Abdominal injury assessment method and system based on artificial intelligence

Abstract

The invention relates to an abdominal injury assessment method and system based on artificial intelligence. The method comprises the following steps: acquiring multi-modal data of the abdomen of a patient; wherein the multi-modal data comprises image data and clinical data; based on a Vit model, extracting image data features in the multi-modal data to obtain a structured image feature set; based on a Wesseling algorithm, clinical data features in the multi-modal data are extracted, and a streaming physiological index set is obtained; based on a cross-modal Transform model, fusing the structured image feature set and the dynamic physiological index set to obtain a multi-modal fusion decision matrix; and based on the multi-modal fusion decision matrix, carrying out parallel scoring on an abdomen region distinguished by an abdomen nine-score method to obtain an organ-specific trauma scoring table. By adopting the method, dynamic parallel scoring can be realized, and the processing speed is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of information technology assisted medical diagnosis, and particularly relates to an abdominal injury assessment method and system based on artificial intelligence. BACKGROUND

[0002] With the development of artificial intelligence technology, computer-aided abdominal injury assessment technology has emerged. Traditional computer-aided abdominal injury assessment technology is mainly realized through the following processes: first, based on medical images, image processing algorithms such as threshold segmentation and region growing are used to perform rough segmentation of organs, and morphological parameters of the damaged area are extracted in combination with artificial design features; then, a shallow machine learning model such as a support vector machine or a decision tree is used to classify the damage type; finally, a trauma grade is generated by matching the features according to a pre-defined static scoring rule. This technology relies on artificial setting of segmentation thresholds and classification rules, and has problems such as weak feature expression ability, rigid scoring, and slow processing speed. SUMMARY

[0003] Therefore, it is necessary to provide an abdominal injury assessment method and system based on artificial intelligence which can dynamically and in parallel score, greatly speeding up the processing speed.

[0004] In a first aspect, the present application provides an abdominal injury assessment method based on artificial intelligence, comprising:

[0005] obtaining multi-modal data of the patient's abdomen; wherein the multi-modal data includes image data and clinical data;

[0006] extracting image data features in the multi-modal data based on a Vit model to obtain a structured image feature set;

[0007] extracting clinical data features in the multi-modal data based on a Wesseling algorithm to obtain a streaming physiological indicator set;

[0008] fusing the structured image feature set and the dynamic physiological indicator set based on a cross-modal Transformer model to obtain a multi-modal fusion decision matrix;

[0009] based on the multi-modal fusion decision matrix, scoring the abdominal regions divided by the abdominal nine-part method in parallel to obtain an organ-specific trauma score table.

[0010] Further, based on the multi-modal fusion decision matrix, the abdominal regions divided by the abdominal nine-part method are scored in parallel to obtain an organ-specific trauma score table, comprising:

[0011] segmenting the multi-modal fusion decision matrix into nine abdominal regions by the abdominal nine-part method to obtain nine-part boundary coordinates;

[0012] Based on the nine-partition boundary coordinates, the volume proportion of each organ in the partition is calculated by the following formula to obtain the organ region weight distribution table:

[0013]

[0014] where w i,j is the volume proportion of the i-th organ in the j-th partition, V i,j is the volume of organ i in partition j, is the total volume of organ i in all partitions;

[0015] Based on the organ region weight distribution table and the preset model matching rule, a model is assigned to each region to obtain a matching relationship list;

[0016] Based on the matching relationship list, the model is loaded and the score of each abdominal region is calculated by the model to obtain an organ-specific trauma score table.

[0017] Further, based on the matching relationship list, the model is loaded and the score of each abdominal region is calculated by the model to obtain an organ-specific trauma score table, including:

[0018] Based on the matching relationship list, the features of the liver and spleen region are extracted by the 3D DenseNet model to obtain a liver and spleen injury feature vector;

[0019] The features of the intestinal tract region are extracted by the combined hybrid model of U-Net3D and random forest, and are spliced with the liver and spleen injury feature vector to obtain a structured feature set;

[0020] Based on the cross-modal Transformer model, the structured feature set and real-time physiological data are subjected to attention fusion processing to obtain a fusion feature vector;

[0021] Based on the matching relationship list, the fusion feature vector is assigned to the model and scored in parallel to obtain a baseline ISS score table; wherein the liver and spleen region uses a gradient boosting decision tree model for scoring, and the intestinal tract region uses a combined model of LSTM model and logistic regression for scoring;

[0022] Based on the continuous physiological data, the score in the baseline ISS score table is adjusted by the following formula to obtain an organ-specific trauma score table:

[0023]

[0024] where S final is the adjusted score, S n is the baseline ISS score, a is the blood loss sensitivity coefficient, is the hemoglobin change rate.

[0025] Further, based on the continuous physiological data, the score of the baseline ISS score table is adjusted in real time to obtain an organ-specific trauma score table, which further comprises:

[0026] Based on the preset anatomical logic rule, the organ-specific trauma score table is verified to obtain an anatomical verification report;

[0027] Based on the total blood loss rule, the organ-specific trauma score table is verified by the following formula to obtain a physiological verification report:

[0028]

[0029] Wherein a is the theoretical blood loss amount, which is compared with the actual blood loss amount calculated by the decrease of hemoglobin to verify the ISS i is the regional score, and each ISS classification corresponds to a baseline blood loss amount of 150ml;

[0030] The anatomical verification report and the physiological verification report are combined to obtain a regional conflict report;

[0031] Based on the regional conflict report, regions with a confidence greater than 85% and no conflict mark are extracted to obtain a list of trusted regions; the list of trusted regions is used to generate a digital signature to ensure audit tracking;

[0032] Based on the regional conflict report, a high-resolution model is called to reevaluate and verify the regions with a confidence lower than the threshold in the regional conflict report to obtain a review report;

[0033] Based on the review report and the list of trusted regions, a comprehensive trauma report is generated.

[0034] Further, based on the cross-modal Transformer model, the structured image feature set and the dynamic physiological index set are fused to obtain a multi-modal fusion decision matrix, which comprises:

[0035] The structured image feature set and the dynamic physiological index set are time and space matched by dynamic time warping to obtain a spatio-temporal aligned multi-modal feature matrix;

[0036] The spatio-temporal aligned multi-modal feature matrix is passed through the cross-modal Transformer model to generate attention weights to obtain a cross-modal attention weight matrix;

[0037] Based on the gradient reversal layer, the dynamic weights of the cross-modal attention weight matrix are adjusted to obtain balanced multi-modal fusion features;

[0038] The organ damage probability matrix is obtained by calculating the organ damage probability corresponding to the balanced multi-modal fusion features through a multi-layer perception machine; each multi-layer perception machine is independently trained for each anatomical organ;

[0039] Based on the integrated gradient algorithm, the organ damage probability matrix and the cross-modal attention weight matrix are fused to obtain a multi-modal fusion decision matrix.

[0040] Further, based on the Wesseling algorithm, clinical data features in multi-modal data are extracted to obtain a set of streaming physiological indicators, including:

[0041] Align the time of the clinical data and the laboratory discrete data through the PTP protocol to obtain a time-aligned signal stream;

[0042] Based on the adaptive wavelet threshold method, the time-aligned signal stream is denoised to obtain a denoised physiological waveform;

[0043] Through the Wesseling algorithm, features are recognized from the denoised physiological waveform to obtain a beat-by-beat blood pressure parameter sequence; the beat-by-beat blood pressure parameter sequence is used to represent real-time blood pressure dynamic changes;

[0044] Based on the beat-by-beat blood pressure parameter sequence, based on the cross-correlation coefficient, the correlation between the blood pressure parameter and the electrocardiogram parameter is calculated to obtain a blood pressure electrocardiogram correlation coefficient;

[0045] Based on the blood pressure electrocardiogram correlation coefficient and the laboratory discrete data, a compensation parameter is calculated to obtain a set of dynamic compensation parameters;

[0046] Based on the set of dynamic compensation parameters, an XGBoost regression model is used to obtain a blood loss rate prediction value;

[0047] Integrate the beat-by-beat blood pressure parameter sequence, the blood pressure electrocardiogram correlation coefficient, the set of dynamic compensation parameters, and the blood loss rate prediction value to obtain the set of streaming physiological indicators.

[0048] Further, based on the Vit model, image data features in multi-modal data are extracted to obtain a set of structured image features, including:

[0049] Based on the ITK geometric transformation model, the image data coordinate system is unified to obtain standardized three-dimensional image volume data;

[0050] The standardized three-dimensional image volume data is segmented by a pre-trained adaptive model to obtain an organ mask;

[0051] Among them, the 3D U-Net model is used for segmentation processing for liver and spleen organs, and the nnU-Net model is used for segmentation processing for intestines, blood vessels, and kidneys;

[0052] The organ mask and the standardized three-dimensional image volume data are extracted through the ITK-SNAP feature calculation engine to obtain an organ-level traditional feature table;

[0053] The organ mask and the standardized three-dimensional image data are subjected to global semantic extraction by a Vit model to obtain an organ-level Vit feature table;

[0054] Based on principal component analysis, the organ-level traditional feature table and the organ-level Vit feature table are spliced to obtain a fusion feature vector;

[0055] The fusion feature vector is subjected to abnormality detection to obtain an abnormality detection report;

[0056] The fusion feature vector and the abnormality detection report are integrated to obtain a structured image feature set.

[0057] In a second aspect, the present application further provides an abdominal injury assessment system based on artificial intelligence, comprising:

[0058] An alignment module is configured to acquire multi-modal data of a patient's abdomen; wherein the multi-modal data comprises image data and clinical data;

[0059] An image module is configured to extract image data features in the multi-modal data based on a Vit model to obtain a structured image feature set;

[0060] A physiological module is configured to extract clinical data features in the multi-modal data based on a Wesseling algorithm to obtain a set of dynamic physiological indicators;

[0061] A fusion module is configured to fuse the structured image feature set and the set of dynamic physiological indicators based on a cross-modal Transformer model to obtain a multi-modal fusion decision matrix;

[0062] A scoring module is configured to score abdominal regions divided by abdominal nine-part method in parallel based on the multi-modal fusion decision matrix to obtain an organ-specific trauma score table.

[0063] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method provided in the first aspect of the present application when executing the computer program.

[0064] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the method provided in the first aspect of the present application.

[0065] The aforementioned AI-based abdominal injury assessment method and system acquires multimodal data from the patient's abdomen, including imaging and clinical data. Based on the Vit model, it extracts imaging features from the multimodal data to obtain a structured imaging feature set. Using the Wesseling algorithm, it extracts clinical features from the multimodal data to obtain a flow cytometry physiological indicator set. Based on the cross-modal Transformer model, it fuses the structured imaging feature set and the dynamic physiological indicator set to obtain a multimodal fusion decision matrix. Based on this multimodal fusion decision matrix, it scores abdominal regions differentiated by the nine-point abdominal classification method in parallel to obtain an organ-specific trauma scoring scale. This technique enables parallel dynamic assessment of the patient's abdominal injury, significantly improving the processing speed of injury assessment. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a flowchart of the artificial intelligence-based abdominal injury assessment method of the present invention;

[0068] Figure 2 This is a diagram of the artificial intelligence-based abdominal injury assessment system of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0070] In one embodiment, such as Figure 1 As shown, an artificial intelligence-based method for assessing abdominal injuries is provided. This embodiment illustrates the application of this method to a terminal. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0071] Step 101: Obtain multimodal data of the patient's abdomen; wherein, the multimodal data includes imaging data and clinical data.

[0072] Specifically, the terminal can connect the monitor and the laboratory information system, and collect the streaming physiological waveform and the discrete laboratory index from the monitor and the laboratory information system, and integrate the multi-modal data.

[0073] For example, the multi-modal data can include: image data: CT scan (Computed Tomography), MRI (Nuclear Magnetic Resonance Imaging), ultrasound, etc.; clinical data: hemodynamic signal, electrocardiogram signal, respiratory parameter, organ perfusion index, blood index, biochemical marker, etc.

[0074] Step 102, based on the Vit model, extracting image data features in the multi-modal data to obtain a structured image feature set.

[0075] Specifically, ViT (Vision Transformer) is a visual model based on self-attention mechanism, which is used to divide the image into blocks and extract features through global context modeling. The structured image feature set is a quantitative data table containing organ shape features and damage features. Through the ITK (Insight Segmentation and Registration Toolkit) tool, the coordinate system of different image devices is standardized, and the 3D U-Net model is used to segment the liver and spleen, and the nnU-Net model is used to segment the intestinal tract. The organ mask is cut into voxel blocks, input into the ViT model, and the spatial relationship between organs is captured through the multi-head self-attention mechanism to obtain the structured image feature set.

[0076] Step 103, based on the Wesseling algorithm, extracting clinical data features in the multi-modal data to obtain a streaming physiological index set.

[0077] Specifically, the Wesseling algorithm is a hemodynamic analysis method that extracts dynamic parameters within the cardiac cycle through beat-by-beat blood pressure waveform analysis; the streaming physiological index set can include real-time physiological parameters and laboratory-derived dynamic compensation parameters.

[0078] For example, the terminal aligns the timestamps of the monitor and laboratory data through the PTP protocol, and then performs adaptive threshold denoising on the original blood pressure waveform, retains the effective physiological frequency band of 5-20Hz, further divides the continuous blood pressure waveform into single cardiac cycle, calculates the systolic pressure variation and pulse pressure difference per beat, integrates the XGBoost (Extreme Gradient Boosting) model, blood lactic acid, and hemoglobin change rate to predict the current bleeding rate, and obtains the streaming physiological index set.

[0079] Step 104, based on the cross-modal Transformer model, the structured image feature set and the dynamic physiological indicator set are fused to obtain a multi-modal fusion decision matrix.

[0080] Specifically, the cross-modal Transformer is a multi-modal fusion architecture that aligns the feature spaces of different modalities through cross-attention mechanisms. The multi-modal fusion decision matrix is a matrix that contains organ injury probability and injury severity, where each element corresponds to a specific anatomical region. The terminal can map the ViT features and physiological indicators to a unified spatio-temporal grid, balance the gradient contributions of the two types of modalities in training through the gradient inversion layer, prevent physiological noise from dominating image features, input image features as Query and physiological features as Key / Value into the cross-attention head to calculate attention weights, output injury probability through organ-specific multi-layer perceptron, and integrate gradient algorithm to visualize key feature contribution. The multi-modal fusion decision matrix is obtained by integration.

[0081] Step 105, based on the multi-modal fusion decision matrix, the abdominal regions divided by the abdominal nine-part method are scored in parallel to obtain an organ-specific trauma score table.

[0082] Specifically, the abdominal nine-part method divides the abdomen into nine regions, each corresponding to a specific organ combination. The organ-specific trauma score table is a structured report that can include injury scores and confidence levels for each region. The liver region uses 3D DenseNet to analyze hematoma volume, the intestinal region uses U-Net3D+Random Forest to evaluate intestinal wall integrity, and the baseline score is dynamically adjusted according to the hemoglobin drop rate to obtain the organ-specific trauma score table.

[0083] The abdominal injury assessment method based on artificial intelligence provided in this embodiment acquires multi-modal data of the patient's abdomen; wherein the multi-modal data includes image data and clinical data; based on the Vit model, image data features in the multi-modal data are extracted to obtain a structured image feature set; based on the Wesseling algorithm, clinical data features in the multi-modal data are extracted to obtain a flow physiological indicator set; based on the cross-modal Transformer model, the structured image feature set and the dynamic physiological indicator set are fused to obtain a multi-modal fusion decision matrix; based on the multi-modal fusion decision matrix, the abdominal regions divided by the abdominal nine-part method are scored in parallel to obtain an organ-specific trauma score table. Through the above technical means, the patient's abdominal injury can be dynamically evaluated in parallel, greatly improving the processing speed of injury assessment.

[0084] In one embodiment, based on the multi-modal fusion decision matrix, the abdominal regions divided by the abdominal nine-part method are scored in parallel to obtain an organ-specific trauma score table, including:

[0085] Step 201, divide the multi-modal fusion decision matrix into nine abdominal regions by the nine-part method to obtain the nine-part boundary coordinates.

[0086] Specifically, the nine-part boundary coordinates are a set of three-dimensional space coordinates that define the geometric range of each partition. The multi-modal fusion decision matrix is registered with the standard abdominal three-dimensional grid, the coordinate system is aligned through affine transformation, the partition boundary is generated based on anatomical landmarks, the decision matrix is cut according to the nine-part boundary to generate nine sub-matrices, and each sub-matrix contains organ injury data for the corresponding region.

[0087] For example, the abdominal nine-part method divides the abdomen into nine regions: upper abdomen: right hypochondriac region, upper abdominal region, left hypochondriac region; middle abdomen: right lumbar region, umbilical region, left lumbar region; lower abdomen: right iliac region, pubic region, left iliac region.

[0088] Step 202, based on the nine-part boundary coordinates, calculate the volume proportion of each organ in the partition by the following formula to obtain the organ region weight distribution table:

[0089]

[0090] where w i,j is the volume proportion of the i-th organ in the j-th partition, V i,j is the volume of organ i in partition j, is the total volume of organ i in all partitions.

[0091] Specifically, the organ region weight distribution table is a structured table that records the volume proportion of each organ in each partition. Based on the segmentation results, the number of voxels of each organ in each partition is calculated and converted to physical volume. If new bleeding areas are found during the operation, the volume data is corrected in real time. For important organs, set a volume threshold, such as a partition with a weight greater than 5% needs to participate in scoring, and filter irrelevant regions.

[0092] Step 203, based on the organ region weight distribution table and the pre-set model matching rules, assign a model to each region to obtain a matching relationship list.

[0093] Specifically, the matching relationship list is a structured list that records the model and its parameters that need to be loaded for each partition. If the volume proportion of a certain organ in a partition is greater than 20%, assign a dedicated model. When a partition contains multiple high-weight organs, load multiple models in parallel and set priorities, and update the model selection according to real-time data.

[0094] Step 204, load the model based on the matching relationship list and score each abdominal region through the model to obtain an organ-specific trauma score table.

[0095] Specifically, the organ-specific trauma score table is a structured report that can include the injury level of each partitioned organ, confidence. The 3D DenseNet is used to extract hematoma volume and density heterogeneity features in the liver and spleen area, and the gradient boosting decision tree is used to calculate the ISS score. The U-Net3D is used to segment the intestinal wall damage range in the intestinal tract area, and the random forest is combined with the intestinal sound feature to judge the severity of ischemia. The score is adjusted according to the real-time physiological data.

[0096] The present embodiment realizes the effect of reducing damage positioning error and greatly improving calculation efficiency through accurate grading and parallelized calculation.

[0097] In one of the embodiments, the model is loaded based on the matching relationship list, and each abdominal area is scored by the model to obtain an organ-specific trauma score table, including:

[0098] Step 301, based on the matching relationship list, the 3D DenseNet model is used to extract features in the liver and spleen area to obtain a liver and spleen injury feature vector.

[0099] Specifically, the liver and spleen injury feature vector is a high-dimensional numerical vector output by the model, representing the key attributes of liver and spleen injury. The model extracts local texture features layer by layer through three-dimensional convolution kernels, and preserves low-level details through dense skip connection to prevent information loss in deep network, obtains the feature vector encoding the anatomical-functional combined features of liver and spleen injury, and generates a feature vector sequence through continuous scanning data to capture the hematoma expansion rate.

[0100] Step 302, the U-Net3D and random forest combined hybrid model is used to extract features in the intestinal tract area, and the liver and spleen injury feature vector is spliced to obtain a structured feature set.

[0101] Specifically, the U-Net3D segment input intestinal tract area CT dynamic enhancement sequence, output three-dimensional mask and damage probability graph of intestinal wall damage area, extract morphological features; input laboratory data and physiological waveform, through random forest feature fusion, calculate intestinal ischemia risk index, splice with U-Net3D features, generate intestinal injury comprehensive feature vector; the liver and spleen area and intestinal tract area feature vectors are spliced in the channel dimension to form a structured feature set.

[0102] Step 303, based on the cross-modal Transformer model, the structured feature set and real-time physiological data are subjected to attention fusion processing to obtain a fusion feature vector.

[0103] Specifically, the cross-modal Transformer is a model based on self-attention mechanism, aligns and fuses image features and physiological time series data. The structured feature set is mapped to a query vector (Query), and the real-time physiological data is encoded into a key-value vector (Key-Value) through LSTM, the response weight of the image feature to the physiological change is calculated, and a fusion feature vector containing cross-modal spatio-temporal correlation features is output.

[0104] Step 304, based on the matching relationship list, the fusion feature vector is assigned a model parallel score to obtain a baseline ISS score table; wherein the liver and spleen area uses a gradient boosting decision tree model for scoring, and the intestinal area uses a combination model of LSTM model and logistic regression for scoring.

[0105] Specifically, the baseline ISS score table is a score table representing the severity of injury in each region of the abdomen. In the liver and spleen area, the liver and spleen related part in the input fusion feature vector is calculated by GBDT to obtain the baseline ISS score; in the intestinal area, the LSTM processes the intestinal feature sequence to output the injury progression trend, and the logistic regression maps the LSTM hidden state to the baseline ISS score.

[0106] Step 305, based on the continuous physiological data, adjust the score in the baseline ISS score table by the following formula to obtain an organ-specific trauma score table:

[0107]

[0108] Wherein, S final is the adjusted score, S n is the baseline ISS score, a is the blood loss sensitivity coefficient, is the hemoglobin change rate.

[0109] Specifically, the blood loss sensitivity coefficient is an empirical parameter, which is dynamically adjusted according to the patient's basic state, such as patient age and underlying disease, and the hemoglobin change rate reflects the real-time blood loss rate.

[0110] This embodiment realizes real-time physiological driven decision making through the closed loop of anatomical feature extraction, cross-modal fusion and dynamic physiological correction, and provides fast response intelligent decision support for abdominal injury assessment.

[0111] In one of the embodiments, after obtaining the organ-specific trauma score table by adjusting the score of the baseline ISS score table based on the continuous physiological data, it further comprises:

[0112] Step 401, based on the preset anatomical logical rule, verifying the organ-specific trauma score table to obtain an anatomical verification report.

[0113] Specifically, the anatomical logic rules are a set of constraints based on common sense of human anatomy, and the anatomical verification report is a structured document used to record the entries in the scoring table that violate the anatomical rules. Each entry in the organ-specific scoring table is compared with the rule base, and if a contradiction is detected, a conflict flag and a confidence score are generated and integrated into the anatomical verification report.

[0114] Step 402, based on the total blood loss rule, the organ-specific trauma scoring table is verified to obtain a physiological verification report by the following formula:

[0115]

[0116] Where a is the theoretical blood loss amount, which is compared with the actual blood loss amount calculated from the actual hemoglobin drop to verify the ISS i is the regional score, and each ISS classification corresponds to a baseline blood loss of 150 ml.

[0117] Specifically, the physiological verification report is used to compare the theoretical blood loss amount with the actual blood loss amount and record the difference rate. The theoretical blood loss is calculated by accumulating the ISS scores of each zone, and the actual blood loss amount is calculated by the actual hemoglobin drop rate. If the difference rate between the theoretical blood loss and the actual blood loss amount is greater than the threshold value, it is marked as a physiological conflict, and the threshold value can be adjusted according to the actual situation. When the difference rate is greater than 50%, it prompts the undiscovered source of bleeding.

[0118] Step 403, combine the anatomical verification report and the physiological verification report to obtain a regional conflict report.

[0119] Specifically, the regional conflict report is a document that integrates anatomical and physiological conflicts and can include the regions that violate anatomical rules or total blood loss rules and specific descriptions. The conflict entries in the anatomical verification report are associated and integrated with the difference entries in the physiological verification report, and the conflict level is divided according to the confidence and difference rate to generate a comprehensive conflict list.

[0120] Step 404, based on the regional conflict report, extract the regions with a confidence greater than 85% and no conflict flag to obtain a trusted region list; the trusted region list is used to generate a digital signature to ensure audit tracking.

[0121] Specifically, the trusted region list is a list of regions with high confidence and no conflict selected from the regional conflict report, which can include: region number, organ damage score, digital signature, etc. The regions with a confidence greater than 85% and no conflict flag in the regional conflict report are retained, and the content of the trusted region list is hashed and encrypted and stored in the blockchain to ensure that the audit log cannot be tampered with.

[0122] Step 405, based on the regional conflict report, call a high-resolution model to re-evaluate the regions with a confidence lower than the threshold value in the regional conflict report and perform verification to obtain a review report.

[0123] Specifically, the review report is a revised regional conflict report after re-evaluation and verification by a high-resolution model, which is a refined analysis model called for low confidence or conflict regions. The conflict region calls 1mm layer thickness CT data, uses super-resolution U-Net to re-segment, and starts continuous hemoglobin monitoring in the blood loss difference area to obtain the review report.

[0124] Step 406, based on the review report and the list of trusted regions, generate a comprehensive trauma report.

[0125] Specifically, the comprehensive trauma report is the final clinical decision document integrating the list of trusted regions, the review report and real-time physiological data. The list of trusted regions, the review report and real-time physiological data are integrated, the damage regions are sorted according to clinical priority, and the processing suggestions and basis are marked to obtain the comprehensive trauma report.

[0126] The embodiment realizes intelligent allocation of resources, reduces the probability of misdiagnosis, and improves the efficiency of allocation of computing resources and efficiency of injury assessment through anatomical-physiological dual verification, dynamic conflict processing and high-resolution review.

[0127] In one embodiment, based on a cross-modal Transformer model, a multi-modal fusion decision matrix is obtained by fusing a structured image feature set and a dynamic physiological index set, including:

[0128] Step 501, time and space matching of the structured image feature set and the dynamic physiological index set is performed by dynamic time warping to obtain a spatio-temporally aligned multi-modal feature matrix.

[0129] Specifically, the spatio-temporally aligned multi-modal feature matrix is an aligned data matrix containing the mapping relationship between image features and physiological indicators. The physiological signals are nonlinearly scaled to align the key events with the image acquisition time, and the image features and physiological signal source positions are aligned through a three-dimensional coordinate transformation matrix to obtain a multi-modal fusion decision matrix.

[0130] Step 502, the spatio-temporally aligned multi-modal feature matrix is input into the cross-modal Transformer model to generate attention weights to obtain a cross-modal attention weight matrix.

[0131] Specifically, the cross-modal attention weight matrix is a numerical matrix representing the response strength of image features to physiological changes. In the cross-modal Transformer model, Q (Query) is the image feature vector, K (Key) is the physiological feature vector, and V (Value) is the original numerical value of the physiological feature. The attention score of each image voxel with all time point physiological features is calculated to generate a weight matrix of voxel and time point products.

[0132] Step 503, for the cross-modal attention weight matrix, adjust the dynamic weight based on the gradient inversion layer to obtain the balanced multi-modal fusion feature.

[0133] Specifically, the gradient inversion layer is an adversarial training technique that reverses the gradient sign during backpropagation to balance the contribution of different modalities to model training. During training, the gradient inversion layer multiplies the physiological feature gradient by a negative coefficient to prevent physiological noise from dominating image feature learning, and automatically adjusts the inversion intensity according to the modality confidence to obtain the balanced fusion feature.

[0134] Step 504, calculate the organ damage probability corresponding to the balanced multi-modal fusion feature through a multi-layer perception, to obtain an organ damage probability matrix; each multi-layer perception is independently trained for each anatomical organ.

[0135] Specifically, the multi-layer perception is a fully connected neural network that maps high-dimensional features to damage probabilities through nonlinear transformation. The organ damage probability matrix is a matrix element representing the probability that a specific voxel belongs to a certain organ damage. Each organ uses an independent multi-layer perception, with an input dimension of the fusion feature dimension and an output dimension of 1. For each voxel, the damage probability is calculated through the multi-layer perception of the corresponding organ.

[0136] Step 505, based on the integrated gradient algorithm, fuse the organ damage probability matrix and the cross-modal attention weight matrix to obtain a multi-modal fusion decision matrix.

[0137] Specifically, the integrated gradient algorithm is an explainable AI technique that fuses multi-source data decision-making by calculating the contribution of input features to the prediction result. For each element in the organ damage probability matrix, integrate the gradient along the path from the baseline to the current feature to obtain a feature importance score. The contribution score is weighted and summed with the cross-modal attention weight to obtain a multi-modal fusion decision matrix.

[0138] This embodiment realizes precise fusion, multi-source data collaborative decision-making, and significantly improves the accuracy of injury assessment through multi-modal spatio-temporal alignment, dynamic weight balancing, and explainable fusion.

[0139] In one of the embodiments, based on the Wesseling algorithm, the clinical data features in the multi-modal data are extracted to obtain a set of streaming physiological indicators, including:

[0140] Step 601, align the time of the clinical data and laboratory discrete data through the PTP protocol to obtain time-aligned signal streams.

[0141] Specifically, the PTP protocol is a network time synchronization protocol for achieving microsecond-level time synchronization in a distributed system, ensuring the time axis consistency of multi-device data acquisition; the time-aligned signal stream is a data set obtained by unifying the timestamps of clinical data and laboratory data through the PTP protocol, eliminating the clock deviation between devices. All devices are connected to the PTP master clock, and accurate timestamps are attached during data acquisition. Linear interpolation is performed on the laboratory discrete data to generate continuous signals matching the time axis of the clinical data.

[0142] For example, the clinical data is a real-time monitored physiological signal, which can include, for example, arterial blood pressure waveform, electrocardiogram, respiratory rate, etc., and the sampling rate is usually 100-1000 Hz. The laboratory discrete data is a biochemical index obtained by blood sampling, which can include, for example, hemoglobin, lactic acid, coagulation function, etc., and the update frequency is minute to hour level.

[0143] Step 602, based on the adaptive wavelet threshold method, denoising the time-aligned signal stream to obtain the denoised physiological waveform.

[0144] Specifically, the adaptive wavelet threshold method is a denoising method based on dynamic adjustment of wavelet coefficient threshold according to local signal characteristics, which can preserve effective physiological frequency bands and suppress high-frequency noise; the denoised physiological waveform is the blood pressure, electrocardiogram, etc. signal after denoising. The Daubechies wavelet basis is used to decompose the signal into 5 layers to obtain different frequency subbands. The threshold is dynamically set according to the statistical distribution of each layer of coefficients, the coefficients below the threshold are set to zero, the effective components are preserved, the denoised signal is reconstructed by inverse transform, and the denoised physiological waveform is obtained.

[0145] Step 603, identifying features of the denoised physiological waveform by Wesseling algorithm to obtain a beat-by-beat blood pressure parameter sequence; the beat-by-beat blood pressure parameter sequence is used to represent the real-time blood pressure dynamic change.

[0146] Specifically, the beat-by-beat blood pressure parameter sequence is a set of blood pressure features in each heart beat cycle. The Wesseling algorithm is a waveform morphology-based hemodynamic analysis method that calculates beat-by-beat parameters by identifying feature points in a single heart beat cycle. The blood pressure waveform is segmented according to the R-wave of the electrocardiogram signal, the single heart beat interval is extracted, and the first derivative of the waveform is located by using the difference threshold method. The point where the first derivative changes from positive to negative and the diastolic period minimum point are located, and the pulse pressure, waveform rising slope, etc. are calculated for each heart beat cycle to form a time series.

[0147] Step 604, based on the beat-by-beat blood pressure parameter sequence, based on the cross-correlation coefficient, calculating the correlation between the blood pressure parameters and the electrocardiogram parameters to obtain the blood pressure electrocardiogram correlation coefficient.

[0148] Specifically, the cross-correlation coefficient is used to measure the linear correlation between the blood pressure parameters and the electrocardiogram parameters in the time domain. Through sliding window analysis, the correlation coefficient of the blood pressure parameter sequence and the electrocardiogram parameter sequence in the window is calculated.

[0149] In step 605, based on the blood pressure-electrocardiogram correlation coefficient and the laboratory discrete data, the compensation parameter is calculated to obtain a dynamic compensation parameter set.

[0150] Specifically, the compensation parameter is an index reflecting the compensatory ability of the body to blood loss. The compensation index is calculated by combining the beat-by-beat blood pressure parameter, the electrocardiogram parameter, and the laboratory data. The compensation parameter is recalculated every 5 seconds to generate a real-time updated parameter set.

[0151] In step 606, based on the dynamic compensation parameter set, an XGBoost regression model is used to obtain a blood loss rate prediction value.

[0152] Specifically, the XGBoost regression model is an optimized implementation of gradient boosting decision trees. It predicts continuous variables by integrating multiple regression trees. The input features can include beat-by-beat parameters, correlation coefficients, and compensation parameters. The model is trained using historical data, and the objective function is mean squared error. The blood loss rate prediction value is obtained through the trained model.

[0153] In step 607, the beat-by-beat blood pressure parameter sequence, the blood pressure-electrocardiogram correlation coefficient, the dynamic compensation parameter set, and the blood loss rate prediction value are integrated to obtain a flow physiological index set.

[0154] Specifically, the flow physiological index set is a real-time data set after integration. The output of each module is aligned by timestamp using a message queue to generate a unified data stream, which is written into a time series database for storage.

[0155] This embodiment realizes real-time abdominal injury assessment through high-precision signal processing and AI prediction model, increases the prediction accuracy of blood loss rate, and improves the accuracy of abdominal injury assessment.

[0156] In one embodiment, based on the Vit model, image data features are extracted from the multi-modal data to obtain a structured image feature set, including:

[0157] In step 701, based on the ITK geometric transformation model, the image data coordinate system is unified to obtain standardized three-dimensional image volume data.

[0158] Specifically, the standardized three-dimensional image volume data is used to unify all image data to a standard anatomical coordinate system to ensure spatial consistency across patients and modalities. The abdominal standard template is used to extract image bone landmarks as registration reference. The best transformation parameters are solved by minimizing the objective function. The original image data is applied to the transformation matrix, and the B-spline interpolation is used to generate standardized volume data to eliminate the difference in resolution between devices.

[0159] Step 702, segmenting the standardized three-dimensional image data by the pre-trained adaptive model to obtain organ masks.

[0160] Among them, the 3D U-Net model is used for segmentation processing for liver and spleen organs, and the nnU-Net model is used for segmentation processing for intestines, blood vessels and kidneys.

[0161] Specifically, the organ mask is a binary image in medical image processing, which is used to accurately mark the anatomical position of a specific organ. The input liver and spleen standardized CT arterial phase image, the 3D U-Net outputs the liver and spleen mask, and the morphological post-processing eliminates artifacts, the nnU-Net automatically adjusts the training strategy to segment the intestines, blood vessels and kidneys, and outputs the multi-organ joint mask.

[0162] Step 703, extracting features from the organ mask and standardized three-dimensional image data through the ITK-SNAP feature calculation engine to obtain an organ-level traditional feature table.

[0163] Specifically, the organ-level traditional feature table is used to record the quantitative indicators of each organ. For each voxel covered by each organ mask, the ITK-SNAP feature calculation engine calculates its shape, gray scale and texture parameters.

[0164] Step 704, performing global semantic extraction on the organ mask and standardized three-dimensional image data through the Vit model to obtain an organ-level Vit feature table.

[0165] Specifically, the organ-level ViT feature table is a high-dimensional vector representing the semantic attributes of the organ. The standardized image is divided into 16 3 size voxel blocks, each block is linearly mapped to a 256-dimensional vector, and the Transformer encoder is used to calculate the correlation between blocks to generate a global feature vector.

[0166] Step 705, based on the principal component analysis method, concatenating the organ-level traditional feature table and the organ-level Vit feature table to obtain a fusion feature vector.

[0167] Specifically, the traditional feature table and the ViT feature table are concatenated by column to form a high-dimensional original feature vector, and then the covariance matrix is calculated, the principal components retaining 95% of the variance are retained, the dimension is reduced, and the redundancy is eliminated.

[0168] Step 706, performing anomaly detection on the fusion feature vector to obtain an anomaly detection report.

[0169] Specifically, based on Isolation Forest or One-Class Support Vector Machine (OC-SVM), outliers in the feature vector are identified, the detector is trained using healthy population data, the normal feature distribution range is defined, and the abnormal score is calculated for the test sample. The threshold can be set to P less than 0.01, i.e. 99% confidence.

[0170] In step 707, the fusion feature vector and the anomaly detection report are integrated to obtain a structured image feature set.

[0171] Specifically, the fusion feature vector after dimension reduction and the anomaly detection result are stored according to organ classification, which can include a timestamp, device information and a digital signature.

[0172] The embodiment improves the accuracy of abdominal injury assessment through millimeter-level accurate assessment, and balances real-time decision support and high-quality data sedimentation.

[0173] It should be understood that, although each step in the flowchart involved in each embodiment as described above is displayed in sequence according to the arrow, these steps are not necessarily executed in sequence according to the arrow. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other sequences. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution sequence of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or steps or stages in other steps.

[0174] Based on the same inventive concept, the embodiments of the present application also provide an artificial intelligence-based abdominal injury assessment system for implementing the above-mentioned artificial intelligence-based abdominal injury assessment method. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme described in the above method, so the specific limitations in one or more artificial intelligence-based abdominal injury assessment system embodiments provided below can refer to the limitations of the artificial intelligence-based abdominal injury assessment method described above, which will not be repeated here.

[0175] In one exemplary embodiment, as shown in Figure 2 An artificial intelligence-based abdominal injury assessment system 800 is provided, comprising:

[0176] A source module 801 is configured to obtain multi-modal data of a patient's abdomen; wherein the multi-modal data comprises image data and clinical data;

[0177] The image module 802 is configured to extract image data features in the multi-modal data based on a Vit model to obtain a structured image feature set.

[0178] The physiological module 803 is configured to extract clinical data features in the multi-modal data based on a Wesseling algorithm to obtain a dynamic physiological index set.

[0179] The fusion module 804 is configured to fuse the structured image feature set and the dynamic physiological index set based on a cross-modal Transformer model to obtain a multi-modal fusion decision matrix.

[0180] The scoring module 805 is configured to score the abdominal regions divided by the nine-part method based on the multi-modal fusion decision matrix to obtain an organ-specific trauma score table.

[0181] Further, the scoring module 805 is further configured to:

[0182] divide the multi-modal fusion decision matrix into nine abdominal regions by the nine-part method to obtain nine-part boundary coordinates;

[0183] calculate the volume proportion of each organ in the partition based on the nine-part boundary coordinates by the following formula to obtain an organ region weight distribution table:

[0184]

[0185] wherein, w i,j is the volume proportion of the i-th organ in the j-th partition, V i,j is the volume of the organ i in the partition j, is the total volume of the organ i in all partitions;

[0186] based on the organ region weight distribution table and a preset model matching rule, assign a model to each region to obtain a matching relationship list;

[0187] based on the matching relationship list, load the model and score each abdominal region by the model to obtain the organ-specific trauma score table.

[0188] Further, the scoring module 805 is further configured to:

[0189] based on the matching relationship list, extract features in the liver and spleen region by a 3D DenseNet model to obtain a liver and spleen injury feature vector;

[0190] extract features in the intestinal region by a hybrid model combining a U-Net3D and a random forest, and splice the features with the liver and spleen injury feature vector to obtain a structured feature set;

[0191] Based on the cross-modal Transformer model, the structured feature set and the real-time physiological data are processed by attention fusion to obtain a fusion feature vector;

[0192] Based on the matching relationship list, the fusion feature vector is assigned a model parallel score to obtain a baseline ISS score table; wherein the liver and spleen region uses a gradient boosting decision tree model for scoring, and the intestinal tract region uses a combination model of LSTM model and logistic regression for scoring;

[0193] Based on the continuous physiological data, the scores in the baseline ISS score table are adjusted by the following formula to obtain an organ-specific trauma score table:

[0194]

[0195] Wherein, S final is the adjusted score, S n is the baseline ISS score, and a is the blood loss sensitivity coefficient, is the hemoglobin change rate.

[0196] Further, the system further comprises a verification module for:

[0197] Based on the preset anatomical logic rule, the organ-specific trauma score table is verified to obtain an anatomical verification report;

[0198] Based on the total blood loss rule, the organ-specific trauma score table is verified by the following formula to obtain a physiological verification report:

[0199]

[0200] Wherein, a is the theoretical blood loss amount, and the actual hemoglobin drop calculated blood loss amount is compared for verification, and ISS i is the regional score, and each ISS classification corresponds to 150ml of reference blood loss amount;

[0201] The anatomical verification report and the physiological verification report are combined to obtain a regional conflict report;

[0202] Based on the regional conflict report, regions with a confidence level greater than 85% and no conflict mark are extracted to obtain a list of trusted regions; the list of trusted regions is used to generate a digital signature to ensure audit tracking;

[0203] Based on the regional conflict report, a high-resolution model is called to reevaluate and verify regions with a confidence level below the threshold in the regional conflict report to obtain a review report;

[0204] Based on the review report and the list of trusted regions, a comprehensive trauma report is generated.

[0205] Further, the fusion module 804 is further used for:

[0206] The structured image feature set and the dynamic physiological indicator set are matched in time and space by dynamic time warping to obtain a spatio-temporal aligned multi-modal feature matrix;

[0207] The spatio-temporal aligned multi-modal feature matrix is used to generate attention weights by a cross-modal Transformer model to obtain a cross-modal attention weight matrix;

[0208] The cross-modal attention weight matrix is used to adjust dynamic weights based on a gradient reversal layer to obtain balanced multi-modal fusion features;

[0209] The balanced multi-modal fusion features are used to calculate organ damage probabilities by a multi-layer perceptron to obtain an organ damage probability matrix; each multi-layer perceptron is independently trained for each anatomical organ;

[0210] The organ damage probability matrix and the cross-modal attention weight matrix are fused based on an integrated gradient algorithm to obtain a multi-modal fusion decision matrix.

[0211] Further, the physiological module 803 is further configured to:

[0212] The clinical data and the laboratory discrete data are aligned in time by a PTP protocol to obtain a time-aligned signal stream;

[0213] The time-aligned signal stream is denoised based on an adaptive wavelet threshold method to obtain denoised physiological waveforms;

[0214] The denoised physiological waveforms are used to identify features by a Wesseling algorithm to obtain a beat-by-beat blood pressure parameter sequence; the beat-by-beat blood pressure parameter sequence is used to represent real-time blood pressure dynamic changes;

[0215] Based on the beat-by-beat blood pressure parameter sequence, a blood pressure-electrocardiogram correlation coefficient is calculated based on a cross-correlation coefficient;

[0216] Based on the blood pressure-electrocardiogram correlation coefficient and the laboratory discrete data, a compensation parameter is calculated to obtain a dynamic compensation parameter set;

[0217] Based on the dynamic compensation parameter set, an XGBoost regression model is used to obtain a blood loss rate prediction value;

[0218] The beat-by-beat blood pressure parameter sequence, the blood pressure-electrocardiogram correlation coefficient, the dynamic compensation parameter set, and the blood loss rate prediction value are integrated to obtain a flow physiological indicator set.

[0219] Further, the image module 802 is further configured to:

[0220] Based on an ITK geometric transformation model, an image data coordinate system is unified to obtain standardized three-dimensional image volume data;

[0221] segment the standardized three-dimensional image volume data through a pre-trained adaptive model to obtain an organ mask;

[0222] wherein the 3D U-Net model is used for segmenting the liver and spleen organs, and the nnU-Net model is used for segmenting the intestines, blood vessels and kidneys;

[0223] extract features from the organ mask and the standardized three-dimensional image volume data through an ITK-SNAP feature calculation engine to obtain an organ-level traditional feature table;

[0224] extract global semantics from the organ mask and the standardized three-dimensional image volume data through a Vit model to obtain an organ-level Vit feature table;

[0225] based on principal component analysis, splice the organ-level traditional feature table and the organ-level Vit feature table to obtain a fusion feature vector;

[0226] perform anomaly detection on the fusion feature vector to obtain an anomaly detection report;

[0227] integrate the fusion feature vector and the anomaly detection report to obtain a structured image feature set.

[0228] In one embodiment, a computer device is provided, comprising a memory and a processor, the memory storing a computer program, and the processor implementing the steps of the artificial intelligence-based abdominal injury assessment method as described above when executing the computer program.

[0229] In one embodiment, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps of the above method embodiments.

[0230] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts are described in the method embodiment. The device embodiments described above are only schematic, and the components described as separate components can or can not be physically separate, and the components displayed as a unit can or can not be a physical unit, i.e. they can be located in one place, or distributed on multiple network units. According to actual needs, some or all of the modules can be selected to achieve the purpose of the present disclosure. Those skilled in the art can understand and implement it without creative labor.

[0231] The above-described embodiments only express several implementation manners of the application, the description is more specific and detailed, but it cannot be understood as the limitation of the patent scope of the application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the application, several modifications and improvements can be made, which are within the protection scope of the application.

Claims

1. An artificial intelligence-based method for assessing abdominal injuries, characterized in that, The method includes: Acquire multimodal data of the patient's abdomen; wherein, the multimodal data includes imaging data and clinical data; Based on the Vit model, image data features are extracted from the multimodal data to obtain a structured image feature set; Based on the Wesseling algorithm, clinical data features are extracted from the multimodal data to obtain a set of flow cytometry physiological indicators; Based on the cross-modal Transformer model, the structured image feature set and the dynamic physiological index set are fused to obtain a multimodal fusion decision matrix; Based on the multimodal fusion decision matrix, the abdominal regions distinguished by the abdominal nine-point method are scored in parallel to obtain an organ-specific trauma scoring table.

2. The method for assessing abdominal injury based on artificial intelligence according to claim 1, characterized in that, Based on the multimodal fusion decision matrix, parallel scoring is performed on abdominal regions differentiated by the abdominal nine-point method to obtain an organ-specific trauma scoring table, including: The multimodal fusion decision matrix is ​​divided into nine abdominal regions using the abdominal nine-division method to obtain the boundary coordinates of the nine regions; Based on the boundary coordinates of the nine zones, the volume proportion of each organ within a zone is calculated using the following formula to obtain the organ region weight allocation table: Among them, w i,j V represents the volume percentage of the i-th organ in the j-th partition. i,j Let i be the volume of organ i in partition j. Let i be the total volume of organ i across all partitions; Based on the organ region weight allocation table and the preset model matching rules, a model is assigned to each region to obtain a list of matching relationships; Based on the matching relationship list, a model is loaded, and each abdominal region is scored using the model to obtain the organ-specific trauma scoring table.

3. The artificial intelligence-based abdominal injury assessment method according to claim 2, characterized in that, The process of loading a model based on the matching relationship list and scoring each abdominal region using the model to obtain the organ-specific trauma scoring table includes: Based on the matching relationship list, features are extracted from the liver and spleen regions using a 3D DenseNet model to obtain liver and spleen injury feature vectors; Features were extracted from the intestinal region using a hybrid model combining U-Net3D and random forest, and then concatenated with the liver and spleen injury feature vectors to obtain a structured feature set. Based on the cross-modal Transformer model, attention fusion processing is performed on the structured feature set and real-time physiological data to obtain a fused feature vector; Based on the matching relationship list, the fusion feature vector allocation model is scored in parallel to obtain the baseline ISS score table; wherein, the liver and spleen area is scored using a gradient boosting decision tree model, and the intestinal area is scored using a combination model of LSTM model and logistic regression. Based on continuous physiological data, the organ-specific trauma scoring scale is obtained by adjusting the scores in the baseline ISS scoring scale using the following formula: Among them, S final For the adjusted rating, S n The baseline ISS score is given, and α is the blood loss sensitivity coefficient. This represents the rate of change in hemoglobin levels.

4. The artificial intelligence-based abdominal injury assessment method according to claim 3, characterized in that, After obtaining the organ-specific trauma score by adjusting the baseline ISS score in real time based on continuous physiological data, the process further includes: Based on preset anatomical logic rules, the organ-specific trauma scoring table is verified to obtain an anatomical verification report; Based on the total blood loss rule, the organ-specific trauma scoring scale is validated using the following formula to obtain a physiological validation report: Where 'a' represents the theoretical blood loss, which is verified by comparing it with the blood loss estimated from the actual decrease in hemoglobin. (ISS) i For regional scoring, each ISS grade corresponds to a baseline blood loss of 150ml. By combining the anatomical verification report and the physiological verification report, a regional conflict report is obtained; Based on the regional conflict report, regions with a confidence level greater than 85% and no conflict markers are extracted to obtain a list of trusted regions; the list of trusted regions is used to generate digital signatures to ensure audit trails. Based on the regional conflict report, a high-resolution model is invoked to re-evaluate and verify the regions in the regional conflict report whose confidence level is lower than the threshold, resulting in a review report. A comprehensive trauma report is generated based on the review report and the list of trusted areas.

5. The method for assessing abdominal injury based on artificial intelligence according to claim 1, characterized in that, The multimodal fusion decision matrix, based on the cross-modal Transformer model and fusing the structured image feature set and the dynamic physiological index set, includes: By performing dynamic time warping, the structured image feature set and the dynamic physiological indicator set are matched in time and space to obtain a spatiotemporally aligned multimodal feature matrix. The spatiotemporally aligned multimodal feature matrix is ​​used to generate attention weights through a cross-modal Transformer model to obtain a cross-modal attention weight matrix; For the cross-modal attention weight matrix, the dynamic weights are adjusted based on the gradient reversal layer to obtain the balanced multimodal fusion features; The organ damage probability matrix is ​​obtained by calculating the organ damage probability corresponding to the balanced multimodal fusion features using a multilayer perceptron; each multilayer perceptron is trained independently for each anatomical organ. Based on the ensemble gradient algorithm, the organ damage probability matrix and the cross-modal attention weight matrix are fused to obtain the multimodal fusion decision matrix.

6. The method for assessing abdominal injuries based on artificial intelligence according to claim 1, characterized in that, The clinical data features extracted from the multimodal data using the Wesseling algorithm are used to obtain a set of flow cytometry physiological indicators, including: The clinical data and laboratory discrete data are time-aligned using the PTP protocol to obtain a time-aligned signal stream; Based on the adaptive wavelet thresholding method, the time-aligned signal stream is denoised to obtain the denoised physiological waveform. The Wesseling algorithm is used to identify features of the denoised physiological waveform to obtain a pulse-wave blood pressure parameter sequence; the pulse-wave blood pressure parameter sequence is used to characterize real-time dynamic changes in blood pressure. Based on the aforementioned pulse-wave blood pressure parameter sequence, and based on the cross-correlation coefficient, the correlation between blood pressure parameters and electrocardiogram parameters is calculated to obtain the blood pressure-electrocardiogram correlation coefficient. Based on the blood pressure-electrocardiogram correlation coefficient and the laboratory discrete data, compensation parameters are calculated to obtain a dynamic compensation parameter set. Based on the dynamic compensation parameter set, the predicted blood loss rate is obtained through the XGBoost regression model. The flow cytometry physiological index set is obtained by integrating the pulse-wave blood pressure parameter sequence, the blood pressure-ECG correlation coefficient, the dynamic compensation parameter set, and the predicted blood loss rate.

7. The method for assessing abdominal injury based on artificial intelligence according to any one of claims 1 to 6, characterized in that, The Vit model is used to extract image data features from the multimodal data to obtain a structured image feature set, including: Based on the ITK geometric transformation model, the coordinate system of the image data is unified to obtain standardized three-dimensional image volume data; The standardized 3D image volume data is segmented using a pre-trained adaptive model to obtain organ masks. Specifically, the liver and spleen were segmented using a 3D U-Net model, while the intestines, blood vessels, and kidneys were segmented using an nnU-Net model. The organ mask and the standardized three-dimensional image volume data are used to extract features through the ITK-SNAP feature calculation engine to obtain an organ-level traditional feature table. The organ mask and the standardized three-dimensional image volume data are used to perform global semantic extraction through the Vit model to obtain an organ-level Vit feature table; Based on principal component analysis, the features of the organ-level traditional feature table and the organ-level Vit feature table are concatenated to obtain a fused feature vector; Anomaly detection is performed on the fused feature vector to obtain an anomaly detection report; The structured image feature set is obtained by integrating the fused feature vector and the anomaly detection report.

8. An artificial intelligence-based abdominal injury assessment system, characterized in that, The system includes: The source module is used to acquire multimodal data of the patient's abdomen; wherein, the multimodal data includes imaging data and clinical data; The image module is used to extract image data features from the multimodal data based on the Vit model to obtain a structured image feature set. The physiological module is used to extract clinical data features from the multimodal data based on the Wesseling algorithm to obtain a set of flow cytometry physiological indicators. The fusion module is used to fuse the structured image feature set and the dynamic physiological index set based on the cross-modal Transformer model to obtain a multimodal fusion decision matrix; The scoring module is used to score the abdominal regions distinguished by the abdominal nine-point method in parallel based on the multimodal fusion decision matrix, so as to obtain an organ-specific trauma scoring table.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images