An auxiliary diagnosis method for children acute abdomen based on multi-modal data fusion

Abstract

The application discloses a kind of children acute abdomen auxiliary diagnosis method based on multi-modal data fusion, it is related to artificial intelligence medical field, the method includes: the double-probe ultrasonic video sequence of sick child, clinical structured information and laboratory inflammation-related index are collected;Bayesian group LASSO priori and diagnostic priori graph constraint are used to carry out predictor screening, obtain diagnostic correlation factor and its regression coefficient posterior distribution;Diagnostic knowledge graph is constructed and instruction text is injected into training triple;Visual features are extracted using graphic-text feature extraction network and evidence fusion is carried out with clinical statistical characteristics, to obtain fusion characteristics;LoRA low-rank adaptation layer parameters in graphic-text feature extraction network and evidence head parameters are jointly optimized, to obtain auxiliary diagnosis model;According to auxiliary diagnosis model, two-stage cascade diagnosis is carried out, backtracking knowledge graph diagnostic reasoning path, generates structured auxiliary diagnosis report, improves the accuracy and explainability of children acute abdomen auxiliary diagnosis, provides auxiliary reference for clinical decision-making.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence medical technology, and in particular to an auxiliary diagnostic method for acute abdominal pain in children based on multimodal data fusion. Background Technology

[0002] Children presenting with acute abdominal pain as their chief complaint are very common in pediatric outpatient clinics. Children's physiological and anatomical structures differ significantly from adults, resulting in a much wider spectrum of acute abdominal conditions. These conditions are complex in etiology and progress rapidly, with common causes including acute appendicitis, intussusception, gallstones, and ovarian cyst torsion. Treatment approaches vary depending on the disease; some require emergency surgical intervention, while others can be managed conservatively. Therefore, the ability to quickly and accurately identify critical and severe cases at the initial consultation is crucial for clinical decision-making.

[0003] Ultrasound examination, due to its advantages such as being non-invasive, radiation-free, and bedside operation, has become the preferred imaging method for pediatric acute abdominal conditions. Its high-frequency linear array probe offers high resolution, clearly displaying superficial abdominal wall structures and small lesions; its low-frequency convex array probe has strong penetrating power, suitable for exploring deep abdominal organs and large-scale lesions. Existing studies have shown that dual-probe combined scanning can improve diagnostic accuracy and reduce misdiagnosis rates. However, current clinical practice still faces the following technical challenges: First, the fusion analysis of dual-probe ultrasound information highly relies on the individual experience of the sonographer, lacking automated feature complementarity and joint reasoning mechanisms; second, ultrasound image features, clinical symptoms, and laboratory indicators are fragmented, and existing medical image analysis methods based on large-scale visual language models mostly focus on single image modalities, failing to effectively fuse dual-probe ultrasound images with clinical symptoms and laboratory indicators in a multimodal manner, and lacking uncertainty quantification and evidence conflict measurement mechanisms during the fusion process; third, diagnostic conclusions lack traceable reasoning paths, resulting in insufficient interpretability. Summary of the Invention

[0004] In view of the above-mentioned prior art, the present invention provides an auxiliary diagnostic method for acute abdominal pain in children based on multimodal data fusion, which mainly solves the technical problems existing in the background art.

[0005] To achieve the above objectives, the technical solution of this invention is implemented as follows: This invention provides an auxiliary diagnostic method for acute abdominal pain in children based on multimodal data fusion, the method comprising the following steps: S1: Collect dual-probe ultrasound video sequences, clinical structured information, disease type diagnostic labels, and laboratory inflammation-related indicators of the child; S2: Based on the clinical structured information, the laboratory inflammation-related indicators, and the disease type diagnostic label, predictive factors are screened using Bayesian group LASSO prior and diagnostic prior graph constraints to obtain the diagnostic association factors and their posterior distribution of regression coefficients; S3: Construct a diagnostic knowledge graph and training triples, generate instruction text based on the diagnostic reasoning path in the diagnostic knowledge graph, and inject the instruction text into the question text of the training triples to obtain the enhanced training triples. S4: Based on the ultrasound video sequence, visual features are extracted using a graphic feature extraction network; clinical statistical features are obtained based on the posterior distribution of the regression coefficients of the diagnostic correlation factors and the clinical structured information; the visual features and the clinical statistical features are fused using an evidence head to obtain fused features. S5: Based on the fusion features and the enhanced training triples, jointly optimize the parameters of the LoRA low-rank adaptation layer and the parameters of the evidence head configured in the image and text feature extraction network to obtain the optimized image and text feature extraction network and use it as an auxiliary diagnostic model. S6: Perform a two-stage cascaded diagnosis based on the auxiliary diagnostic model, output the diagnostic results, and based on the diagnostic results, trace back the diagnostic reasoning path in the diagnostic knowledge graph to generate a structured auxiliary diagnostic report.

[0006] As a preferred embodiment of the present invention, step S2 specifically includes: S2-1: Based on the clinical structured information and the laboratory inflammation-related indicators, the clinical feature vector is obtained through encoding and standardization. S2-2: Construct a multinomial logistic regression model based on the clinical feature vector and the disease type diagnosis label. Apply a Bayesian group LASSO prior distribution to each regression coefficient in the multinomial logistic regression model and perform posterior sampling to obtain the posterior distribution of each regression coefficient. S2-3: Calculate the posterior inclusion probability of each predictor based on the posterior distribution of each regression coefficient, calculate the coefficient of variation of each predictor based on the posterior inclusion probability, and remove unstable predictors based on the coefficient of variation to obtain a set of stable predictors. S2-4: Match the stable predictor set with the association paths of each disease node in the diagnostic prior graph, retain the predictors with association paths as the diagnostic association factors, and extract the corresponding subsets from the posterior distribution of each regression coefficient according to the variable index of the diagnostic association factors to obtain the posterior distribution of the regression coefficients of the diagnostic association factors.

[0007] As a preferred embodiment of the present invention, step S2-2 specifically includes: S2-2-1: The specific formula for constructing a multinomial logistic regression model is as follows:

[0008] in, In clinical feature vectors The following child belongs to the first category The probability of such diseases; For the first Regression coefficients corresponding to disease categories This represents the total number of disease types. For the first Regression coefficients corresponding to disease categories; S2-2-2: Applying the Bayesian LASSO prior distribution to each regression coefficient in the multinomial logistic regression model and performing posterior sampling, the specific formula for obtaining the posterior distribution of the regression coefficients is as follows:

[0009] in, To obtain the posterior distribution of the regression coefficients through posterior sampling, For the observation dataset, Let be the likelihood function, representing the likelihood given the regression coefficients. All observed below Disease labels of training samples The joint probability, For the first The likelihood value of each sample. For the Bayesian group LASSO prior distribution, The sign indicates proportionality.

[0010] As a preferred embodiment of the present invention, step S3 specifically includes: S3-1: Construct the diagnostic knowledge graph using ultrasound signs, disease types, and laboratory inflammation-related indicators as nodes, and differential diagnostic relationships and conflict relationships as edges; S3-2: Construct the training triplet consisting of ultrasound images, question text, and answer text; S3-3: Traverse the diagnostic knowledge graph, starting from each ultrasound sign node, extract the complete diagnostic reasoning path to the disease type node along the differential diagnosis relationship edge, and convert the node sequence and edge relationship in the diagnostic reasoning path into instruction text; S3-4: The instruction text is appended as a prefix to the question text of the training triple, and the appended question text is reorganized with the corresponding ultrasound image and answer text to obtain the enhanced training triple.

[0011] As a preferred embodiment of the present invention, step S4 specifically includes: S4-1: Extract high-frequency keyframes and low-frequency keyframes from the ultrasound video sequence using the visual encoder of the image and text feature extraction network to obtain high-frequency aggregation vectors and low-frequency aggregation vectors. S4-2: Based on the clinical structured information and laboratory inflammation-related indicators of the current sample, obtain the clinical feature vector, and concatenate the clinical feature vector with the mean vector of the posterior distribution of the regression coefficients of the diagnostic association factors to obtain the clinical statistical feature vector. S4-3: Input the high-frequency aggregation vector, the low-frequency aggregation vector, and the clinical statistical feature vector into the trainable evidence head for nonlinear mapping to obtain the high-frequency evidence vector, the low-frequency evidence vector, and the clinical evidence vector, and construct their respective Dirichlet distributions. S4-4: Based on their respective Dirichlet distributions, the high-frequency evidence vector, the low-frequency evidence vector, and the clinical evidence vector are converted into high-frequency opinions, low-frequency opinions, and clinical opinions, respectively. S4-5: The high-frequency opinions, low-frequency opinions, and clinical opinions are iteratively aggregated using the DS evidence combination rule to obtain fusion features.

[0012] As a preferred embodiment of the present invention, step S4-4 specifically includes: S4-4-1: Using the Dirichlet distribution as the conjugate prior, the high-frequency evidence vector, the low-frequency evidence vector, and the clinical evidence vector are respectively used as the base values ​​of the concentration parameters of the Dirichlet distribution, and are summed with the preset prior concentration parameters to obtain the Dirichlet concentration parameters corresponding to each evidence vector. S4-4-2: Divide the component value of each dimension in each evidence vector by the total Dirichlet strength of the corresponding evidence vector to obtain the class belief quality of each evidence vector in each category. S4-4-3: The uncertainty mass of each evidence vector is obtained by dividing the product of the prior concentration parameter of the Dirichlet distribution and the total number of categories by the total Dirichlet intensity. S4-4-4: Combine the class belief quality and uncertainty quality of each evidence vector in each category to obtain the high-frequency opinion, the low-frequency opinion and the clinical opinion.

[0013] As a preferred embodiment of the present invention, step S4-5 specifically includes: S4-5-1: Combine the high-frequency opinion with the low-frequency opinion as DS evidence to obtain the intermediate image opinion; combine the intermediate image opinion with the clinical opinion as DS evidence to obtain the fusion opinion; S4-5-2: Set corresponding reliability coefficients based on the uncertainty quality of intermediate imaging opinions and clinical opinions respectively; S4-5-3: Calculate the conflict coefficient based on the joint Dirichlet strength of the fusion opinions; S4-5-4: When the conflict coefficient is greater than the preset threshold, the quality of the category belief is weighted and discounted according to the corresponding reliability coefficient. The DS evidence is recombined according to the weighted and discounted quality of the category belief to obtain the final fusion opinion. The fusion feature is obtained according to the final fusion opinion.

[0014] As a preferred embodiment of the present invention, step S5 specifically includes: S5-1: Input the fused features into the language decoder of the image-text feature extraction network, and use the word sequence of the answer text in the enhanced training triples as the supervision target to calculate the model loss; S5-2: Calculate the mean squared error loss based on the difference between the unique heat code of the disease type diagnosis label and the category belief quality of each evidence vector; S5-3: Calculate the variance loss based on the concentration parameters of the Dirichlet distribution corresponding to each evidence vector; S5-4: Calculate the KL divergence penalty loss based on the concentration parameters of the Dirichlet distribution corresponding to the non-real disease type category component in each evidence vector; S5-5: The mean square error loss, the variance loss and the KL divergence penalty loss are weighted and summed to obtain the evidence calibration loss; S5-6: Obtain the total loss based on the model loss and the evidence calibration loss; S5-7: Based on the total loss, the parameters of the pre-configured LoRA low-rank adaptation layer and the parameters of each evidence head in the image feature extraction network are jointly optimized through backpropagation to obtain the auxiliary diagnostic model.

[0015] As a preferred embodiment of the present invention, in step S5-5, the mean square error loss, the variance loss, and the KL divergence penalty loss are weighted and summed to obtain the specific formula for the evidence calibration loss as follows:

[0016] in, To calibrate the loss for evidence, For the unique heat code of the disease type diagnostic label, For disease category indexing, For the first The total Dirichlet strength of the path evidence vector, For the first The Dirichlet distribution corresponding to the path evidence vector is in the th... Concentration parameters for disease categories To follow the training rounds Increasing regularization weight coefficients, KL divergence is used to measure the difference between two probability distributions. Indicates the calibration concentration parameter The Dirichlet distribution for the concentration parameter; For the first The probability distribution vector of the road evidence vector across each category. This indicates a uniform Dirichlet distribution.

[0017] As a preferred embodiment of the present invention, step S6 specifically includes: S6-1: Input the fusion features into the language decoder of the auxiliary diagnostic model to obtain specific ultrasound signs; S6-2: In the disease type determination stage, retrieve differential diagnosis paths that match the specific ultrasound signs from the diagnostic knowledge graph, use the differential diagnosis paths as the decoding context, and adjust the probability of candidate disease types in combination with the child's gender to obtain the disease type; S6-3: Under the constraint of disease type, obtain the severity level corresponding to the disease type based on the laboratory inflammation-related indicators and the specific ultrasound signs.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention uses Bayesian group LASSO priors and diagnostic prior graph constraints to screen diagnostic association factors, ensuring that the injected features have statistical robustness and cognitive uncertainty; it generates diagnostic reasoning paths through a diagnostic knowledge graph; it quantifies the uncertainty of multi-source evidence and measures cross-modal consistency conflicts through evidence deep learning and DS evidence combination rule iterative aggregation, improving the reliability of multimodal fusion; it performs two-stage cascaded diagnosis based on the auxiliary diagnostic model, backtracks the diagnostic reasoning path of the knowledge graph, and generates a structured auxiliary diagnostic report to improve the accuracy and interpretability of auxiliary diagnosis of acute abdomen in children, providing auxiliary reference for clinical decision-making. Attached Figure Description

[0019] Figure 1 This is a schematic diagram illustrating the steps of an auxiliary diagnostic method for acute abdominal pain in children based on multimodal data fusion. Detailed Implementation

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. In the following description, the expression "some embodiments" refers to a subset of all possible embodiments; however, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0021] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0022] It should be understood that the present invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Furthermore, the terminology used herein is intended only to describe particular embodiments and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “compose” and / or “comprising,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.

[0023] It should also be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "inner," "outer," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0024] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.

[0025] Please refer to the attached document. Figure 1 This application provides a method for auxiliary diagnosis of acute abdominal pain in children based on multimodal data fusion, the method comprising the following steps: S1: Collect dual-probe ultrasound video sequences, clinical structured information, disease type diagnostic labels, and laboratory inflammation-related indicators for the child.

[0026] In this embodiment, a high-frequency linear array probe and a low-frequency convex array probe were used to scan the child's abdomen, acquiring two ultrasound video sequences. The high-frequency linear array probe has high resolution and can clearly display the superficial structures of the abdominal wall and small lesions; the low-frequency convex array probe has strong penetration and is suitable for exploring deep abdominal organs and large-scale lesions. The acquisition time of each ultrasound video sequence is 3 to 10 seconds, and the number of effective frames in each ultrasound video sequence is not less than 20 frames.

[0027] Furthermore, keyframe extraction is performed on each ultrasound video sequence, specifically including: calculating local contrast, edge sharpness, and signal-to-noise ratio frame by frame; weightedly fusing local contrast, edge sharpness, and signal-to-noise ratio to obtain a three-dimensional quality score; removing frames with a three-dimensional quality score below a preset threshold; calculating the structural similarity of adjacent frames in chronological order for the remaining frames; determining redundant frames when the structural similarity exceeds a preset threshold, and retaining only the frame with the highest three-dimensional quality score in the redundant segment; clustering the deredundant frames based on grayscale histogram statistical features, and selecting the top-ranked frames from each cluster according to their three-dimensional quality scores. frame, A value of 5 can be chosen to improve the coverage of different image appearance types. This yields high-frequency and low-frequency keyframes.

[0028] In this embodiment, the clinical structured information includes demographic information and symptom and sign information, specifically including age, gender, whether there is abdominal pain, whether there is fever, whether there is rectal bleeding, whether there is vomiting, and whether there is an abdominal mass.

[0029] In this embodiment, the laboratory inflammation-related indicators include white blood cell count (WBC), C-reactive protein (CRP), and interleukin-6 (IL-6). Simultaneously, disease type diagnostic labels are also collected, including appendicitis, intussusception, gallstones, ovarian cyst torsion, and other types of acute abdominal conditions.

[0030] S2: Based on the clinical structured information, the laboratory inflammation-related indicators, and the disease type diagnostic label, predictive factors are screened using Bayesian group LASSO prior and diagnostic prior graph constraints to obtain the diagnostic association factors and their posterior distribution of regression coefficients.

[0031] In this embodiment, step S2 specifically includes the following steps: S2-1: Based on the clinical structured information and the laboratory inflammation-related indicators, encoding and standardization are performed to obtain a clinical feature vector. The specific process is as follows: The presence or absence of fever, rectal bleeding, vomiting, and abdominal mass in the clinical structured information are used as categorical variables and converted into numerical variables using one-heat coding. For example, "yes" is encoded as 1 and "no" is encoded as 0. For continuous variables (such as age) in the clinical structured information and each indicator variable in the laboratory inflammation-related indicators, Z-score standardization is used for standardization. All standardized variables are concatenated to obtain the clinical feature vector.

[0032] S2-2: Construct a multinomial logistic regression model based on the clinical feature vector and the disease type diagnosis label. Apply a Bayesian LASSO prior distribution to each regression coefficient in the multinomial logistic regression model and perform posterior sampling to obtain the posterior distribution of each regression coefficient; the specific process is as follows: Diagnosis label based on disease type To provide a supervisory signal, a multinomial logistic regression model is constructed, with the following formula:

[0033] in, In clinical feature vectors The following child belongs to the first category The probability of such diseases; For the first Regression coefficients corresponding to disease categories Total number of disease types For the first Regression coefficients corresponding to disease categories.

[0034] Unlike traditional logistic regression, which provides a set of regression coefficients... As a deterministic point estimate, this invention applies a Bayesian group LASSO prior distribution to each regression coefficient. The posterior distribution of the regression coefficients is obtained through posterior sampling. :

[0035] in, For the observation dataset, Let be the likelihood function, representing the likelihood given the regression coefficients. All observed below Disease labels of training samples The joint probability, For the first The likelihood value of each sample, The sign indicates proportionality. The posterior distribution... It carries cognitive uncertainty with regression coefficients, rather than just providing point estimates. Specifically, the Bayesian group's LASSO prior distribution... The specific mathematical expression is as follows:

[0036] in, Grouping indexes for variables, The total number of groups, For the first The group's regularization weights, For belonging to the first The subvectors of the regression coefficients of the group for Norm, This is the regularization hyperparameter.

[0037] S2-3: Calculate the posterior inclusion probability of each predictor based on the posterior distribution of the regression coefficients, calculate the coefficient of variation of each predictor based on the posterior inclusion probability, and remove unstable predictors based on the coefficient of variation to obtain a set of stable predictors. The specific process is as follows: First, to avoid the screening results from over-reliance on a single regularization hyperparameter. The present invention performs sensitivity analysis on the prior parameter values, specifically including: performing K-fold cross-validation on the training set, where K is 5 or 10, and in this embodiment, it is 5. This is done on a preset hyperparameter grid. Sampling at logarithmic intervals One value, Take 10-20, The optimal value is selected as the one that maximizes the expected log-likelihood of the cross-validation. Other candidate values ​​are then selected centered around this optimal value to form a candidate hyperparameter value set. .

[0038] Calculate the posterior inclusion probability of each predictor under the set of candidate hyperparameter values. The formula for calculating the total number of cases is as follows:

[0039] in, This indicates that given the observed data D and the hyperparameters... Next, the Regression coefficients of predictors on disease type c The posterior mean, For hyperparameter index, The preset effect size threshold is taken in this embodiment. ; Let be an indicator function, meaning that the value is 1 if the condition within the parentheses is true, and 0 otherwise; for the Each predictor factor is used to determine the posterior mean of its regression coefficients for various diseases. Is the absolute value greater than the preset effect size threshold? The mean of the judgment results across all categories is taken as the posterior inclusion probability of the predictor.

[0040] Furthermore, the coefficient of variation is calculated based on the posterior inclusion probability of each predictor under different hyperparameters, and the calculation formula is as follows:

[0041] in, For the first The standard deviation of the posterior inclusion probability of each predictor under different hyperparameters Let be the mean of the posterior inclusion probability, when the coefficient of variation If the result of the selection of the predictor is unstable, it is rejected, thus obtaining a stable predictor set.

[0042] S2-4: Match the stable predictor set with the association paths of each disease node in the diagnostic prior graph, retain the predictors with associated paths as the diagnostic association factors, and extract corresponding subsets from the posterior distributions of the regression coefficients based on the variable indexes of the diagnostic association factors to obtain the posterior distributions of the regression coefficients of the diagnostic association factors. Specifically, this includes: In this embodiment, a diagnostic prior graph is constructed based on the generative relationship between disease state and clinical manifestations to constrain the association between symptoms, signs, laboratory indicators, and disease types. The diagnostic prior graph uses disease type and clinical manifestations as nodes and causal relationships as edges.

[0043] Furthermore, based on the variable index of the diagnostic association factor, the regression coefficient component corresponding to the index position is extracted from the posterior distribution of each regression coefficient. Specifically, during the posterior sampling process, for each regression coefficient posterior distribution obtained in each sampling iteration, only the component value corresponding to the variable index set is retained, and other components that were not selected as diagnostic association factors are discarded. The retained components are then reorganized into the posterior distribution of the regression coefficients corresponding to the diagnostic association factor.

[0044] S3: Construct a diagnostic knowledge graph and training triples, generate instruction text based on the diagnostic reasoning path in the diagnostic knowledge graph, and inject the instruction text into the question text of the training triples to obtain the enhanced training triples.

[0045] In this embodiment, step S3 specifically includes: S3-1: Construct the diagnostic knowledge graph using ultrasound signs, disease types, and laboratory inflammation-related indicators as nodes, and differential diagnostic relationships and conflict relationships as edges.

[0046] Specifically, the nodes of the diagnostic knowledge graph include ultrasound sign nodes (such as blind-ended tubular structures, concentric circle sign, hyperechoic mass with acoustic shadowing, vortex sign, etc.), disease type nodes (such as appendicitis, intussusception, gallstones, ovarian cyst torsion, etc.), and laboratory inflammation-related indicator nodes (such as elevated WBC, elevated CRP, etc.). Edges include differential diagnostic relationships (such as "concentric circle sign, suggesting intussusception") and conflicting relationships (such as "hyperechoic mass with acoustic shadowing" and "blind-ended tubular structure" pointing to different disease types).

[0047] S3-2: Construct the training triplet consisting of ultrasound images, question text, and answer text.

[0048] Specifically, the training triples are represented as ,in, For ultrasound imaging, For the question text, The answer text; the question text The answer text is obtained by splicing together the clinical symptom description and the laboratory test indicators. The data is presented in a structured format, sequentially including a description of ultrasound findings, a preliminary diagnosis, and a pathological classification.

[0049] S3-3: Traverse the diagnostic knowledge graph, starting from each ultrasound sign node, extract the complete diagnostic reasoning path to the disease type node along the differential diagnosis relationship edge, and convert the node sequence and edge relationship in the diagnostic reasoning path into instruction text.

[0050] S3-4: The instruction text is appended as a prefix to the question text of the training triple, and the appended question text is reorganized with the corresponding ultrasound image and answer text to obtain the enhanced training triple.

[0051] S4: Based on the ultrasound video sequence, visual features are extracted using a graphic feature extraction network; based on the posterior distribution of the regression coefficients of the diagnostic correlation factors and the clinical structured information, clinical statistical features are obtained; the visual features and the clinical statistical features are fused using an evidence head to obtain fused features.

[0052] Specifically, the image-text feature extraction network includes a visual encoder with a multi-layer attention mechanism to encode the input image into a sequence of visual features, and a language decoder with a multi-layer attention mechanism for autoregressive text generation. The image-text feature extraction network uses Qwen3-VL as its base model and can simultaneously receive image and text input.

[0053] In this embodiment, step S4 specifically includes: S4-1: Extract high-frequency keyframes and low-frequency keyframes from the ultrasound video sequence using the visual encoder of the image and text feature extraction network to obtain high-frequency aggregation vectors and low-frequency aggregation vectors.

[0054] Specifically, will The high-frequency keyframes of a frame are input into a visual encoder to obtain high-frequency visual features. These high-frequency visual features are then global pooled and mapped, and a correlation discount factor based on inter-frame structural similarity is applied. Aggregates into high-frequency aggregated vectors The specific formula is as follows:

[0055]

[0056] in, For the first The maximum structural cosine similarity between the frame and the selected representative frames. The preset weight lower limit ranges from 0.1 to 0.4; in this embodiment, it is set to 0.2. The same process is applied to the low-frequency keyframes of the frame to obtain the low-frequency aggregate vector. .

[0057] S4-2: Based on the current sample's clinical structured information and laboratory inflammation-related indicators, obtain a clinical feature vector. Concatenate the clinical feature vector with the mean vector of the posterior distribution of the regression coefficients of the diagnostic association factors to obtain a clinical statistical feature vector. .

[0058] In this embodiment, following the same method as step S2-1, the symptom characteristic variables and laboratory inflammation-related indicators in the clinical structured information of the current sample are encoded and standardized. All the encoded and standardized variables are then concatenated in a preset order to obtain the clinical feature vector of the current sample.

[0059] Furthermore, the posterior distribution of the regression coefficients of the diagnostic association factors is obtained, and the posterior mean vector of the regression coefficients corresponding to each disease category is calculated. Specifically, for each disease category, the average of all posterior sampled values ​​of the diagnostic association factors under that disease category is calculated to obtain the posterior mean vector of the regression coefficients of the diagnostic association factors corresponding to that disease category.

[0060] The clinical statistical feature vector is obtained by concatenating the clinical feature vector of the current sample with the posterior mean vector of the regression coefficients of the diagnostic association factors corresponding to all disease categories in sequence.

[0061] S4-3: Input the high-frequency aggregated vector, the low-frequency aggregated vector, and the clinical statistical feature vector into the trainable evidence head for nonlinear mapping to obtain the high-frequency evidence vector, low-frequency evidence vector, and clinical evidence vector, respectively, and construct their respective Dirichlet distributions; wherein, the specific nonlinear mapping formula is as follows:

[0062] in, Indicates the first The evidence vector output by the road evidence header. , indicates taking the subscript symbol. It is the ReLU activation function. and The first The trainable weight matrix and bias vector of the path evidence header. The set represents the high-frequency aggregated vector, the low-frequency aggregated vector, and the clinical statistical feature vector. By substituting the high-frequency aggregated vector, the low-frequency aggregated vector, and the clinical statistical feature vector into the nonlinear mapping formula, the high-frequency evidence vector, the low-frequency evidence vector, and the clinical evidence vector can be obtained respectively.

[0063] Specifically, taking the first Taking the path evidence vector as an example, the first The probability density function of the Dirichlet distribution corresponding to the path evidence vector for:

[0064] in, For the first The probability distribution vectors of the path evidence vectors in each category satisfy the following conditions: ; For the first The Dirichlet concentration parameter vector of the path evidence vector, Indicates the first The path evidence vector is at the th Dirichlet concentration parameters corresponding to each disease category; Let be the gamma function, representing the generalized form of the factorial over the real number field. The expected value of this probability density function distribution is:

[0065] in, For the first The total Dirichlet strength of the path evidence vector, For the first The path evidence vector is at the th Quality of category beliefs in each disease category For the first The path evidence vector is at the th Component values ​​for each disease category.

[0066] S4-4: Based on their respective Dirichlet distributions, the high-frequency evidence vector, the low-frequency evidence vector, and the clinical evidence vector are transformed into high-frequency opinions, low-frequency opinions, and clinical opinions, respectively. This specifically includes the following steps: S4-4-1: Using the Dirichlet distribution as the conjugate prior, the high-frequency evidence vector, the low-frequency evidence vector, and the clinical evidence vector are respectively used as the base values ​​of the concentration parameters of the Dirichlet distribution, and then summed with the preset prior concentration parameters to obtain the concentration parameters of the Dirichlet distribution corresponding to each evidence vector. The specific formula is as follows:

[0067] in, This is the prior concentration parameter.

[0068] S4-4-2: Divide the component value of each dimension in each evidence vector by the total Dirichlet strength of the corresponding evidence vector to obtain the class belief quality of each evidence vector in each category. The specific formula is as follows:

[0069] in, For the first The total Dirichlet strength of the road evidence vector.

[0070]

[0071] in, For the first The quality of the category belief of the road evidence vector.

[0072] S4-4-3: The uncertainty mass of each evidence vector is obtained by dividing the product of the prior concentration parameter of the Dirichlet distribution and the total number of categories by the total Dirichlet intensity. The specific formula is as follows:

[0073] in, Indicates the first Uncertainty quality of the path evidence vector.

[0074] S4-4-4: Combine the category belief quality and uncertainty quality of each evidence vector in each category to obtain the high-frequency opinion, the low-frequency opinion and the clinical opinion, wherein the high-frequency opinion, the low-frequency opinion and the clinical opinion each include their respective category belief quality and uncertainty quality.

[0075] S4-5: The high-frequency opinions, low-frequency opinions, and clinical opinions are iteratively aggregated using the DS evidence combination rule to obtain fusion features, specifically including the following steps: S4-5-1: Combine the high-frequency opinion with the low-frequency opinion as DS evidence to obtain the intermediate image opinion; combine the intermediate image opinion with the clinical opinion as DS evidence to obtain the fusion opinion; In this embodiment, the basic form of the DS evidence combination rule is as follows:

[0076] in, For the first Road opinion in the first Quality of category beliefs in each disease category For the first The uncertainty quality of road opinions For the first Road opinion in the first Quality of category beliefs in each disease category For the first The uncertainty quality of road opinions For the combined Dirichlet strength of integrated opinions; and The opinions on integration were respectively in the first Category belief quality and uncertainty quality across disease categories.

[0077] S4-5-2: Set corresponding reliability coefficients based on the uncertainty quality of intermediate imaging opinions and clinical opinions respectively. The specific formula is as follows:

[0078] S4-5-3: Calculate the conflict coefficient based on the joint Dirichlet strength of the fusion opinions. The specific formula is as follows:

[0079] S4-5-4: When the conflict coefficient is greater than the preset discount threshold At that time, a preset discount threshold is set. The value ranges from 0.05 to 0.15, and in this embodiment, it is taken as 0.10. The quality of the category belief is then weighted and discounted according to the corresponding reliability coefficient:

[0080] in, This represents the quality of category beliefs after weighted discounting.

[0081] If the quality of the category belief after the discount is satisfied , This represents the uncertainty after the discount.

[0082] Based on the weighted discounted quality of categorical beliefs, the DS evidence is recombined to obtain the final fusion opinion, and the fusion uncertainty quality in the final fusion opinion is extracted; the corresponding conflict coefficient is used as the fusion conflict coefficient.

[0083] Furthermore, the fusion weights of each opinion are determined based on the uncertainty quality of the high-frequency opinion, the low-frequency opinion, and the clinical opinion, respectively. :

[0084] in, This indicates that the source of all evidence vectors is traversed and summed. This indicates a temporary traversal index number. Take in sequence .

[0085] The high-frequency aggregation vector, the low-frequency aggregation vector, and the clinical statistical feature vector are weighted and summed according to each fusion weight. The fusion uncertainty quality is used as a gate coefficient to scale the weighted summation result to obtain the fusion features. The specific formula is as follows:

[0086] in, To integrate uncertain quality.

[0087] S5: Based on the fusion features and the enhanced training triples, jointly optimize the parameters of the LoRA low-rank adaptation layer and the parameters of the evidence head configured in the image and text feature extraction network to obtain the optimized image and text feature extraction network and use it as an auxiliary diagnostic model.

[0088] In this embodiment, step S5 specifically includes: S5-1: Input the fused features into the language decoder of the image-text feature extraction network, and use the word sequence of the answer text in the enhanced training triples as the supervision target to calculate the model loss. The specific formula is as follows:

[0089] in, For model loss, This represents the expectation of the cross-entropy loss over all samples in the training set. Represents a given fusion feature At that time, the model processed the real word sequence. The predicted probability, This is the set of trainable parameters.

[0090] Due to the parameters of each evidence header and bias vector The gradient is only indirectly backpropagated through the language decoder and fused features, lacking direct supervision of its category belief quality and uncertainty quality. Therefore, this invention adds an evidence calibration loss for each evidence header, as follows: S5-2: Calculate the mean squared error loss based on the difference between the unique heat code of the disease type diagnosis label and the category belief quality of each evidence vector; S5-3: Calculate the variance loss based on the concentration parameters of the Dirichlet distribution corresponding to each evidence vector; S5-4: Calculate the KL divergence penalty loss based on the concentration parameters of the Dirichlet distribution corresponding to the non-real disease type category component in each evidence vector; S5-5: The weighted sum of the mean square error loss, the variance loss, and the KL divergence penalty loss yields the evidence calibration loss, the specific formula of which is as follows:

[0091] in, To calibrate the loss for evidence, For the unique heat code of the disease type diagnostic label, For disease category indexing, For the first The total Dirichlet strength of the path evidence vector, For the first The Dirichlet distribution corresponding to the path evidence vector is in the th... Concentration parameters for each disease category To follow the training rounds Increasing regularization weight coefficients, KL divergence is used to measure the difference between two probability distributions. Indicates the calibration concentration parameter For the Dirichlet distribution of the concentration parameter, For the first The probability distribution vector of the road evidence vector across each category. This indicates a uniform Dirichlet distribution.

[0092] S5-6: Obtain the total loss based on the model loss and the evidence calibration loss. The specific formula is as follows:

[0093] in, The weighting coefficient for evidence calibration loss is used to balance the contribution of evidence calibration loss. Its value ranges from 0.01 to 0.1, and in this embodiment, it is 0.05.

[0094] S5-7: Based on the total loss, the parameters of the pre-configured LoRA low-rank adaptation layer and the parameters of each evidence head in the image feature extraction network are jointly optimized through backpropagation to obtain the auxiliary diagnostic model.

[0095] In this embodiment, the LoRA low-rank adaptation layer parameters include low-rank matrix A and low-rank matrix B injected next to the query matrix and value matrix of each attention layer; The evidence header parameters include the weight parameters of each evidence header. and bias vector .

[0096] The jointly optimized backpropagation path includes: total loss The gradients of the LoRA parameters are backpropagated to each attention layer via the language decoder; the total loss... The gradients of the evidence header parameters are backpropagated through two paths: one path indirectly via the language decoder and fused features, and the other path directly via the evidence calibration loss. The gradients from both paths converge at the evidence header, achieving end-to-end joint optimization. Throughout the optimization process, the original weights of the image-text feature extraction network (i.e., the base weights of the visual encoder and language decoder) are completely frozen and do not participate in gradient updates.

[0097] S6: Perform a two-stage cascaded diagnosis based on the aforementioned auxiliary diagnostic model, output the diagnostic results, and, based on the diagnostic results, backtrack the diagnostic reasoning path in the diagnostic knowledge graph to generate a structured auxiliary diagnostic report. The two-stage cascaded diagnosis includes a first stage, namely the disease type determination stage, which outputs the disease type; and a second stage, which, under the constraint of the disease type, outputs the severity level of the disease type.

[0098] In this embodiment, step S6 specifically includes: S6-1: Input the fusion features into the language decoder of the auxiliary diagnostic model to obtain specific ultrasound signs.

[0099] In this embodiment, the specific ultrasound signs include: blind-end tubular structures, increased diameter, no deformation under compression, thickened appendix wall, and enhanced echogenicity of the mesentery surrounding the appendix corresponding to appendicitis; concentric circle sign in transverse section and sleeve sign or pseudo-kidney sign in longitudinal section corresponding to intussusception; strong echo mass with posterior acoustic shadowing and thickened gallbladder wall corresponding to gallstones; and cystic mass and vortex sign in the adnexal region corresponding to ovarian cyst torsion.

[0100] S6-2: In the disease type determination stage, retrieve differential diagnosis paths that match the specific ultrasound signs from the diagnostic knowledge graph, use the differential diagnosis paths as the decoding context, and adjust the probability of candidate disease types in combination with the child's gender to obtain the disease type; Specifically, the fusion uncertainty quality As a confidence reference, the fusion conflict coefficient As a cross-modal consistency measure, when the fusion conflict coefficient If the threshold is not exceeded, the process directly proceeds to the disease type determination stage. When the fusion conflict coefficient... When the preset threshold is exceeded, the direction of conflict is determined based on the quality of category belief between intermediate image opinions and clinical opinions in each category:

[0101] in, The intermediate opinions of images obtained during the iterative aggregation of DS evidence combination are in the first... Quality of category beliefs in each disease category For clinical opinions in the first Quality of typological beliefs in each disease category This indicates finding the disease category index value with the highest probability. and These are the disease category indexes that best represent the quality of the two category beliefs, when... At that time, the conflict arises from differences in uncertainty, and the two paths discount the quality of their respective category beliefs according to their own uncertainty quality; when At that time, the category belief quality vector with the smaller maximum category belief quality in the intermediate image opinion and the clinical opinion is multiplied by an additional attenuation factor. The fusion conflict coefficient is then recalculated. If the fusion conflict coefficient drops below a preset threshold, the sample proceeds to the second stage based on the recalculated fusion conflict coefficient. If the fusion conflict coefficient still exceeds the preset threshold, the sample is marked as low confidence and does not proceed to the second stage. The disease types output in the first stage include appendicitis, intussusception, gallstones, ovarian cyst torsion, and other acute abdominal conditions.

[0102] S6-3: Under the constraint of disease type, obtain the severity level corresponding to the disease type based on the laboratory inflammation-related indicators and the specific ultrasound signs.

[0103] In this embodiment, the output severity level corresponding to the disease type is divided into mild, moderate and severe.

[0104] For example, when the disease type is appendicitis, the severity level corresponding to the disease type is determined by combining the content levels of white blood cell count (WBC), C-reactive protein (CRP), and interleukin-6 (IL-6), and at the same time combining whether specific ultrasonic signs are accompanied by periappendiceal effusion, abscess, fecal stone or interrupted wall to confirm the inflammation degree. The specific judgment indicators are as follows: Mild: WBC < 12, CRP < 20 mg / L, and IL-6 < 10 pg / mL; Moderate: 12 < WBC < 18, or CRP 20 ~ 60 mg / L, or IL-6 10 ~ 50 pg / mL; Severe: WBC > 18, or CRP > 60 mg / L, or IL-6 > 50 pg / mL.

[0105] Further, according to the disease type and the severity level, the diagnostic reasoning path in the diagnostic knowledge graph is backtracked to generate a structured auxiliary diagnosis report.

[0106] Specifically, starting from the specific ultrasonic sign node, tracing back to the disease type node along the differential diagnosis edge of the diagnostic knowledge graph, and then tracing back to the grading node corresponding to the severity level, the reasoning path formed by the passed nodes and edges is converted into structured text. The auxiliary diagnosis model generates a structured auxiliary diagnosis report based on the structured text. Each conclusion in the diagnosis report can be traced back to the specific nodes and edges in the diagnostic knowledge graph, so that the diagnostic basis is clear and the reasoning path can be reproduced, thereby helping to improve the accuracy of auxiliary diagnosis of acute abdominal pain in children and providing auxiliary reference for clinical decision-making.

[0107] The above is only the specific implementation of the present invention. However, the protection scope of the present invention is not limited thereto. Any person skilled in the art who is within the technical scope disclosed by the present invention can easily think of changes or substitutions, which shall be covered within the protection scope of the present invention. The protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for auxiliary diagnosis of acute abdominal pain in children based on multimodal data fusion, characterized in that, The method includes the following steps: S1: Collect dual-probe ultrasound video sequences, clinical structured information, disease type diagnostic labels, and laboratory inflammation-related indicators of the child; S2: Based on the clinical structured information, the laboratory inflammation-related indicators, and the disease type diagnostic label, predictive factors are screened using Bayesian group LASSO prior and diagnostic prior graph constraints to obtain the diagnostic association factors and their posterior distribution of regression coefficients; S3: Construct a diagnostic knowledge graph and training triples, generate instruction text based on the diagnostic reasoning path in the diagnostic knowledge graph, and inject the instruction text into the question text of the training triples to obtain the enhanced training triples. S4: Based on the ultrasound video sequence, visual features are extracted using a graphic feature extraction network; clinical statistical features are obtained based on the posterior distribution of the regression coefficients of the diagnostic correlation factors and the clinical structured information; the visual features and the clinical statistical features are fused using an evidence head to obtain fused features. S5: Based on the fusion features and the enhanced training triples, jointly optimize the parameters of the LoRA low-rank adaptation layer and the parameters of the evidence head configured in the image and text feature extraction network to obtain the optimized image and text feature extraction network and use it as an auxiliary diagnostic model. S6: Perform a two-stage cascaded diagnosis based on the auxiliary diagnostic model, output the diagnostic results, and based on the diagnostic results, trace back the diagnostic reasoning path in the diagnostic knowledge graph to generate a structured auxiliary diagnostic report.

2. The method for auxiliary diagnosis of acute abdominal pain in children based on multimodal data fusion according to claim 1, characterized in that, Step S2 specifically includes: S2-1: Based on the clinical structured information and the laboratory inflammation-related indicators, the clinical feature vector is obtained through encoding and standardization. S2-2: Construct a multinomial logistic regression model based on the clinical feature vector and the disease type diagnosis label. Apply a Bayesian group LASSO prior distribution to each regression coefficient in the multinomial logistic regression model and perform posterior sampling to obtain the posterior distribution of each regression coefficient. S2-3: Calculate the posterior inclusion probability of each predictor based on the posterior distribution of each regression coefficient, calculate the coefficient of variation of each predictor based on the posterior inclusion probability, and remove unstable predictors based on the coefficient of variation to obtain a set of stable predictors. S2-4: Match the stable predictor set with the association paths of each disease node in the diagnostic prior graph, retain the predictors with association paths as the diagnostic association factors, and extract the corresponding subsets from the posterior distribution of each regression coefficient according to the variable index of the diagnostic association factors to obtain the posterior distribution of the regression coefficients of the diagnostic association factors.

3. The method for auxiliary diagnosis of acute abdominal pain in children based on multimodal data fusion according to claim 2, characterized in that, Step S2-2 specifically includes: S2-2-1: The specific formula for constructing a multinomial logistic regression model is as follows: in, In clinical feature vectors The following child belongs to the first category The probability of such diseases; For the first Regression coefficients corresponding to disease categories This represents the total number of disease types. For the first Regression coefficients corresponding to disease categories; S2-2-2: Applying the Bayesian LASSO prior distribution to each regression coefficient in the multinomial logistic regression model and performing posterior sampling, the specific formula for obtaining the posterior distribution of the regression coefficients is as follows: in, To obtain the posterior distribution of the regression coefficients through posterior sampling, For the observation dataset, Let be the likelihood function, representing the likelihood given the regression coefficients. All observed below Disease labels of training samples The joint probability, For the first The likelihood value of each sample. For the Bayesian group LASSO prior distribution, The sign indicates proportionality.

4. The method for auxiliary diagnosis of acute abdomen in children based on multimodal data fusion according to claim 1, characterized in that, Step S3 specifically includes: S3-1: Construct the diagnostic knowledge graph using ultrasound signs, disease types, and laboratory inflammation-related indicators as nodes, and differential diagnostic relationships and conflict relationships as edges; S3-2: Construct the training triplet consisting of ultrasound images, question text, and answer text; S3-3: Traverse the diagnostic knowledge graph, starting from each ultrasound sign node, extract the complete diagnostic reasoning path to the disease type node along the differential diagnosis relationship edge, and convert the node sequence and edge relationship in the diagnostic reasoning path into instruction text; S3-4: The instruction text is appended as a prefix to the question text of the training triple, and the appended question text is reorganized with the corresponding ultrasound image and answer text to obtain the enhanced training triple.

5. The method for auxiliary diagnosis of acute abdominal pain in children based on multimodal data fusion according to claim 3, characterized in that, Step S4 specifically includes: S4-1: Extract high-frequency keyframes and low-frequency keyframes from the ultrasound video sequence using the visual encoder of the image and text feature extraction network to obtain high-frequency aggregation vectors and low-frequency aggregation vectors. S4-2: Based on the clinical structured information and laboratory inflammation-related indicators of the current sample, obtain the clinical feature vector, and concatenate the clinical feature vector with the mean vector of the posterior distribution of the regression coefficients of the diagnostic association factors to obtain the clinical statistical feature vector. S4-3: Input the high-frequency aggregation vector, the low-frequency aggregation vector, and the clinical statistical feature vector into the trainable evidence head for nonlinear mapping to obtain the high-frequency evidence vector, the low-frequency evidence vector, and the clinical evidence vector, and construct their respective Dirichlet distributions. S4-4: Based on their respective Dirichlet distributions, the high-frequency evidence vector, the low-frequency evidence vector, and the clinical evidence vector are converted into high-frequency opinions, low-frequency opinions, and clinical opinions, respectively. S4-5: The high-frequency opinions, low-frequency opinions, and clinical opinions are iteratively aggregated using the DS evidence combination rule to obtain fusion features.

6. The method for auxiliary diagnosis of acute abdominal pain in children based on multimodal data fusion according to claim 5, characterized in that, Step S4-4 specifically includes: S4-4-1: Using the Dirichlet distribution as the conjugate prior, the high-frequency evidence vector, the low-frequency evidence vector, and the clinical evidence vector are respectively used as the base values ​​of the concentration parameters of the Dirichlet distribution, and are summed with the preset prior concentration parameters to obtain the Dirichlet concentration parameters corresponding to each evidence vector. S4-4-2: Divide the component value of each dimension in each evidence vector by the total Dirichlet strength of the corresponding evidence vector to obtain the class belief quality of each evidence vector in each category. S4-4-3: The uncertainty mass of each evidence vector is obtained by dividing the product of the prior concentration parameter of the Dirichlet distribution and the total number of categories by the total Dirichlet intensity. S4-4-4: Combine the class belief quality and uncertainty quality of each evidence vector in each category to obtain the high-frequency opinion, the low-frequency opinion and the clinical opinion.

7. The method for auxiliary diagnosis of acute abdomen in children based on multimodal data fusion according to claim 6, characterized in that, Step S4-5 specifically includes: S4-5-1: Combine the high-frequency opinion with the low-frequency opinion as DS evidence to obtain the intermediate image opinion; combine the intermediate image opinion with the clinical opinion as DS evidence to obtain the fusion opinion; S4-5-2: Set corresponding reliability coefficients based on the uncertainty quality of intermediate imaging opinions and clinical opinions respectively; S4-5-3: Calculate the conflict coefficient based on the joint Dirichlet strength of the fusion opinions; S4-5-4: When the conflict coefficient is greater than the preset threshold, the quality of the category belief is weighted and discounted according to the corresponding reliability coefficient. The DS evidence is recombined according to the weighted and discounted quality of the category belief to obtain the final fusion opinion. The fusion feature is obtained according to the final fusion opinion.

8. The method for auxiliary diagnosis of acute abdominal pain in children based on multimodal data fusion according to claim 7, characterized in that, Step S5 specifically includes: S5-1: Input the fused features into the language decoder of the image-text feature extraction network, and use the word sequence of the answer text in the enhanced training triples as the supervision target to calculate the model loss; S5-2: Calculate the mean squared error loss based on the difference between the unique heat code of the disease type diagnosis label and the category belief quality of each evidence vector; S5-3: Calculate the variance loss based on the concentration parameters of the Dirichlet distribution corresponding to each evidence vector; S5-4: Calculate the KL divergence penalty loss based on the concentration parameters of the Dirichlet distribution corresponding to the non-real disease type category component in each evidence vector; S5-5: The mean square error loss, the variance loss and the KL divergence penalty loss are weighted and summed to obtain the evidence calibration loss; S5-6: Obtain the total loss based on the model loss and the evidence calibration loss; S5-7: Based on the total loss, the parameters of the pre-configured LoRA low-rank adaptation layer and the parameters of each evidence head in the image feature extraction network are jointly optimized through backpropagation to obtain the auxiliary diagnostic model.

9. A method for auxiliary diagnosis of acute abdominal pain in children based on multimodal data fusion according to claim 8, characterized in that, In step S5-5, the mean square error loss, the variance loss, and the KL divergence penalty loss are weighted and summed to obtain the specific formula for the evidence calibration loss as follows: in, To calibrate the loss for evidence, For the unique heat code of the disease type diagnostic label, For disease category indexing, For the first The total Dirichlet strength of the path evidence vector, For the first The Dirichlet distribution corresponding to the path evidence vector is in the th... Concentration parameters for disease categories To follow the training rounds Increasing regularization weight coefficients, KL divergence is used to measure the difference between two probability distributions. Indicates the calibration concentration parameter The Dirichlet distribution for the concentration parameter; For the first The probability distribution vector of the road evidence vector across each category. This indicates a uniform Dirichlet distribution.

10. A method for auxiliary diagnosis of acute abdomen in children based on multimodal data fusion according to claim 9, characterized in that, Step S6 specifically includes: S6-1: Input the fusion features into the language decoder of the auxiliary diagnostic model to obtain specific ultrasound signs; S6-2: In the disease type determination stage, retrieve differential diagnosis paths that match the specific ultrasound signs from the diagnostic knowledge graph, use the differential diagnosis paths as the decoding context, and adjust the probability of candidate disease types in combination with the child's gender to obtain the disease type; S6-3: Under the constraint of disease type, obtain the severity level corresponding to the disease type based on the laboratory inflammation-related indicators and the specific ultrasound signs.

Images