Prediction system and method for postoperative cognitive impairment of gastrointestinal tumor surgery patient

Abstract

The invention provides a postoperative cognitive impairment prediction system and method for gastrointestinal tumor surgery patients, and belongs to the technical field of postoperative cognitive impairment prediction of patients based on machine learning. The system comprises a data receiving and obtaining module, an abdomen CT feature extraction module, a patient clinical and information data coding module and a gastrointestinal tumor patient postoperative cognitive impairment prediction module. The data receiving and obtaining module extracts three-dimensional features of the abdomen image data to obtain an abdomen CT image fusion feature map; the patient clinical and information data coding module performs feature coding on the patient clinical data and the information data to obtain clinical feature vectors; and the gastrointestinal tumor patient postoperative cognitive impairment prediction module takes the abdominal CT image fusion feature map and the clinical feature vector as input and outputs a quantitative evaluation result of the postoperative cognitive impairment of the patient. According to the method, effective classification of cognitive impairment is realized, and powerful auxiliary support is provided for clinical intervention and management of postoperative patients.

Description

Technical Field

[0001] The present invention belongs to the technical field of prediction of postoperative cognitive dysfunction in patients based on machine learning, and particularly relates to a prediction system and method for postoperative cognitive dysfunction in patients undergoing gastrointestinal tumor surgery. Background Art

[0002] Postoperative cognitive dysfunction in patients undergoing gastrointestinal tumor surgery has attracted much attention due to its high incidence and clinical harm. Its occurrence is closely related to the intestinal-brain axis disorder caused by surgical stress, the neurotoxicity of chemotherapy drugs (such as oxaliplatin), and metabolic abnormalities. Cognitive dysfunction is a common neurological complication in patients undergoing gastrointestinal tumor surgery, especially in elderly patients, mainly manifested as memory loss, decline in executive function, attention disorder, and impaired social ability. Studies have shown that the incidence of cognitive dysfunction within 1 week after non-cardiac surgery is relatively high, and some patients still have cognitive impairment 3 months after surgery. This cognitive decline not only prolongs the hospital stay and increases medical expenses, but also may accelerate the process of postoperative dementia and even increase the long-term mortality rate, posing a severe challenge to the quality of life of patients, the burden of family care, and the social medical system.

[0003] At present, the prevention and treatment strategies for postoperative cognitive dysfunction in patients undergoing gastrointestinal tumor surgery are mainly divided into two categories: drug intervention and non-drug intervention. Non-drug measures include optimizing the surgical method (such as minimally invasive technology), maintaining intraoperative organ perfusion, refined anesthesia depth management, strengthening postoperative analgesia, and early rehabilitation training. Drug intervention involves the application of brain protection drugs such as dexmedetomidine, anti-inflammatory drugs, and neuro-metabolic regulators. Although these measures can reduce the risk of cognitive dysfunction to a certain extent, their effects are still limited by individual differences and the complexity of pathological mechanisms. Therefore, establishing a precise risk prediction system to achieve early identification of high-risk patients has become the core direction for optimizing the management of cognitive dysfunction.

[0004] The prediction methods for postoperative cognitive dysfunction in patients undergoing gastrointestinal tumor surgery are mainly divided into four categories: clinical assessment tools, biomarker detection, imaging techniques, and artificial intelligence models. Clinical assessment tools include neuropsychological test batteries (such as HVLT, TMT, MoCA, etc.) and risk prediction models (such as the ISPOCD model). Although they are authoritative evaluation indicators, they are time-consuming and greatly affected by the clinical experience of doctors. Biomarker detection covers blood (IL-6, NfL, etc.), cerebrospinal fluid (Aβ42 / tau), and gene (APOE ε4) analysis, which can objectively reflect the risk of nerve injury. However, some of the detections are invasive and costly. Imaging techniques reveal brain structure / function abnormalities through MRI, fMRI, etc. The emerging near-infrared spectroscopy (NIRS) can also monitor the brain oxygen status in real time during surgery. However, the above methods have limitations such as strong subjectivity in clinical assessment, insufficient specificity of biomarkers, and lack of timeliness.

[0005] The artificial intelligence model integrates multi-modal data (clinical indicators + MRI + EEG), and uses algorithms such as random forest and deep learning to improve the prediction efficiency. It has been widely used in the prediction of postoperative cognitive dysfunction and achieved results consistent with the evaluation of clinical experts. However, the existing technologies have the following problems: Over-reliance on brain imaging data: Existing artificial intelligence models mainly integrate neuroimaging data such as brain MRI and EEG, ignoring the correlation between the structural changes of abdominal organs and cognitive impairment, resulting in the inability to use the non-invasive data source of routine preoperative abdominal CT of gastrointestinal tumor patients for prediction; Inefficient multi-modal fusion mechanism: Existing methods for fusing imaging and clinical data mostly use simple splicing or shallow feature interaction, which are difficult to capture cross-modal high-order associations (such as the synergistic effect between intraoperative drug dosage and tumor morphology), restricting the model's expressive ability; Insufficient generalization ability: The model relies on brain-specific detections (such as EEG), which are limited by the popularity of equipment and patient cooperation, and are difficult to promote in primary hospitals; moreover, prior knowledge of organ anatomy is not introduced, and it is sensitive to image noise. Summary of the Invention

[0006] In view of the above problems, the first aspect of the present invention provides a prediction system for postoperative cognitive dysfunction in gastrointestinal tumor surgery patients, including a data reception and acquisition module, an abdominal CT feature extraction module, a patient clinical and information data encoding module, and a prediction module for postoperative cognitive dysfunction in gastrointestinal tumor patients; The data reception and acquisition module is used to receive and acquire the patient's abdominal imaging data, patient clinical data, and patient information data; The abdominal CT feature extraction module includes a tumor region feature extraction network, an organ-level local feature extraction network, and a multi-level feature fusion network; the tumor region feature extraction network performs three-dimensional convolutional feature extraction on the patient's preoperative abdominal imaging to capture the deep information of the tumor in terms of spatial distribution, density, and morphology, and obtains a tumor morphology feature map; the organ-level local feature extraction network takes the patient's preoperative abdominal imaging and the patient's postoperative abdominal imaging as inputs, segments the stomach, colon, and rectum, and obtains a key organ spatial position feature map; the multi-level feature fusion network is used to fuse the tumor morphology feature map and the key organ spatial position feature map to obtain an abdominal CT image fusion feature map; The patient clinical and information data encoding module performs feature encoding on the patient's clinical data and information data to obtain a clinical feature vector; The prediction module for postoperative cognitive dysfunction in gastrointestinal tumor patients takes the abdominal CT image fusion feature map and the clinical feature vector as inputs and outputs a quantitative evaluation result of the patient's postoperative cognitive dysfunction.

[0007] Preferably, the abdominal imaging data includes the patient's preoperative abdominal imaging and the postoperative abdominal images of the patient ; the clinical data of the patient includes the perioperative related data of patients with gastric organ tumors, colon organ tumors and rectal organ tumors, specifically the preoperative neuropsychological assessment results of the patient 、the postoperative neuropsychological assessment results of the patient 、anesthesia method 、surgical method 、the intraoperative drug dosage 、time-related parameters and brain function monitoring indicators ; the patient information data includes the patient's past medical history information 、patient attribute indicators 、gender, height, weight, education level, and the current description of the patient's gastrointestinal tumor condition .

[0008] Preferably, the tumor region feature extraction network is an improved architecture based on 3D ResNet-50, including a pre-trained 3D Mask R-CNN layer, a dynamic density histogram equalization layer, a residual attention backbone network, and a fully connected layer; the th preoperative abdominal image of the patient is input into the pre-trained 3D Mask R-CNN layer to locate the tumor region in, and is cropped into a tumor region cube; the tumor region cube is input into the dynamic density histogram equalization layer to enhance the contrast of the tumor boundary in the tumor region cube, obtaining an enhanced tumor region cube; the enhanced tumor region cube is input into the residual attention backbone network to obtain the tumor morphology depth features, and the residual attention backbone network includes four residual blocks, a 3D attention layer, and a global adaptive average pooling layer connected in sequence. The number of channels of the four residual blocks connected in sequence is 64, 128, 256, and 512 respectively. Each residual block contains 3 3×3×3 convolutional layers for extracting tumor region features layer by layer. A 3D attention layer is connected behind each residual block, and the 3D attention layer is used to perform weighted fusion on the input features and output features of each residual block. The global adaptive average pooling layer is used to compress the feature dimension of the output of the last residual block through the 3D attention layer, obtaining a dimensionality-reduced feature; the dimensionality-reduced feature is input into the fully connected layer for dropout operation to obtain the final tumor morphology feature map.

[0009] Preferably, the organ-level local feature extraction network is an improved architecture based on 3D nnU-Net, including an image registration layer, a pre-trained 3D-Unet model, and a transformer encoder; the specific data processing process is as follows: The th postoperative abdominal image of the patient Input image registration layer and the patient's preoperative abdominal image Image registration is performed. The image registration layer extracts SIFT feature points using the Scale-Invariant Feature Transform algorithm and the key matching points of the SIFT feature points of are used to calculate the affine transformation matrix. Using the affine transformation matrix, the pixel points of are mapped to the pixel point coordinate system of to achieve and cross-modal spatial registration, and the th postoperative abdominal registered image of the patient is obtained ; Using the pre-trained 3D-Unet model, the th postoperative abdominal registered image of the patient is semantically segmented to obtain the th patient's gastric feature map, colon feature map and rectal feature map; The th patient's gastric feature map, colon feature map and rectal feature map are respectively input into the transformer encoder to obtain the th patient's gastric organ spatial position feature map, colon organ spatial position feature map and rectal organ spatial position feature map; The transformer encoder is used to model the spatial context relationship inside the organ to obtain more localized spatial feature information that is more relevant to the risk of cognitive impairment.

[0010] Preferably, the multi-level feature fusion network fuses the th patient's tumor morphological feature map, gastric organ spatial position feature map, colon organ spatial position feature map and rectal organ spatial feature map; The multi-level feature fusion network includes a feature alignment layer, a channel attention weighting layer and a first feature splicing layer; The feature alignment layer reduces the dimension of the tumor morphological feature map; The feature alignment layer includes a fully connected layer and a 3D transposed convolution; The channel attention weighting layer takes the spatially aligned tumor morphological feature map and organ feature map as inputs; The channel attention weighting layer includes a global average pooling layer, two fully connected layers and a sigmoid activation function layer; The first feature splicing layer takes the channel-weighted tumor morphological feature map, gastric organ spatial position feature map, colon organ spatial position feature map and rectal organ spatial feature map as inputs, and performs horizontal feature splicing to obtain an abdominal CT image fusion feature map with 1024 channels.

[0011] Preferably, the patient clinical and information data encoding module includes a CLIP model, a numerical encoding model and a second feature splicing layer; Input the text data in the patient's clinical and patient information data into the CLIP model, project different types of structured texts into the same latent semantic space to obtain text semantic feature vectors; First input the numerical data in the patient's clinical and patient information data into a numerical encoding model to convert the numerical data into numerical feature vectors; then input the numerical feature vectors into the encoder of the CLIP model to convert the numerical feature vectors into numerical semantic feature vectors with the same feature dimension as the text semantic feature vectors; The second feature concatenation layer horizontally concatenates the numerical semantic feature vectors and the text semantic feature vectors to obtain clinical feature vectors.

[0012] Preferably, the postoperative cognitive dysfunction prediction module for gastrointestinal tumor patients includes a global average pooling layer, a dual-head self-attention mechanism network, and a cognitive dysfunction prediction network; The global average pooling layer reduces the dimension of the abdominal CT image fusion feature map, and calculates the average value of all spatial positions in each channel of the abdominal CT image fusion feature map to obtain an abdominal CT image fusion feature vector; The dual-head self-attention mechanism network takes the abdominal CT image fusion feature vector and the clinical feature vector as inputs; the dual-head self-attention mechanism network includes a linear mapping layer, a residual feature concatenation layer, and a dual-head self-attention mechanism layer; input the abdominal CT image fusion feature vector into the linear mapping layer to map it into a query vector and a key vector , input the clinical feature vector into the linear mapping layer to map it into a value vector ; secondly, input the clinical feature vector into the linear mapping layer to map it into a query vector and a key vector , input the abdominal CT image fusion feature vector into the linear mapping layer to map it into a value vector ; again, input ; ; and ; ; into the dual-head self-attention mechanism layer at the same time to obtain the weight of the abdominal CT image fusion feature vector relative to the clinical feature vector and the weight of the clinical feature vector relative to the abdominal CT image fusion feature vector ; again, perform a dot product of the abdominal CT image fusion feature vector and the weight to obtain a weighted abdominal CT image fusion feature vector, and perform a dot product of the clinical feature vector and the weight Perform dot product to obtain a weighted clinical feature vector; finally, add the weighted abdominal CT image fusion feature vector and the weighted clinical feature vector, and then input them into the residual feature splicing layer together with the abdominal CT image fusion feature vector and the clinical feature vector for horizontal feature splicing to obtain the fusion feature; The cognitive dysfunction prediction network includes three fully connected layers, three dropout layers, and one softmax activation function layer; where each fully connected layer is followed by a dropout layer. The fully connected layer is used to reduce the dimension of the fusion feature, and the dropout layer is used to freeze the parameters of some neurons in the fully connected layer. The fusion feature processed by the last fully connected layer is input into the RELU activation function layer to output the postoperative cognitive function score of the patient.

[0013] Preferably, use the organ prior constraint training strategy to train the 3D-Unet model in the abdominal CT feature extraction module. The specific process is as follows: Rely on the prior knowledge of the stomach, colon, and rectum images provided by the public dataset to obtain the average shape feature map of the stomach organ, the average shape feature map of the colon organ, and the average shape feature map of the rectum organ; first, for an image data sample in the public dataset, extract the three-dimensional masks of the stomach, colon, and rectum in the image, and use Procrustes analysis to align all organ instances to the same coordinate system to eliminate image rotation and translation differences; secondly, perform vertex-based non-rigid registration on each organ, and vectorize the registered shapes of the stomach, colon, and rectum organs to obtain the shape feature map of the stomach organ, the shape feature map of the colon organ, and the shape feature map of the rectum organ; thirdly, use the principal component analysis method to calculate the shape principal components of the shape feature map of the stomach organ, the shape feature map of the colon organ, and the shape feature map of the rectum organ respectively, and retain the first k principal components that explain 95% of the shape variance to obtain the reduced-dimensional shape feature map of the stomach organ, the reduced-dimensional shape feature map of the colon organ, and the reduced-dimensional shape feature map of the rectum organ; finally, perform the above steps on all the stomach organs, colon organs, and rectum organs in the public dataset, and input all the reduced-dimensional shape feature maps of the stomach organs, the reduced-dimensional shape feature maps of the colon organs, and the reduced-dimensional shape feature maps of the rectum organs into the average pooling layer to obtain the average shape feature map of the stomach organ, the average shape feature map of the colon organ, and the average shape feature map of the rectum organ; During the training process, for the stomach feature map, colon feature map, and rectum feature map obtained by semantic segmentation of the postoperative abdominal registered images of the patient by the 3D-Unet model in the abdominal CT feature extraction module, calculate the Frobenius norm between the stomach feature map and the average shape feature map of the stomach organ the Frobenius norm between the colon feature map and the average shape feature map of the colon organ and the Frobenius norm between the rectum feature map and the average shape feature map of the rectum organ and adopt the loss function Fine-tune the parameters of the 3D-Unet model: ; wherein is the scale of the training set.

[0014] Preferably, the binary cross-entropy loss function is used to perform overall training on the abdominal CT feature extraction module, the patient clinical and information data encoding module, and the postoperative cognitive dysfunction prediction module for gastrointestinal cancer patients. During the training process, the 3D-Unet model and the CLIP model are frozen, and only other neural network structures outside the 3D-Unet model and the CLIP model are trained. The loss function is the mean root mean square error between the patient cognitive impairment prediction result output by the postoperative cognitive dysfunction prediction module for gastrointestinal cancer patients and the true result.

[0015] The second aspect of the present invention provides a method for predicting postoperative cognitive dysfunction in patients undergoing gastrointestinal cancer surgery, deploying the postoperative cognitive dysfunction prediction system for gastrointestinal cancer patients as described in the first aspect on a detection terminal; and including the following processes: Real-time obtain the standard input data of the patient's abdominal image, patient clinical conditions, and patient information; Input the above standard input data into the deployed postoperative cognitive dysfunction prediction system for gastrointestinal cancer patients; Output the prediction score, and the doctor comprehensively diagnoses whether the patient has postoperative cognitive dysfunction based on the postoperative cognitive dysfunction prediction result of the gastrointestinal cancer patient and clinical experience.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. Deeply integrate preoperative abdominal CT image features (such as pancreatic fat infiltration, intestinal wall thickness variation) with perioperative clinical data (anesthesia method, drug dosage, cerebral oxygen monitoring, etc.), breaking through the limitation of a single data source. Dynamically weight different modal features through a cross-modal attention mechanism to enhance the feature expression ability; 2. The improved 3D ResNet-50 network captures the tumor spatial distribution / morphological features, combines the nnU-Net to segment key organs, and improves the anatomical structure recognition accuracy through organ prior constraint training; 3. The feature interaction network realizes the deep coupling of images and clinical data, avoiding fusion noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is the overall technical route flowchart of the present invention.

[0018] Figure 2 is the structural diagram of the abdominal CT feature extraction module of the present invention.

[0019] Figure 3This is the structural diagram of the clinical and information data coding module for the patients of the present invention.

[0020] Figure 4 This is the structural diagram of the prediction module for postoperative cognitive dysfunction in patients with gastrointestinal tumors of the present invention.

[0021] Figure 5 This is the display diagram of the tumor region segmentation effect in the embodiment of the present invention.

[0022] Figure 6 This is the confusion matrix result diagram in the embodiment of the present invention.

[0023] Figure 7 This is the ROC curve result diagram in the embodiment of the present invention. Detailed implementation manners

[0024] The present invention utilizes deep learning technology to fully utilize the abdominal organ image information, physiological data and cognitive status of patients during the perioperative period, and proposes a method for predicting postoperative cognitive dysfunction in patients undergoing gastrointestinal tumor surgery. The overall process is as Figure 1 shown: First, collect the perioperative abdominal imaging data, clinical data and information data of patients with gastric cancer, colon cancer and rectal cancer. Use the preoperative neuropsychological assessment results and postoperative neuropsychological assessment results in the clinical data to evaluate the cognitive behavior of patients to obtain the postoperative cognitive function scores of patients as the labels of the data set, and use the abdominal imaging data, the remaining clinical data and information data as data samples to construct a training set and a test set. Secondly, construct an abdominal CT feature extraction module to extract the three-dimensional features of the abdominal imaging data to obtain an abdominal CT image fusion feature map, which covers the global image features, tumor morphological features and key organ spatial position features. Thirdly, construct a patient clinical and information data coding module to perform feature coding on the patient clinical data and information data to obtain a clinical feature vector. Thirdly, construct a prediction module for postoperative cognitive dysfunction in patients with gastrointestinal tumors, and use the abdominal CT image fusion feature map and the clinical feature vector as inputs to achieve a quantitative evaluation result of the postoperative cognitive dysfunction of patients. Finally, design a multi-modal model training strategy to train the abdominal CT feature extraction module, the patient clinical and information data coding module and the prediction module for postoperative cognitive dysfunction in patients with gastrointestinal tumors.

[0025] The invention will be further described below in conjunction with specific embodiments.

[0026] I. Dataset construction First, collect the pre-operative and post-operative clinical information, abdominal CT images, intraoperative data, and neuropsychological scale results of gastric cancer, colon cancer, and rectal cancer patients at different stages (early, middle, and late), and perform preprocessing and privacy protection. Secondly, experts annotate the cognitive behavior of patients according to the results of the neuropsychological scale, including normal cognition and cognitive impairment. Use the pre-operative clinical information, post-operative clinical information, abdominal CT images, and intraoperative data of patients as samples, and use the post-operative cognitive function scores of patients as labels to obtain a completely annotated dataset; Construct a data system for postoperative cognitive dysfunction in gastrointestinal tumor surgery patients covering multi-source heterogeneous information, including three types: patient abdominal imaging data, patient clinical data, and patient information data. The patient abdominal imaging data is the patient's pre-operative abdominal imaging and the patient's post-operative abdominal imaging ; The patient clinical data includes peri-operative related data of patients with gastric organ tumors, colon organ tumors, and rectal organ tumors. Specifically, the patient's pre-operative neuropsychological assessment results (MMSE score), the patient's post-operative neuropsychological assessment results (MMSE score), anesthesia method (general anesthesia or combined epidural block), surgical method (open abdomen or laparoscopic surgery), intraoperative drug dosage (dexmedetomidine dosage, remifentanil dosage, propofol dosage), time-related parameters (surgery time and anesthesia time) and brain function monitoring indicators (duration of intraoperative cerebral oxygen saturation (rSO2) below 35 and duration of its relative change amplitude ΔrSO2 greater than 13%); The patient information data includes the patient's previous medical history information (whether suffering from hypertension and diabetes), patient attribute indicators (age (0 represents male, 1 represents female), gender, height and weight, education level (0 represents below bachelor's degree, 1 represents bachelor's degree, 2 represents postgraduate degree)) and the description of the patient's current gastrointestinal tumor condition . Through systematic collection, collation, and privacy desensitization processing of the above multi-source heterogeneous data, a complete and high-quality dataset can be constructed, providing a solid data foundation for feature extraction of subsequent models and risk prediction of postoperative cognitive impairment; Experts evaluate the cognitive behavior of patients based on the patient's pre-operative neuropsychological assessment results and the patient's post-operative neuropsychological assessment results to obtain the patient's post-operative cognitive function score. The score ranges from 1 to 10 points, and the higher the score, the higher the degree of cognitive impairment of the patient; The patient's pre-operative abdominal imaging , Postoperative abdominal images of the patient , Anesthesia method , Surgical method , Dosage of drugs used during the operation , Time-related parameters (Surgical time and anesthesia time), Brain function monitoring indicators , Past medical history information of the patient , Patient attribute indicators and the current description of the patient's gastrointestinal tumor condition Using the above as data samples and the patient's cognitive behavior assessment category as a label to construct a training set and a test set; in order to improve the generalization ability of the model, the patient data collected in the present invention covers the perioperative data of three types of patients, namely gastric cancer patients, colon cancer patients, and rectal cancer patients, at different stages (early, middle, and late).

[0027] II. Construct an abdominal CT feature extraction module The abdominal CT feature extraction module includes a tumor region feature extraction network, an organ-level local feature extraction network, and a multi-level feature fusion network, as Figure 2 shown; among them, the tumor region feature extraction network performs three-dimensional convolutional feature extraction on the patient's preoperative abdominal images to capture the deep information of the tumor in terms of spatial distribution, density, and morphology, and obtain a tumor morphology feature map; the patient's preoperative abdominal images and the patient's postoperative abdominal images are input into the organ-level local feature extraction network to segment the stomach, colon, and rectum, and obtain a key organ spatial position feature map; the multi-level feature fusion network is used to fuse the tumor morphology feature map and the key organ spatial position feature map to obtain an abdominal CT image fusion feature map; the specific steps for extracting the abdominal CT three-dimensional features of the rd patient using the abdominal CT feature extraction module are as follows: 1. The tumor region feature extraction network is an improved architecture based on 3D ResNet-50, including a pre-trained 3D Mask R-CNN layer, a dynamic density histogram equalization layer, a residual attention backbone network, and a fully connected layer; the rd patient's preoperative abdominal images are input into the pre-trained 3D Mask R-CNN layer to locate the tumor region in Crop the tumor region into a cube of 64×64×32; input the tumor region cube into the dynamic density histogram equalization layer to enhance the contrast of the tumor boundary in the tumor region cube, obtaining an enhanced tumor region cube; input the enhanced tumor region cube into the residual attention backbone network to obtain the tumor morphology depth features. The residual attention backbone network includes four residual blocks, a 3D attention layer, and a global adaptive average pooling layer connected in sequence. The number of channels of the four residual blocks connected in sequence are 64, 128, 256, and 512 respectively. Each residual block contains 3 3×3×3 convolutional layers for extracting tumor region features layer by layer. A 3D attention layer is connected behind each residual block, and the 3D attention layer is used to perform weighted fusion of the input features and output features of each residual block to prevent gradient explosion during model training. The global adaptive average pooling layer is used to compress the feature dimension of the output of the last residual block after passing through the 3D attention layer, obtaining a dimensionality-reduced feature; input the dimensionality-reduced feature into the fully connected layer for dropout operation to obtain the final tumor morphology feature map (H×W×D×512), where H, W, and D are the length, width, and depth of the tumor morphology feature map respectively; this feature map not only covers three static structure information of tumor volume, edge morphology, and texture distribution, but also effectively retains the organ-level spatial hierarchy and the relationship between tissues, and can describe the overall abdominal structure and the tumor region distribution pattern, providing a global information basis for subsequent cognitive impairment risk prediction; 2. The organ-level local feature extraction network is an improved architecture based on 3D nnU-Net, including an image registration layer, a pre-trained 3D-Unet model, and a transformer encoder, specifically as follows: 1) Input the postoperative abdominal image of the th patient into the image registration layer to perform image registration with the preoperative abdominal image of the patient . The image registration layer uses the scale-invariant feature transform algorithm to extract the key matching points of the SIFT feature points of and the SIFT feature points of . Calculate the affine transformation matrix using the key matching points, and use the affine transformation matrix to map the pixel points of to the pixel point coordinate system of to achieve cross-modal spatial registration between and , obtaining the postoperative abdominal registered image of the th patient ; 2) Use the pre-trained 3D-Unet model on the TotalSegmentator public dataset to perform semantic segmentation on the postoperative abdominal registered image of the th patient to obtain the Gastric feature maps, colonic feature maps, and rectal feature maps of a patient; The 3D-Unet model pre-trained on the TotalSegmentator public dataset has strong transfer learning and adaptive modeling capabilities, enabling high-precision segmentation results across different anatomical structures; 3) Use a transformer encoder to perform organ-level local feature extraction on the gastric feature maps, colonic feature maps, and rectal feature maps of the patient. The Transformer encoder is used to model the spatial context relationship inside the organ to obtain more localized spatial feature information that is more relevant to the risk of cognitive impairment; Input the gastric feature maps, colonic feature maps, and rectal feature maps of the patient into the transformer encoder respectively to obtain the gastric organ spatial position feature map (H1×W1×D1×256), colonic organ spatial position feature map (H1×W1×D1×256), and rectal organ spatial position feature map (H1×W1×D1×256) of the patient; H1, W1, and D1 are the length, width, and depth of the feature map output by the Transformer encoder respectively; 3. Construct a multi-level feature fusion network to fuse the tumor morphological feature map, gastric organ spatial position feature map, colonic organ spatial position feature map, and rectal organ spatial feature map of the patient; The multi-level feature fusion network includes a feature alignment layer, a channel attention weighting layer, and a feature horizontal splicing layer. The specific steps are as follows: 1) Since the dimension of the tumor morphological feature map is H×W×D×256, and the dimension of the organ feature map is H1×W1×D1×256, in order to avoid introducing noise due to dimension mismatch when fusing the tumor morphological feature map and the organ feature map, it is necessary to align the dimensions of the tumor morphological feature map and the organ feature map. Therefore, a feature alignment layer is constructed to reduce the dimension of the tumor morphological feature map; The feature alignment layer includes a fully connected layer and a 3D transposed convolution; First, input the tumor morphological feature map into the fully connected layer to reduce its dimension to a tumor morphological feature map with 256 channels (H×W×D×256), and then use a 3D transposed convolution (kernel size 3×3×3, stride 1) to convert the tumor morphological feature map with 256 channels into the same spatial size as the organ feature map to obtain the spatially aligned tumor morphological feature map (H1×W1×D1×256); 2) Input the spatially aligned tumor morphological feature map and organ feature map into the channel attention weighting layer; the channel attention weighting layer includes a global average pooling layer, two fully connected layers, and a sigmoid activation function layer; first, input the spatially aligned tumor morphological feature map, gastric organ spatial position feature map, colonic organ spatial position feature map, and rectal organ spatial feature map into the global average pooling layer respectively, perform global average pooling on each channel in the feature map to obtain the tumor morphological feature vector, gastric organ spatial position feature vector, colonic organ spatial position feature vector, and rectal organ spatial feature vector; secondly, input the tumor morphological feature vector, gastric organ spatial position feature vector, colonic organ spatial position feature vector, and rectal organ spatial feature vector into two fully connected layers and then into the sigmoid activation function layer to obtain the channel weights of the tumor morphological feature map, gastric organ spatial position feature map, colonic organ spatial position feature map, and rectal organ spatial feature map; the sigmoid activation function layer restricts the weights within [0,1] to avoid gradient explosion; multiply the spatially aligned tumor morphological feature map, gastric organ spatial position feature map, colonic organ spatial position feature map, and rectal organ spatial feature map by their corresponding channel attention weights respectively to obtain the channel-weighted tumor morphological feature map, channel-weighted gastric organ spatial position feature map, channel-weighted colonic organ spatial position feature map, and channel-weighted rectal organ spatial feature map; 3) Input the channel-weighted tumor morphological feature map, gastric organ spatial position feature map, colonic organ spatial position feature map, and rectal organ spatial feature map into the feature splicing layer for horizontal feature splicing to obtain an abdominal CT image fusion feature map with 1024 channels, which covers the global image features, tumor morphological features, and key organ spatial position features.

[0028] III. Construct a patient clinical and information data encoding module Construct a patient clinical and information data encoding module to perform feature encoding on patient clinical data and patient information data, as Figure 3 shown; the patient clinical and information data encoding module includes a CLIP model, a numerical encoding model, and a feature splicing layer; the CLIP model is a multi-modal pre-trained neural network that can effectively capture the potential correlations and semantic shifts in language information; the specific steps for using the patient clinical and information data encoding module to perform feature encoding on the th patient's clinical data and information data are as follows: 1. Use the CLIP model to process the anesthesia method surgical method previous medical history information and current gastrointestinal tumor condition description After performing feature encoding, tokenization, and embedding, the input is fed into the CLIP model. The CLIP model projects different types of structured text into the same latent semantic space to obtain text semantic feature vectors. 2. Since the traditional CLIP model cannot directly process numerical data, the present invention designs a numerical encoding model to convert numerical data into numerical feature vectors. The numerical encoding model includes a normalization layer and the encoder of a pre-trained BERT model. The intraoperative drug dosage , time-related parameters , brain function monitoring indicators , and patient attribute indicators are input into the normalization layer for data normalization and then input into the encoder of the pre-trained BERT model to obtain numerical feature vectors. 3. The numerical feature vectors are input into the encoder of the CLIP model to convert them into numerical semantic feature vectors with the same feature dimension as the text semantic feature vectors, enabling them to be fused in the same feature space as the text semantic feature vectors. 4. The numerical semantic feature vectors and the text semantic feature vectors are input and horizontally concatenated to obtain clinical feature vectors.

[0029] IV. Constructing a prediction module for postoperative cognitive dysfunction in gastrointestinal tumor patients Constructing a prediction module for postoperative cognitive dysfunction in gastrointestinal tumor patients to achieve a quantitative evaluation result of the patient's postoperative cognitive dysfunction. The model structure is as Figure 4 shown. The prediction module for postoperative cognitive dysfunction in gastrointestinal tumor surgery patients includes an interactive feature fusion network and a cognitive dysfunction prediction network. The fused feature map of the abdominal CT image and the clinical feature vector of the th patient are input into the prediction module for postoperative cognitive dysfunction in gastrointestinal tumor patients. The specific steps are as follows: To more fully explore the potential relationship between the fused features of abdominal CT images and clinical features, the present invention designs an interactive feature fusion network. Its core lies in a dual-head self-attention structure, which can reflect the relationship between any two positions of the input feature sequence in a weighted form, thereby capturing the deep coupling effect between different features. The interactive feature fusion network includes a global average pooling layer and a dual-head self-attention mechanism network. Since the fused features of abdominal CT images are 3D feature maps while the clinical feature vectors are vectors, in order not to introduce additional noise during the fusion process, the present invention uses a global average pooling layer to reduce the dimension of the fused feature map of abdominal CT images, calculating the average value of all spatial positions in each channel of the fused feature map of abdominal CT images to obtain the fused feature vector of abdominal CT images. Input the abdominal CT image fusion feature vector and the clinical feature vector into the dual-head self-attention mechanism network; the dual-head self-attention mechanism network includes a linear mapping layer, a residual feature splicing layer, and a dual-head self-attention mechanism layer; first, input the abdominal CT image fusion feature vector into the linear mapping layer to map it into a query vector and a key vector , input the clinical feature vector into the linear mapping layer to map it into a value vector ; secondly, input the clinical feature vector into the linear mapping layer to map it into a query vector and a key vector , input the abdominal CT image fusion feature vector into the linear mapping layer to map it into a value vector ; thirdly, input ; ; and ; ; into the dual-head self-attention mechanism layer simultaneously to obtain the weight of the abdominal CT image fusion feature vector relative to the clinical feature vector and the weight of the clinical feature vector relative to the abdominal CT image fusion feature vector ; thirdly, perform a dot product of the abdominal CT image fusion feature vector and the weight to obtain a weighted abdominal CT image fusion feature vector, perform a dot product of the clinical feature vector and the weight to obtain a weighted clinical feature vector; finally, add the weighted abdominal CT image fusion feature vector and the weighted clinical feature vector and input them into the residual feature splicing layer together with the abdominal CT image fusion feature vector and the clinical feature vector for horizontal feature splicing to obtain a fused feature, ensuring that the model balances deep interaction and information fidelity; Input the fused feature into the postoperative cognitive dysfunction prediction network of gastrointestinal tumor patients to obtain the postoperative cognitive dysfunction prediction result of the th patient; the postoperative cognitive dysfunction prediction network of gastrointestinal tumor patients includes three fully connected layers, three dropout layers, and one softmax activation function layer; among them, each fully connected layer is followed by a dropout layer. The fully connected layer is used to reduce the dimension of the fused feature, and the dropout layer is used to freeze the parameters of some neurons in the fully connected layer to prevent the model from exploding in gradients during the training process. Input the fused feature processed by the last fully connected layer into the RELU activation function layer to output the postoperative cognitive function score of the patient.

[0030] V. Training Strategy Design a multi-modal model training strategy to train the abdominal CT feature extraction module, the patient clinical and information data encoding module, and the postoperative cognitive dysfunction prediction module of gastrointestinal tumor patients; the multi-modal model training strategy includes an organ prior constraint training strategy and a cross-entropy training strategy, specifically as follows: Train the 3D-Unet model in the abdominal CT feature extraction module using the designed organ prior constraint training strategy. The specific process is as follows: Rely on the prior knowledge of gastric, colon, and rectal images provided by the public dataset AMOS to obtain the average shape feature maps of the gastric organ, colon organ, and rectal organ. First, for an image data sample in the public dataset, extract the three-dimensional masks of the stomach, colon, and rectum in the image, and use Procrustes analysis to align all organ instances to the same coordinate system to eliminate image rotation and translation differences. Second, perform vertex-based non-rigid registration on each organ, and vectorize the registered shapes of the stomach, colon, and rectum organs to obtain the shape feature maps of the gastric organ, colon organ, and rectal organ. Third, use the principal component analysis method to calculate the shape principal components of the shape feature maps of the gastric organ, colon organ, and rectal organ respectively, and retain the first k principal components that explain 95% of the shape variance to obtain the reduced-dimensional shape feature maps of the gastric organ, colon organ, and rectal organ. Finally, perform the above steps on all gastric, colon, and rectal organs in the public dataset AMOS, and input all the reduced-dimensional shape feature maps of the gastric organ, colon organ, and rectal organ into the average pooling layer to obtain the average shape feature maps of the gastric organ, colon organ, and rectal organ; During the training process, for the gastric feature map, colon feature map, and rectal feature map obtained by semantic segmentation of the postoperative abdominal registration images of patients using the 3D-Unet model in the abdominal CT feature extraction module, calculate the Frobenius norm between the gastric feature map and the average shape feature map of the gastric organ the Frobenius norm between the colon feature map and the average shape feature map of the colon organ and the Frobenius norm between the rectal feature map and the average shape feature map of the rectal organ and use the loss function to fine-tune the parameters of the 3D-Unet model: ; where is the size of the training set; The cross entropy loss function was used to train the abdominal CT feature extraction module, the patient clinical and information data encoding module, and the gastrointestinal tumor patient postoperative cognitive dysfunction prediction module as a whole. The 3D-Unet model and the CLIP model were frozen during the training process, and only the other neural network structures other than the 3D-Unet model and the CLIP model were trained. The loss function was the average root mean square error between the patient cognitive dysfunction prediction results output by the gastrointestinal tumor patient postoperative cognitive dysfunction prediction module and the actual results. Doctors combined the gastrointestinal tumor patient postoperative cognitive dysfunction prediction results and clinical experience to diagnose whether the patient had postoperative cognitive dysfunction.

[0031] 6. Experimental Results The present invention cooperated with a number of medical centers to collect abdominal imaging data, clinical data and information data of patients with gastrointestinal tumors during the perioperative period, and performed unified preprocessing and privacy protection on them; among them, there were 398 samples of patients with gastric organ tumors, 486 samples of patients with colon organ tumors and 438 samples of patients with rectal organ tumors, with a total of 1322 samples in the data set. Subsequently, the patient's postoperative cognitive function score was calculated based on the neuropsychological scale or expert diagnostic criteria to form a well-labeled data set to facilitate subsequent feature extraction and model training. The 1322 samples were divided into a training set of 1058 cases and a test set of 264 cases. Among them, there were 601 / 457 cases with no cognitive dysfunction / cognitive dysfunction in the training set. There were 133 / 131 cases with no cognitive dysfunction / cognitive dysfunction in the test set.

[0032] Figure 5 The effect of using the segmentation model to segment intestinal tumors is demonstrated. The fine-tuned TotalSegmentator can accurately segment gastrointestinal tumors.

[0033] In order to better evaluate and optimize the overall performance of the model, the present invention selects a series of commonly used classification and prediction indicators for quantitative evaluation. First, based on the matching of the true label and the predicted result, the model can measure the accuracy of its classification of the sample as a whole, and use the recall rate to evaluate the completeness of identifying patients undergoing gastrointestinal tumor surgery, and combine the precision rate or F1 score to balance the recall rate and precision rate. In addition, constructing the ROC curve and calculating the area under it helps to demonstrate the classification performance of the model under different risk thresholds, providing an objective basis for the selection of thresholds. Through the comprehensive analysis of these evaluation indicators, the present invention can continuously adjust the model structure and optimize the training process, so that it can accurately identify the risk of postoperative cognitive dysfunction in patients undergoing gastrointestinal tumor surgery while minimizing false positives and negatives, providing more reliable decision support for clinical applications.

[0034] Table 1 Model prediction effect display

[0035] A total of 264 cases of postoperative patients' multimodal data were included in this invention for evaluating the prediction model of postoperative cognitive dysfunction in patients. Among them, 133 patients had no cognitive dysfunction and 131 patients had cognitive dysfunction, and the label distribution was relatively balanced. The overall performance was good, initially verifying the feasibility and effectiveness of doctors in the prediction task of postoperative cognitive dysfunction in patients with the assistance of the model.

[0036] As Figure 6 shown, from the confusion matrix, among the 133 patients with no cognitive dysfunction, doctors with relatively shallow clinical experience successfully identified 111 cases with the assistance of the model, and 22 cases were misjudged as having cognitive dysfunction; among the 131 patients with cognitive dysfunction, 108 cases were correctly identified as having cognitive dysfunction, and 23 cases were misjudged as not having cognitive dysfunction. This result indicates that the recognition ability of clinicians in the two categories is relatively balanced with the assistance of the model, and no obvious bias occurs.

[0037] Further analyzing from the perspective of performance indicators: the precision rate of the non-cognitive dysfunction category is 0.828, the recall rate is 0.835, and the F1 score is 0.831; while the precision rate of the cognitive dysfunction category is 0.831, the recall rate is 0.824, and the F1 score is 0.828. The macro-average F1 score is 0.829, indicating that clinicians maintain a relatively consistent performance between the two categories with the assistance of the model. The weighted average F1 score is also 0.829, indicating that this performance is also stable in the overall sample and is not significantly affected by class imbalance.

[0038] As Figure 7 shown, in terms of probability prediction, the area under the ROC curve drawn is 0.90, indicating that the model has strong discrimination ability. The overall trend of the ROC curve is relatively ideal, suggesting that the model can maintain a good balance between sensitivity and specificity at different probability thresholds.

[0039] Based on the above indicators, the constructed prediction model of postoperative cognitive dysfunction in patients undergoing gastrointestinal tumor surgery shows high accuracy and robustness on this experimental dataset, and has good potential in early identifying patients with cognitive dysfunction. With the assistance of this model, clinicians not only achieved effective classification of cognitive dysfunction, but also maintained the recognition balance in the two types of samples, which can provide strong auxiliary support for the clinical intervention and management of postoperative patients. Follow-up studies can further expand the sample size, improve the generalization ability of the model, and further optimize and interpret it in combination with clinical characteristics.

[0040] This method not only shows higher accuracy and stability in experimental evaluations, but also has good clinical promotion value. It can provide important references for doctors in preoperative stratified management and postoperative intervention decision-making, and promote the prediction of cognitive dysfunction in patients after gastrointestinal cancer surgery towards practicality and intelligence.

[0041] The above are only the preferred embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

[0042] Although the specific implementation manners of the present invention have been described above, they do not limit the protection scope of the present invention. Those skilled in the art should understand that based on the technical solutions of the present invention, various modifications or deformations that can be made without creative efforts by those skilled in the art are still within the protection scope of the present invention.

Claims

1. A prediction system for postoperative cognitive dysfunction in patients undergoing gastrointestinal tumor surgery, characterized in that, It includes a data reception and acquisition module, an abdominal CT feature extraction module, a patient clinical and information data encoding module, and a prediction module for postoperative cognitive dysfunction in gastrointestinal tumor patients; The data reception and acquisition module is used to receive and acquire patient abdominal imaging data, patient clinical data, and patient information data; The abdominal CT feature extraction module includes a tumor region feature extraction network, an organ-level local feature extraction network, and a multi-level feature fusion network; the tumor region feature extraction network performs three-dimensional convolutional feature extraction on the patient's preoperative abdominal imaging to capture the deep information of the tumor in terms of spatial distribution, density, and morphology, and obtains a tumor morphology feature map; the organ-level local feature extraction network takes the patient's preoperative abdominal imaging and the patient's postoperative abdominal imaging as inputs, segments the stomach, colon, and rectum, and obtains a key organ spatial position feature map; the multi-level feature fusion network is used to fuse the tumor morphology feature map and the key organ spatial position feature map to obtain an abdominal CT image fusion feature map; The patient clinical and information data encoding module performs feature encoding on the patient's clinical data and information data to obtain a clinical feature vector; The prediction module for postoperative cognitive dysfunction in gastrointestinal tumor patients takes the abdominal CT image fusion feature map and the clinical feature vector as inputs and outputs a quantitative evaluation result of the patient's postoperative cognitive dysfunction.

2. The postoperative cognitive dysfunction prediction system for gastrointestinal tumor surgery patients according to claim 1, wherein: The abdominal imaging data includes the patient's preoperative abdominal imaging and the patient's postoperative abdominal imaging ; The patient's clinical data includes the perioperative related data of patients with gastric organ tumors, colon organ tumors, and rectal organ tumors, specifically the patient's preoperative neuropsychological assessment results the patient's postoperative neuropsychological assessment results anesthesia method surgical method intraoperative drug dosage time-related parameters and brain function monitoring indicators ; The patient information data includes the patient's past medical history information patient attribute indicators gender, height, weight, education level, and the patient's current description of the gastrointestinal tumor condition .

3. The postoperative cognitive dysfunction prediction system for gastrointestinal tumor surgery patients according to claim 1, characterized in that: The tumor region feature extraction network is an improved architecture based on 3D ResNet-50, including a pre-trained 3D Mask R-CNN layer, a dynamic density histogram equalization layer, a residual attention backbone network, and a fully connected layer; the preoperative abdominal images of the th patient are input into the pre-trained 3D Mask R-CNN layer to locate the tumor region in the, and the is cropped into a tumor region cube; the tumor region cube is input into the dynamic density histogram equalization layer to enhance the contrast of the tumor boundary in the tumor region cube, obtaining an enhanced tumor region cube; the enhanced tumor region cube is input into the residual attention backbone network to obtain the depth features of the tumor morphology. The residual attention backbone network includes four sequentially connected residual blocks, a 3D attention layer, and a global adaptive average pooling layer. The number of channels of the four sequentially connected residual blocks are 64, 128, 256, and 512 respectively. Each residual block contains 3 3×3×3 convolutional layers for extracting tumor region features layer by layer. A 3D attention layer is connected behind each residual block. The 3D attention layer is used to perform weighted fusion of the input features and output features of each residual block. The global adaptive average pooling layer is used to compress the feature dimension of the output of the last residual block through the 3D attention layer, obtaining a dimensionality-reduced feature; the dimensionality-reduced feature is input into the fully connected layer for dropout operation to obtain the final tumor morphology feature map.

4. The postoperative cognitive dysfunction prediction system for gastrointestinal tumor surgery patients according to claim 3, characterized in that: The organ-level local feature extraction network is an improved architecture based on 3D nnU-Net, including an image registration layer, a pre-trained 3D-Unet model, and a transformer encoder; the specific data processing process is as follows: Input the postoperative abdominal image of the th patient into the image registration layer and perform image registration with the preoperative abdominal image of the patient . The image registration layer extracts the SIFT feature points using the Scale-Invariant Feature Transform (SIFT) algorithm and the key matching points of the SIFT feature points of . Calculate the affine transformation matrix using the key matching points, and use the affine transformation matrix to map the pixel points of to the pixel point coordinate system of , achieving and cross-modal spatial registration to obtain the postoperative abdominal registered image of the th patient ; ​ Using the pre-trained 3D-Unet model, for the postoperative abdominal registered image of the patient, perform semantic segmentation to obtain the gastric feature map, colonic feature map and rectal feature map of the patient; Input the gastric feature map, colonic feature map, and rectal feature map of the th patient into the Transformer encoder respectively to obtain the gastric organ spatial location feature map, colonic organ spatial location feature map, and rectal organ spatial location feature map of the th patient; the Transformer encoder is used to model the spatial context relationship inside the organ to obtain more localized spatial feature information that is more relevant to the risk of cognitive impairment.

5. The postoperative cognitive dysfunction prediction system for gastrointestinal tumor surgery patients according to claim 4, characterized in that: The multi-level feature fusion network fuses the tumor morphological feature map, the gastric organ spatial position feature map, the colonic organ spatial position feature map, and the rectal organ spatial feature map of the th patient; the multi-level feature fusion network It includes a feature alignment layer, a channel attention weighting layer, and a first feature splicing layer; The feature alignment layer reduces the dimension of the tumor morphology feature map; the feature alignment layer includes a fully connected layer and a 3D transposed convolution; The channel attention weighting layer takes the spatially aligned tumor morphology feature map and the organ feature map as inputs; the channel attention weighting layer includes a global average pooling layer, two fully connected layers, and a sigmoid activation function layer; The first feature splicing layer takes the channel-weighted tumor morphology feature map, the stomach organ spatial position feature map, the colon organ spatial position feature map, and the rectum organ spatial feature map as inputs and performs horizontal feature splicing to obtain an abdominal CT image fusion feature map with 1024 channels.

6. The postoperative cognitive dysfunction prediction system for gastrointestinal tumor surgery patients according to claim 1, characterized in that: The patient clinical and information data encoding module includes a CLIP model, a numerical encoding model, and a second feature splicing layer; The text data in the patient's clinical and patient information data is input into the CLIP model, and different types of structured text are projected into the same latent semantic space to obtain a text semantic feature vector; The numerical data in the patient's clinical and patient information data is first input into the numerical encoding model to convert the numerical data into a numerical feature vector; then the numerical feature vector is input into the encoder of the CLIP model to convert the numerical feature vector into a numerical semantic feature vector with the same feature dimension as the text semantic feature vector; The second feature splicing layer performs horizontal feature splicing on the numerical semantic feature vector and the text semantic feature vector to obtain a clinical feature vector.

7. A prediction system for postoperative cognitive dysfunction in patients undergoing gastrointestinal tumor surgery according to claim 1, characterized in that: The postoperative cognitive dysfunction prediction module for gastrointestinal tumor patients includes a global average pooling layer, a dual-head self-attention mechanism network, and a cognitive dysfunction prediction network; The global average pooling layer reduces the dimension of the abdominal CT image fusion feature map, and calculates the average value of all spatial positions in each channel of the abdominal CT image fusion feature map to obtain the abdominal CT image fusion feature vector; The dual-head self-attention mechanism network takes the abdominal CT image fusion feature vector and the clinical feature vector as inputs; the dual-head self-attention mechanism network includes a linear mapping layer, a residual feature splicing layer, and a dual-head self-attention mechanism layer; the abdominal CT image fusion feature vector is input into the linear mapping layer to be mapped into a query vector and a key vector , and the clinical feature vector is input into the linear mapping layer to be mapped into a value vector ; secondly, the clinical feature vector is input into the linear mapping layer to be mapped into a query vector and a key vector , and the abdominal CT image fusion feature vector is input into the linear mapping layer to be mapped into a value vector ; again, ; ; and ; ; are simultaneously input into the dual-head self-attention mechanism layer to obtain the weight of the abdominal CT image fusion feature vector relative to the clinical feature vector and the weight of the clinical feature vector relative to the abdominal CT image fusion feature vector ; again, the abdominal CT image fusion feature vector is dot-multiplied with the weight to obtain a weighted abdominal CT image fusion feature vector, and the clinical feature vector is dot-multiplied with the weight to obtain a weighted clinical feature vector; finally, the weighted abdominal CT image fusion feature vector and the weighted clinical feature vector are added and then input into the residual feature splicing layer together with the abdominal CT image fusion feature vector and the clinical feature vector for horizontal feature splicing to obtain a fused feature; The cognitive dysfunction prediction network includes three fully connected layers, three dropout layers, and one softmax activation function layer; a dropout layer is connected after each fully connected layer. The fully connected layer is used to reduce the dimension of the fusion features, and the dropout layer is used to freeze the parameters of some neurons in the fully connected layer. The fusion features processed by the last fully connected layer are input into the RELU activation function layer to output the postoperative cognitive function score of the patient.

8. The postoperative cognitive dysfunction prediction system for gastrointestinal tumor surgery patients according to claim 1, characterized in that: Use the organ prior constraint training strategy to train the 3D-Unet model in the abdominal CT feature extraction module. The specific process is as follows: Rely on the prior knowledge of the stomach, colon, and rectal images provided by the public dataset to obtain the average shape feature map of the stomach organ, the average shape feature map of the colon organ, and the average shape feature map of the rectal organ; first, for an image data sample in the public dataset, extract the three-dimensional masks of the stomach, colon, and rectum in the image, and use Procrustes analysis to align all organ instances to the same coordinate system to eliminate image rotation and translation differences; second, perform vertex-based non-rigid registration on each organ, and vectorize the registered stomach, colon, and rectal organ shapes to obtain the stomach organ shape feature map, the colon organ shape feature map, and the rectal organ shape feature map; third, use the principal component analysis method to calculate the shape principal components of the stomach organ shape feature map, the colon organ shape feature map, and the rectal organ shape feature map respectively, and retain the first k principal components that explain 95% of the shape variance to obtain the reduced-dimensional stomach organ shape feature map, the colon organ shape feature map, and the rectal organ shape feature map; finally, perform the above steps on all the stomach organs, colon organs, and rectal organs in the public dataset, and input all the reduced-dimensional stomach organ shape feature maps, reduced-dimensional colon organ shape feature maps, and reduced-dimensional rectal organ shape feature maps into the average pooling layer to obtain the average shape feature map of the stomach organ, the average shape feature map of the colon organ, and the average shape feature map of the rectal organ; During the training process, for the gastric feature map, colon feature map, and rectal feature map obtained by semantic segmentation of the postoperative abdominal registration images of patients using the 3D-Unet model in the abdominal CT feature extraction module, calculate the Frobenius norm between the gastric feature map and the average shape feature map of the gastric organ , the Frobenius norm between the colon feature map and the average shape feature map of the colon organ and the Frobenius norm between the rectal feature map and the average shape feature map of the rectal organ , and use the loss function to fine-tune the parameters of the 3D-Unet model: ； wherein is the scale of the training set.

9. A postoperative cognitive dysfunction prediction system for gastrointestinal tumor surgery patients according to claim 1, characterized in that: The cross-entropy loss function is used to perform overall training on the abdominal CT feature extraction module, the patient clinical and information data encoding module, and the postoperative cognitive dysfunction prediction module for gastrointestinal tumor patients. During the training process, the 3D-Unet model and the CLIP model are frozen, and only other neural network structures outside the 3D-Unet model and the CLIP model are trained. The loss function is the mean root mean square error between the patient cognitive disorder prediction result output by the postoperative cognitive dysfunction prediction module for gastrointestinal tumor patients and the true result.

10. A method for predicting postoperative cognitive dysfunction in patients undergoing gastrointestinal tumor surgery, characterized in that: Deploy the postoperative cognitive dysfunction prediction system for gastrointestinal tumor surgery patients according to any one of claims 1 to 9 on the detection terminal; And It includes the following processes: Obtain the standard input data of the patient's abdominal image, patient's clinical condition and patient information in real time; Input the above standard input data into the deployed prediction system for postoperative cognitive dysfunction in patients undergoing gastrointestinal tumor surgery; Output the prediction score, and the doctor makes a diagnosis on whether the patient has postoperative cognitive dysfunction by comprehensively considering the prediction results of postoperative cognitive dysfunction in gastrointestinal tumor patients and clinical experience.

Images