the comprehensiveness of patient information and data quality are ensured; a multi-scale vision Transformer model is used to extract multi-level features from abdominal images, effectively improving the recognition ability of complex structure areas; by fusing image features with vital signs and behavioral data, a multi-modal representation reflecting the comprehensive state of the patient is accurately constructed. Based on the intervention strategy network optimized by the trust region strategy, safe and controllable individualized intervention decisions can be realized, and the stability and reliability of the optimization process are ensured through dynamic KL divergence constraints. The system continuously collects feedback data after the intervention execution, and realizes the decision-making closed-loop and continuous optimization by using the reward mechanism and adaptive strategy update. In addition, the image feature response evaluation and Transformer model optimization mechanism driven by feedback further improve the sensitivity of the model to key clinical information and the feature extraction ability. Overall, the present invention can greatly improve the accuracy of intestinal obstruction diagnosis and the individualized level of intervention, enhance the system's response and adaptive optimization ability to abnormal events, and has significant clinical application value and promotion prospects.

[0022] In this embodiment, the multi-source data specifically includes heterogeneous information such as endoscopic images, abdominal imaging data, vital signs data, postoperative behavior record data, physiological parameters, and electronic medical record texts of postoperative gastrointestinal patients.

[0023] In this embodiment, the preprocessing of the multi-source data specifically includes image normalization, text tokenization and encoding, physiological parameter standardization, and data time alignment. The image normalization, text tokenization and encoding, physiological parameter standardization, and data time alignment mean that by performing normalization processing on the image to unify the pixel value distribution, performing tokenization and vector encoding on the text to extract semantic features, standardizing the physiological parameters to eliminate the dimension difference, and aligning various types of data according to the timestamp.

[0024] In this embodiment, S2 specifically includes: S21. Extract abdominal imaging data of the abdominal region from the input dataset to construct an original image tensor , where represents the image height, represents the image width, represents the number of image channels, and R is the set of real numbers; S22. Perform downsampling on the obtained original image tensor by using a multi-scale pyramid strategy to obtain a set of abdominal imaging tensors with multiple different spatial resolutions , specifically, downsample the abdominal image data at different spatial resolutions respectively to obtain multiple abdominal image tensors with different spatial resolutions. Each image tensor at each resolution retains the original abdominal structure information, and the abdominal image tensors at all scales jointly form a multi-scale pyramid structure. Among them, represents the 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. For each scale of the obtained abdominal image tensor , divide it according to a window of a fixed size to obtain a set of non-overlapping windows , and calculate the self-attention features within each window using the local self-attention mechanism. Among them, represents the image sub-block of the i-th window area obtained after dividing the abdominal image tensor at the s-th spatial resolution, represents the total number of windows into which the abdominal image tensor is divided at the s-th scale; S24. For all window areas obtained at each scale, extract the feature representations of each window through the local self-attention mechanism respectively, and splice the features of all windows in the original spatial order to obtain the encoded feature map at the scale, and the encoded feature map completely reflects the spatial structure and fine-grained information of the abdominal image at the scale; S25. For the encoded feature maps at all scales, fuse the encoded feature maps at each scale in the way of channel splicing and weighted summation to obtain a fused multi-scale image feature representation, and the multi-scale image feature representation comprehensively reflects the structural feature information of the abdominal image at different spatial resolutions; S26. Input the fused multi-scale image feature representation into the dual-branch collaborative architecture, and input it into the structure decoding branch and the auxiliary discrimination branch respectively. The structure decoding branch extracts the abdominal structure feature information, and the auxiliary discrimination branch extracts the features related to the risk discrimination of intestinal obstruction; S27. In the decoder of the structure decoding branch , adopt the hierarchical deconvolution and skip connection mechanism to perform multi-level upsampling and spatial feature fusion on the abdominal structure feature information in the hierarchical order from deep to shallow, and reconstruct the abdominal structure prediction map. The adoption of the hierarchical deconvolution and skip connection mechanism to perform multi-level upsampling and spatial feature fusion on the abdominal structure feature information in the hierarchical order from deep to shallow means that through the hierarchical deconvolution and skip connection mechanism, the abdominal structure features at different levels are upsampled step by step and the spatial information is fused to achieve high-precision reconstruction of the abdominal structure. S28. In the auxiliary discrimination branch process the features related to intestinal obstruction risk discrimination using a fully connected layer and an activation function, and output the auxiliary risk discrimination result. The process of processing the features related to intestinal obstruction risk discrimination using a fully connected layer and an activation function means inputting the discrimination features into a fully connected neural network and extracting high-order features through linear transformation and non-linear activation; S29. Further introduce a sliding window local attention mechanism in the multi-scale image feature representation by covering each area of the feature map with a sliding window and calculating local self-attention within each window; S210. During the training process of the multi-scale Vision Transformer model, use a structural loss function and a functional loss function for weighted combination, and optimize the parameters of the multi-scale Vision Transformer model by minimizing the combined loss function of the weighted combination. The structural loss function is constructed using a pixel-level cross-entropy loss function, and the functional loss function is constructed using a binary cross-entropy loss function. Finally, by setting the weighting coefficient, linearly weight and combine the structural loss function and the functional loss function to form a combined loss function; S211. During the training process, use the high-resolution feature branch to perform knowledge distillation on the low-resolution feature branch by minimizing the KL divergence of the output distributions of the high and low resolution branches; S212. After the training of the multi-scale Vision Transformer model is completed, save all the parameter configurations of the multi-scale Vision Transformer model, input the newly acquired abdominal imaging data into the trained multi-scale Vision Transformer model, and obtain the finally output multi-scale image feature representation.

[0025] Through the defined construction and optimization process of the multi-scale vision Transformer model, the present invention realizes multi-scale and multi-level structured feature extraction and discriminant analysis of abdominal images. Through multi-scale pyramid downsampling, local self-attention mechanism and channel fusion operation, it can accurately capture the global and fine-grained structural features of abdominal images at different spatial resolutions, greatly improving the model's recognition ability for complex abdominal structures and tiny lesions. The dual-branch collaborative architecture enables the structure decoding and risk discrimination to be carried out synchronously and efficiently, ensuring both the accuracy of structure reconstruction and the sensitivity to high-risk areas of intestinal obstruction. The use of sliding window local attention mechanism and knowledge distillation effectively improves the model's attention 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 present invention significantly improves the accuracy, robustness and model generalization ability of intelligent analysis of abdominal images, providing a solid feature basis and reliable support for subsequent intelligent diagnosis and intervention.

[0026] In this embodiment, step S3 specifically includes: S31. Extract the vital sign data and postoperative behavior record data at the same time point as the abdominal image data from the input dataset. The vital sign data includes the time series monitoring parameters of body temperature, heart rate, respiratory rate, blood pressure and blood oxygen saturation. The postoperative behavior record data includes the key behavior time series of the patient's exhaust, defecation, eating and activities; S32. Normalize and time-align the extracted vital sign data and postoperative behavior record data to construct a multi-modal structured feature tensor , where represents the number of time steps, represents the number of clinical feature dimensions, is the set of real numbers. The normalization and time-alignment means standardizing the clinical data with different dimensions into a unified numerical range and sorting them according to the time stamp, so that various types of data are synchronously corresponding at the same time step; S33. Perform time series index alignment on the obtained multi-scale image feature representation . For the image features and clinical data corresponding at the same time step, the time series index alignment means matching the multi-scale image features with the clinical data at the corresponding time step according to the time stamp to ensure the alignment of multi-modal information at the same time point in the feature sequence; S34. Perform feature-level fusion on the aligned multi-scale image feature representation and the obtained multi-modal structured feature tensor and obtain the patient multi-modal state representation vector by using the splicing method; S35. Batch organize the patient multi-modal state representation vectors at all time steps to form the patient multi-modal state representation vectors for input to the intervention strategy network. The batch organization means uniformly encoding the multi-modal feature vectors of the patient's physiology, behavior, and environment generated at each time step in chronological order and merging them into a two-dimensional batch tensor.

[0027] Through the extraction and fusion process of multi-modal temporal features, the present invention realizes the accurate modeling of the all-round and dynamic state of gastrointestinal postoperative patients. Through the vital signs and postoperative behavior data extracted at the same time point as the abdominal imaging data, the physiological changes and behavior characteristics of the patients at different postoperative periods can be completely reflected. The normalization and temporal alignment operations ensure the synchronization and comparability of multi-modal data, significantly reducing the information loss caused by data heterogeneity. By fusing multi-scale image features with structured clinical data and constructing the patient multi-modal state representation using the feature splicing method, not only the information at the anatomical, functional, and behavioral levels is comprehensively understood, but also the model's ability to capture individual differences and dynamic changes is enhanced. Finally, the sequence of patient multi-modal state representations obtained by batch organization provides a high-quality and temporally ordered comprehensive representation basis for the input of the subsequent intelligent intervention strategy network, greatly improving the model's representation ability of the complex state of patients and the scientific nature of intelligent decision-making, and laying a solid data foundation for individualized precise intervention and risk prediction.

[0028] In this embodiment, the specific steps of S4 are as follows: S41. Use the obtained patient multi-modal state representation vectors as the current input state and input it into the intervention strategy network based on trust region policy optimization; S42. Set the intervention behavior set in the intervention strategy network, where represents the th intervention behavior, and is the total number of intervention behaviors; S43. Initialize the parameters of the intervention strategy network. Use a deep neural network to input the patient multi-modal state representation vectors into the intervention strategy network, and obtain the probability distribution results of each intervention behavior through the forward calculation of the intervention strategy network; S44. Set the trust region constraint to limit the maximum KL divergence threshold for each policy distribution update; S45. During the policy optimization process, embed medical prior constraints, and incorporate the clinical prior knowledge of medical guidelines, surgical taboos, and expert experience into the policy optimization process in the form of constraint terms. For intervention behaviors that do not conform to medical norms or pose significant medical risks, reduce the weight in the probability distribution of the intervention strategy network; S46. Adopt a trust region dynamic adaptive mechanism to dynamically adjust the KL divergence threshold according to the signal function : ; ; S47. Introduce a multi-objective hierarchical trust region optimization mechanism. During the optimization process of the intervention strategy network, refine the intervention objectives into multiple hierarchical objectives, including maximizing efficacy, minimizing safety risks, and optimizing medical resources. Establish independent trust region constraints for each objective, and balance the policy outputs under different objectives through a hierarchical aggregation mechanism. The maximization of efficacy, minimization of safety risks, and optimization of medical resources specifically refer to maximizing efficacy by quantifying the improvement degree of key indicators after intervention, evaluating the probability and severity of potential side effects to control safety risks, and optimizing the resource allocation efficiency by combining treatment costs and resource consumption rates; S48. During the training and optimization process of the intervention strategy network, define event samples with a historical occurrence probability lower than a preset threshold and a corresponding risk score higher than a preset threshold as rare high-risk event samples. For rare high-risk event samples, set up an experience memory pool and increase the sampling weight of rare high-risk event samples in the training samples. Apply more stringent trust region constraints to the training batches involving rare high-risk events during the policy update process , strengthening the convergence constraint; S49. Collect the feedback after the patient's intervention execution and construct a reward function : ; Among them, represents the clinical feedback reward mapping function, which dynamically adjusts the reward weight according to the individual response of the patient, and divides the reward signal into different levels such as short-term effect, long-term outcome, and risk penalty. At the same time, introduce a feedback traceability and interpretation mechanism during the generation of the reward signal, label and trace the key contributing factors of the reward signal, and enhance the personalization, dynamics, and interpretability of the reward feedback, providing richer feedback information support for the continuous optimization and precise 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 patient's current actual intervention, and collect closed-loop effect feedback and perform dynamic policy update.

[0029] Through the construction and optimization of an intervention strategy network optimized based on the trust region strategy, the scientificity and individualization level of the intervention decision-making for postoperative intestinal obstruction in the gastrointestinal tract have been significantly improved. By introducing multi-modal state representation as the input of the strategy network, it can comprehensively reflect the dynamic changes of patients and provide rich state basis for intervention strategies. Setting the intervention behavior set and the output behavior probability distribution of the deep neural network realizes the intelligent exploration of the high-dimensional and variable medical decision-making space. The trust region constraint and the dynamic adaptive mechanism ensure the safety and stability of the strategy optimization process and prevent the model from outputting extreme intervention suggestions. Integrating medical prior constraints and the multi-objective hierarchical trust region mechanism enables the decision-making network to take into account risk prevention and control and resource allocation while pursuing curative effects, achieving the collaborative optimization of multi-dimensional objectives. The empirical memory and reinforcement constraints on rare high-risk event samples improve the model's response and safety protection capabilities for abnormal clinical situations. Through the multi-level reward function design and the feedback traceability mechanism, the intervention strategy network can continuously adaptively adjust and trace and optimize, significantly improving the personalization, interpretability and closed-loop intelligent optimization capabilities of the intervention suggestions. Overall, the present invention can greatly improve the intelligence, precision and clinical usability of the management of postoperative intestinal obstruction.

[0030] In this embodiment, the S5 specifically includes: S51. According to the intervention behavior decision output by the intervention strategy network, implement corresponding medical intervention measures on the patient, and the medical intervention measures include conservative observation, diet control, drug treatment, gastrointestinal decompression, imaging reexamination or surgical treatment; S52. Collect the feedback data after each intervention measure is implemented, and the feedback data includes the symptom manifestations, vital sign parameters, laboratory test indicators and adverse event information of the patient to form a feedback data set ; S53. Preprocess the feedback data set to obtain an intervention effect change feature vector , and construct a comprehensive reward function according to the intervention effect change feature vector , and the preprocessing of the feedback data set refers to normalizing, time-aligning and denoising the physiological parameters, image indicators and subjective score data of the patient after intervention collected in the feedback data set: ; wherein, is the clinical feedback reward mapping function; S54. The comprehensive reward function And the rare high - risk event indicator function is input into the intervention strategy network as a feedback signal. The rare high - risk event indicator 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 for a patient. When such an event appears in the feedback data, the rare high - risk event indicator function takes a value of 1, otherwise it takes a value of 0. Combine the current state - action pair to perform policy optimization, where represents the 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; S55. Use the trust region policy optimization algorithm to update the parameters, and the optimization objective is : ; And satisfy the dynamic KL divergence constraint: ; Where represents the current intervention policy network parameters under which, for the multi - modal state representation vector of the patient, select the intervention behavior decision with probability, represents the probability of the same state - action pair under the intervention policy network parameters before the previous update, is the reward signal weighting coefficient, is the rare event reward weight coefficient, is the expectation operator, represents taking the parameter corresponding to the maximum value of the objective function, is the KL divergence threshold, represents the KL divergence; In the present invention, the optimization objective Formulas are essentially used to guide the intervention strategy network to continuously improve the scientificity and safety of clinical decisions. The practical significance lies in that after the model receives new patient feedback and reward signals each time, it comprehensively considers the intervention effect and the response to rare high-risk events, uses the two as measurement criteria, and automatically adjusts the intervention strategy to make the subsequent decision-making plan more in line with the individual needs of patients and the requirements of clinical risk prevention and control. During the optimization process, the model will also strictly constrain the change range between the old and new strategy distributions to ensure that there will be no drastic fluctuations or uncontrollable phenomena in each adjustment of the intervention strategy, thus guaranteeing the smoothness of medical decisions and clinical safety. Through the above mechanism, the model can not only continuously accumulate experience, strengthen the learning and protection of rare high-risk events, but also dynamically adapt to the feedback of different patients, realizing the individuation, precision and high safety of intelligent intervention. This optimization method provides a solid guarantee for the clinical practicability and wide popularization and application of the intelligent intervention system for postoperative ileus of the gastrointestinal tract.

[0031] S56. Continuously collect and archive data on each intervention and its feedback, and dynamically adjust the parameters of the reward function and the weights of rare event parameters; S57. Loop through the above steps S51 to S56 to achieve continuous iterative update of the intervention strategy network based on the reward mechanism and dynamic strategy optimization.

[0032] Through the proposed intervention decision-making and continuous optimization process, the present invention realizes a scientific closed loop for the intelligent intervention management of postoperative ileus of the gastrointestinal tract. By applying the decisions output by the intervention strategy network to patients and collecting multi-dimensional feedback data including symptom manifestations, vital signs, and adverse events, etc., the actual effect of the intervention measures can be comprehensively reflected. Using a comprehensive reward function to quantify the intervention effectiveness and combining it with a rare high-risk event indicator function to ensure that the model focuses on responding to rare and high-risk clinical events and strengthens learning. Using the trust region policy optimization algorithm with dynamic KL divergence constraint to achieve safe iterative update of network parameters and effectively avoid extreme strategy risks. The feedback after each round of intervention is continuously input into the model, which can not only dynamically adjust the parameters of the reward function and the weights of rare events, but also realize the individual evolution of the decision-making system and risk adaptive adjustment. Overall, the present invention significantly improves the accuracy, robustness and risk prevention and control ability of intelligent intervention decisions, guarantees the individuation, scientificity and clinical safety of the patient intervention process, and provides strong support for intelligent medical closed-loop management and precision treatment.

[0033] In this embodiment, the specific steps of S6 include: The optimized intervention strategy network is used to process the patient's current multi-modal state representation vector, combined with multi-dimensional information such as the patient's condition, vital signs, and postoperative recovery indicators, to infer and generate personalized intervention behavior decisions in real time. These decisions include but are not limited to drug administration suggestions, dietary intake control, exercise intervention plans, and clinical review reminders. As the patient's state continues to change, the system dynamically inputs the updated state information, continuously iterates and optimizes the intervention strategy, and forms a cyclic closed-loop control mechanism of "perception - decision - feedback - re-decision" to achieve precise intervention and continuous management of the risk of postoperative intestinal obstruction.

[0034] In this embodiment, S7 specifically includes: S71. According to the latest intervention strategy network parameters and the feedback data after the patient's intervention execution, combined with the currently fused multi-scale image feature representation , construct an image feature response evaluation function ; S72. Use the image feature response evaluation function as a supervision signal to input into the multi-scale vision Transformer model, as the basis for adjusting the attention distribution weights, and optimize the attention weight parameters of the multi-scale vision Transformer model at different spatial scales; S73. During the optimization process of the multi-scale vision Transformer model, iteratively update the attention distribution weights of each scale Transformer encoding layer. When the multi-scale vision Transformer model processes new input abdominal image data, it can adaptively enhance the feature extraction ability for key regions highly related to the feedback; S74. Save the optimized multi-scale vision Transformer model parameters to form a continuously self-adaptive updated multi-scale vision Transformer model, and continuously utilize the image feature response evaluation function and the attention guidance mechanism during the abdominal image data analysis process to achieve dynamic supervision optimization of the multi-scale vision Transformer model and improvement of the individual feature extraction ability.

[0035] Through the dynamic supervision optimization process of the defined multi-scale vision Transformer model, the sensitivity and adaptive ability of the model to key regions of abdominal images and clinical feedback are significantly improved. By combining the latest intervention strategy network parameters with patient feedback data, an image feature response evaluation function is constructed, realizing an effective closed-loop between model feature extraction and actual intervention effects. Using the evaluation value as a supervision signal to guide the adjustment of attention distribution weights, the multi-scale vision Transformer model can focus on the regions most relevant to clinical outcomes at different spatial scales, enhancing the ability to identify abnormal structures and early lesions. Through iterative optimization of the attention distribution, the model's adaptive adjustment to new input images and continuous enhancement of feature extraction ability are achieved. Finally, the model parameter preservation and update mechanism ensures that the multi-scale vision Transformer model has the ability of long-term sustainable evolution and individualized feature learning. Overall, the present invention improves the intelligence, precision and dynamic response ability of abdominal image analysis, providing a solid technical foundation and clinical support for the early identification, risk stratification and intervention effect evaluation of intestinal obstruction.

[0036] Reference Figure 2 , an intelligent diagnosis and intervention decision-making system for postoperative intestinal obstruction of the gastrointestinal tract, includes the following modules: A data collection and preprocessing module, used to collect multi-source data of postoperative gastrointestinal patients and perform preprocessing to construct an input data set; A multi-scale vision Transformer feature extraction module, used to extract multi-scale image feature representations from the input data set; A multi-modal state fusion module, used to fuse the multi-scale image feature representations with the input data set to generate a multi-modal state representation vector; An intervention strategy network module, used to output patient intervention behavior decisions based on the trust region policy optimization algorithm and in combination with the multi-modal state representation vector; A reward and rare event discrimination module, used to construct a reward function according to the feedback data after intervention execution, and discriminate whether a rare high-risk event occurs, generating an intervention strategy network for policy optimization; A model supervision and optimization module, used to construct an image feature response evaluation function according to the intervention strategy network parameters and feedback data, and use it as a supervision signal to guide the dynamic optimization of the multi-scale vision Transformer model.

[0037] Example 1: To verify the feasibility of the present invention in implementation, the present invention was applied to the general surgery department of a large tertiary class-A hospital. During the period from May 2024 to May 2025, the hospital fully deployed and actually applied the intelligent diagnosis and intervention decision-making system for postoperative intestinal obstruction of the gastrointestinal tract of the present invention, and the system ran through the entire postoperative hospitalization management process of patients undergoing gastrointestinal surgery. The annual average number of gastrointestinal surgeries in this hospital exceeds 2,000, with a large patient base and complex cases, which is an advanced application scenario in this field.

[0038] Taking a 65-year-old male patient after colorectal cancer surgery as an example, on the third day after surgery, the patient had mild abdominal distension and reduced food intake. In the past, relying on manual judgment by doctors, repeated physical examinations, flipping through paper or scattered electronic records, and waiting for imaging consultations were required, and the average diagnosis and intervention decision-making cycle was 8 to 12 hours. Nowadays, the system automatically collects multi-source data such as the patient's abdominal CT, B-ultrasound, body temperature, heart rate, respiration, blood pressure, blood oxygen, diet, and activities. The data preprocessing module realizes automatic standardization and time series alignment. The multi-scale vision Transformer feature extraction module models the multi-level information of abdominal images, and combines with the multi-modal state fusion module to generate an individual state representation. The intervention strategy network real-time evaluates the patient's risk of intestinal obstruction, and the system recommends to the doctor "restrict diet and closely monitor", and continuously optimizes the intervention plan according to the feedback.

[0039] When the patient had significantly increased abdominal distension, elevated body temperature, and the abdominal image showed intestinal dilation and the appearance of air-fluid levels on the fifth day after surgery, the system automatically identified it as a high-risk state and immediately pushed the suggestions of "nasogastric decompression and emergency intravenous rehydration". After the intervention was carried out, the patient's symptoms improved significantly, and no intestinal perforation or emergency surgery was required. Compared with before the system was enabled, the doctor said: "The condition was detected earlier, treated faster, and the decision-making was well-documented, which greatly reduced the duty pressure and medical risks." During the operation of the system, the hospital respectively counted the core indicators such as the early identification rate of intestinal obstruction, the conversion rate to severe cases, the average length of hospital stay, and the incidence rate of intervention adverse events of postoperative gastrointestinal patients during the period from 2023 (when the system was not applied) to 2024 (when the system was applied).

[0040] Table 1 Comparison table of key indicators of postoperative intestinal obstruction patients in the gastrointestinal tract before and after the application of the intelligent system

[0041] According to the statistical results of the data in Table 1, it can be seen that the intelligent diagnosis and intervention system for postoperative intestinal obstruction of the gastrointestinal tract of the present invention has achieved remarkable clinical results after practical application. During the period from 2023 (when the system was not applied) to 2024 (when the system was applied), the number of patient cases was comparable, both approaching 2,000, so the data is highly comparable. After the application of the system, the early missed diagnosis rate of intestinal obstruction decreased significantly from 9.0% to 2.1%, effectively avoiding the delay of the condition caused by untimely early identification. The rate of progression to severe cases also decreased from 4.6% to 1.2%, indicating that through intelligent risk assessment and dynamic intervention, the risk of patients progressing to severe cases can be significantly reduced.

[0042] In terms of the average length of hospital stay, it was shortened from 13.1 days to 9.5 days, fully reflecting the optimization effect of the system on the timing of intervention and individualized management, helping patients recover earlier and saving medical resources. The incidence of adverse events during intervention decreased from 2.3% to 0.7%, further proving that the intelligent intervention suggestions are safer and can reduce complications and secondary injuries. In addition, the proportion of doctors adopting the system's suggestions is as high as 92%, indicating that the scientific nature and clinical recognition of the system's recommendations are very high. Particularly worthy of attention is that the early warning rate of rare high-risk events by the system increased from 35% to 100%, greatly enhancing the protection ability for critically ill patients. The patient satisfaction rate also increased from 88% to 97%, indicating that the system has optimized the medical experience and treatment confidence.

[0043] Overall, these data clearly prove that the present invention can improve the diagnosis and treatment quality and safety of patients after gastrointestinal surgery, not only reducing missed diagnoses, severe case conversions, and the occurrence of complications, but also optimizing the hospital stay process, obtaining high recognition from both doctors and patients, and showing extremely strong clinical application value and promotion prospects.

[0044] As mentioned above, it is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, makes equivalent substitutions or changes, and should be covered within the protection scope of the present invention.

Claims

1. An intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract, characterized in that, It includes the following steps: S1. Collect multi-source data of patients after gastrointestinal surgery, preprocess the multi-source data, and construct an input data set; S2. Construct a multi-scale vision Transformer model, extract abdominal image data from the input data set, input it into the multi-scale vision Transformer model, and extract multi-scale image feature representations; S3. Integrate the multi-scale image feature representations with the vital sign data and postoperative behavior record data in the input data set to construct a patient multi-modal state representation vector; S4. Input the patient's multi-modal state representation vector into an intervention strategy network optimized based on the trust region policy, set a maximum KL divergence limit threshold, and generate an intervention behavior decision; S5. Implement corresponding intervention measures for the patient according to the intervention behavior decision, collect feedback data after the intervention is executed, calculate the policy reward value between the intervention behavior and the result, and use 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; S6. Process the patient's multi-modal state representation vector with the optimized intervention strategy network, continuously output new intervention behavior decisions, and realize the dynamic update and application cycle of the intervention strategy; S7. According to the parameters of the intervention strategy network and the feedback data, construct an image feature response evaluation function, use it as a supervision signal to input into the multi-scale vision Transformer model, and optimize the multi-scale vision Transformer model.

2. The intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract according to claim 1, wherein The multi-source data specifically includes heterogeneous information such as endoscopic images, abdominal image data, vital sign data, postoperative behavior record data, physiological parameters, and electronic medical record texts of postoperative gastrointestinal patients.

3. The intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract according to claim 1, wherein The preprocessing of the multi-source data specifically includes image normalization, text tokenization and encoding, physiological parameter standardization, and data time alignment.

4. The intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract according to claim 1, characterized in that, The specific content of S2 includes: S21. Extract the abdominal image data of the abdominal region from the input dataset to construct the original image tensor , where represents the image height, represents the image width, represents the number of image channels, and R is the set of real numbers; S22. Downsample the obtained original image tensor using a multi-scale pyramid strategy to obtain a set of abdominal image tensors with multiple different spatial resolutions , where represents the abdominal image tensor at the s-th spatial resolution, and represents the total number of scales of the abdominal image tensors in the multi-scale pyramid structure; S23. For each abdominal image tensor at each scale obtained , divide it according to a window of a fixed size to obtain a set of non-overlapping windows , and calculate the self-attention features within each window using a local self-attention mechanism, where 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 total number of windows into which the abdominal image tensor is divided at the s-th scale; S24. For all window regions obtained at each scale, use the local self-attention mechanism to extract the feature representations of each window respectively, and splice the features of all windows in the original spatial order to obtain the encoded feature map at the scale. ; S25. For the encoded feature maps at all scales, fuse the encoded feature maps at each scale by concatenating them along the channels and performing weighted summation to obtain a fused multi-scale image feature representation; S26. Input the fused multi-scale image feature representation into the dual-branch collaborative architecture and input it into 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 features related to the discrimination of intestinal obstruction risk; S27. In the decoder of the structure decoding branch hierarchical deconvolution and skip connection mechanisms are adopted to perform multi-level upsampling and spatial feature fusion on the abdominal structure feature information in the hierarchical order from deep to shallow, and an abdominal structure prediction map is reconstructed; S28. In the auxiliary discrimination branch the features related to intestinal obstruction risk discrimination are processed using a fully connected layer and an activation function to output an auxiliary risk discrimination result; S29. Further introduce a sliding window local attention mechanism in the multi-scale image feature representation to cover each region of the feature map through a sliding window and calculate local self-attention within each window; S210. During the training process of the multi-scale vision Transformer model, use a structural loss function and a functional loss function for weighted combination, and optimize the parameters of the multi-scale vision Transformer model by minimizing the combined loss function of the weighted combination. The structural loss function is constructed using a pixel-level cross-entropy loss function, and the functional loss function is constructed using a binary cross-entropy loss function. Finally, by setting a weighting coefficient, linearly weight and combine the structural loss function and the functional loss function to form a combined loss function; S211. During the training process, use the high-resolution feature branch to perform knowledge distillation on the low-resolution feature branch by minimizing the KL divergence of the output distributions of the high- and low-resolution branches; S212. After the multi-scale vision Transformer model is trained, save all the parameter configurations of the multi-scale vision Transformer model, and input the newly acquired abdominal image data into the trained multi-scale vision Transformer model to obtain the finally output multi-scale image feature representation.

5. The intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract according to claim 1, wherein The specific steps of S3 are as follows: S31. Extract the vital sign data and postoperative behavior record data at the same time point as the abdominal image data from the input dataset. The vital sign data includes the time series monitoring parameters of body temperature, heart rate, respiratory rate, blood pressure, and blood oxygen saturation. The postoperative behavior record data includes the key behavior time series of the patient's exhaust, defecation, eating, and activities. S32. Normalize and time-align the extracted vital sign data and postoperative behavior record data to construct a multi-modal structured feature tensor , where represents the number of time steps, represents the number of clinical feature dimensions, is the set of real numbers; S33. Perform temporal index alignment on the obtained multi-scale image feature representations so that the image features and clinical data correspond at the same time step; S34. Feature - level fusion is performed on the aligned multi - scale image feature representation and the obtained multi - modal structured feature tensor to obtain the patient multi - modal state representation vector by means of splicing ; S35. Batch process the patient's multi-modal state representation vectors at all time steps to form the patient's multi-modal state representation vector for input to the intervention strategy network.

6. The intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract according to claim 1, characterized in that The specific steps of S4 are as follows: S41. Input the obtained patient multi-modal state representation vector as the current input state into the intervention policy network optimized based on trust region policy; S42. Set an intervention behavior set in the intervention strategy network , where represents the th intervention behavior and is the total number of intervention behaviors; S43. Initialize the parameters of the intervention strategy network , and use a deep neural network to represent the patient's multi-modal state vector as the input to the intervention strategy network, and obtain the probability distribution results of each intervention behavior through the forward calculation of the intervention strategy network; S44. Set a trust region constraint to limit the maximum KL divergence threshold for each update of the policy distribution. ; S45. During the policy optimization process, embed medical prior constraints, and incorporate the clinical medical prior knowledge of medical guidelines, surgical taboos, and expert experience into the policy optimization process in the form of constraint terms. For intervention behaviors that do not conform to medical norms or pose significant medical risks, reduce the weight in the probability distribution of the intervention strategy network. S46. Adopt a trust region dynamic adaptive mechanism to dynamically adjust the KL divergence threshold according to the signal function :[[]]END]]​ ； S47. Introduce a multi-objective hierarchical trust region optimization mechanism. During the optimization process of the intervention strategy network, refine the intervention objectives into multiple hierarchical objectives, including maximizing the curative effect, minimizing the safety risk, and optimizing medical resources. Establish independent trust region constraints for each objective, and balance the policy outputs under different objectives through a hierarchical aggregation mechanism. S48. During the training and optimization process of the intervention strategy network, events with historical occurrence probabilities lower than a preset threshold and corresponding risk scores higher than a preset threshold are defined as rare high-risk event samples. For rare high-risk event samples, an empirical memory pool is set up and the sampling weight of rare high-risk event samples in the training samples is increased. During the policy update process, a more stringent trust region constraint is applied to the training batches involving rare high-risk events to strengthen the convergence constraint; S49. Collect the feedback after the patient's intervention is executed and construct a reward function : ； Among them, represents the clinical feedback reward mapping function, which dynamically adjusts the reward weight according to the individual responses of patients, and divides the reward signals into different levels such as short-term effects, long-term outcomes, and risk penalties; at the same time, a feedback traceability and explanation mechanism is introduced in the process of generating reward signals to label and trace the key contributing factors of the reward signals, improving the personalization, dynamics, and interpretability of the reward feedback, and providing richer feedback information support for the continuous optimization and precise 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 patient's current actual intervention, and collect the closed-loop effect feedback and perform dynamic policy update.

7. The intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract according to claim 1, characterized in that, The specific steps of S5 are as follows: S51. According to the intervention behavior decision output by the intervention strategy network, implement corresponding medical intervention measures for the patient. The medical intervention measures include conservative observation, diet control, drug treatment, gastrointestinal decompression, imaging review, or surgical treatment. S52. Collect the feedback data after each implementation of the intervention measure. The feedback data includes the symptom manifestations, vital sign parameters, laboratory test indicators, and adverse event information of the patient, and form a feedback data set. ; S53. Preprocess the feedback data set to obtain the intervention effect change feature vector , and construct a comprehensive reward function based on the intervention effect change feature vector : ； Among them, is the clinical feedback reward mapping function; S54. Input the comprehensive reward function and the rare high - risk event indication function as feedback signals into the intervention policy network. 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 appears in the feedback data, the rare high - risk event indication function takes a value of 1; otherwise, it takes a value of 0. Combine with the current state - action pair to perform policy optimization, where represents the 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; S55. The trust region strategy optimization algorithm is adopted for parameter update, and the optimization objective is ; S56. Continuously collect and archive each intervention and its feedback data, and dynamically adjust the reward function parameters and rare event weight parameters. S57. Loop through the above steps S51 to S56 to achieve continuous iterative update of the intervention strategy network based on the reward mechanism and dynamic policy optimization.

8. The intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract according to claim 1, wherein The specific steps of S7 are as follows: S71. According to the parameters of the latest intervention strategy network and the feedback data after the patient's intervention execution, combined with the current fused multi-scale image feature representation , construct an image feature response evaluation function ; S72. Input the image feature response evaluation function as a supervision signal into the multi-scale vision Transformer model as the basis for adjusting the attention distribution weights, and optimize the attention weight parameters of the multi-scale vision Transformer model at different spatial scales. S73. During the optimization process of the multi-scale vision Transformer model, iteratively update the attention distribution weights of each scale Transformer encoding layer. When the multi-scale vision Transformer model processes new input abdominal image data, it can adaptively enhance the feature extraction ability for key regions highly relevant to the feedback. Save the parameters of the optimized multi-scale Vision Transformer model to form a multi-scale Vision Transformer model that can be continuously and adaptively updated. During the analysis of abdominal imaging data, continuously use the image feature response evaluation function and the attention guidance mechanism to achieve dynamic supervision and optimization of the multi-scale Vision Transformer model and improvement of the individual feature extraction ability.

9. An intelligent diagnosis and intervention decision-making system for postoperative intestinal obstruction of the gastrointestinal tract, which is applied to the intelligent diagnosis and intervention decision-making method for postoperative intestinal obstruction of the gastrointestinal tract described in any one of claims 1 to 8, and is characterized in that, It includes the following modules: Data acquisition and preprocessing module, which is used to acquire multi-source data of postoperative gastrointestinal patients and perform preprocessing to construct an input data set; Multi-scale Vision Transformer feature extraction module, which is used to extract multi-scale image feature representations from the input data set; Multi-modal state fusion module, which is used to fuse the multi-scale image feature representations with the input data set to generate a multi-modal state representation vector; Intervention strategy network module, which is used to output the patient's intervention behavior decision based on the trust region policy optimization algorithm and combined with the multi-modal state representation vector; Reward and rare event discrimination module, which is used to construct a reward function according to the feedback data after the intervention is executed, and discriminate whether a rare high-risk event occurs, and generate an intervention strategy network for policy optimization; Model supervision and optimization module, which is used to construct an image feature response evaluation function according to the intervention strategy network parameters and feedback data, and use it as a supervision signal to guide the dynamic optimization of the multi-scale Vision Transformer model.

Images