Method and system for identifying mushroom poisoning

Abstract

The application discloses a mushroom poisoning identification method and system, which is applied to the field of artificial intelligence technology, and comprises the following steps: obtaining a mushroom data set; wherein the mushroom data set comprises a plurality of case samples, each case sample comprises a mushroom associated image, chief complaint text information and biochemical test information, and each case sample has a corresponding label; performing model training processing on the mushroom data set to obtain a mushroom poisoning identification model; and performing identification processing on a to-be-tested case sample by using the mushroom poisoning identification model to obtain a mushroom poisoning identification result; wherein the to-be-tested case sample comprises a to-be-tested mushroom associated image, to-be-tested chief complaint text information and to-be-tested biochemical test information. The application can effectively improve the identification accuracy of mushroom poisoning.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and system for identifying mushroom poisoning. Background Technology

[0002] Poisonous mushrooms are those that can cause poisoning if ingested. To improve the accuracy and efficiency of identifying poisonous mushroom poisoning, it is of great significance to realize the state perception of the poisoning process and the prediction of poisoning in various scenarios such as wild picking, family meals, and medical treatment.

[0003] In related technologies, images of poisonous mushrooms are input into a pre-trained neural network model, which outputs the type of poisoning. However, these technologies rely solely on image-based data, making it difficult to comprehensively depict the symptom evolution and damage process during mushroom poisoning, resulting in low accuracy in identifying mushroom poisoning. Summary of the Invention

[0004] This application provides a method and system for identifying poisonous mushroom poisoning, which improves the accuracy of mushroom poisoning identification.

[0005] On one hand, embodiments of this application provide a method for identifying mushroom poisoning, including the following steps: Obtain a mushroom dataset; wherein the mushroom dataset includes multiple case samples, each case sample includes mushroom-related images, chief complaint text information and biochemical test information, and each case sample has a corresponding label; The aforementioned mushroom dataset was used for model training to obtain a mushroom poisoning identification model; The mushroom poisoning identification model is used to identify and process the test case samples to obtain mushroom poisoning identification results; wherein, the test case samples include the mushroom-related image, the chief complaint text information, and the biochemical test information.

[0006] On the other hand, embodiments of this application provide a mushroom poisoning identification system, including: An acquisition module is used to acquire a mushroom dataset; wherein, the mushroom dataset includes multiple case samples, each case sample includes mushroom-related images, chief complaint text information and biochemical test information, and each case sample has a corresponding label; The first processing module is used to perform model training processing using the poisonous mushroom dataset to obtain a poisonous mushroom poisoning identification model; The second processing module is used to identify and process the test case samples using the mushroom poisoning identification model to obtain mushroom poisoning identification results; wherein, the test case samples include the mushroom-related image to be tested, the chief complaint text information to be tested, and the biochemical test information to be tested.

[0007] According to an embodiment of this application, a method and system for identifying mushroom poisoning are provided, comprising: acquiring a mushroom dataset; wherein the mushroom dataset includes multiple case samples, each case sample including a mushroom-associated image, chief complaint text information, and biochemical test information, and each case sample having a corresponding label; using the mushroom dataset for model training to obtain a mushroom poisoning identification model; using the mushroom poisoning identification model to identify test case samples to obtain mushroom poisoning identification results; wherein the test case sample includes a mushroom-associated image to be tested, chief complaint text information to be tested, and biochemical test information to be tested. This application can effectively improve the identification accuracy of mushroom poisoning.

[0008] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0009] Figure 1 This is a flowchart of a method for identifying mushroom poisoning provided in this application; Figure 2 This is a structural diagram of the mushroom poisoning identification model provided in this application; Figure 3 This is a structural diagram of a mushroom poisoning identification system provided in this application; Figure 4 This is a sample diagram of the interactive interface provided in this application. Detailed Implementation

[0010] 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.

[0011] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0012] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0013] 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 application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0014] To address the shortcomings of related technologies, this application provides a method and system for identifying mushroom poisoning. This method and system aim to improve the accuracy of mushroom poisoning identification by analyzing the clinical characteristics, epidemiological features, and blood purification efficacy of mushroom poisoning.

[0015] It should be noted that in all specific embodiments of this application, whenever processing is required based on relevant data such as case samples and samples to be tested, the permission or consent of the subject will be obtained first. Furthermore, the collection, use, and processing of this data will comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require obtaining relevant data such as case samples and samples to be tested, separate permission or consent from the subject will be obtained through pop-ups or redirection to a confirmation page. Only after obtaining the subject's separate permission or consent will the necessary relevant data, such as case samples and samples to be tested, be obtained to enable the embodiments of this application to function properly.

[0016] It is important to emphasize that the case samples and test samples involved in the embodiments of this application are data pre-stored in a database, and not data obtained in real time from patients. That is, the embodiments of this application do not use living human bodies as the object. Furthermore, the final output of the embodiments of this application is intended to provide doctors with diagnostic references, and is not the final diagnostic result; the final diagnosis of mushroom poisoning still needs to be determined by a doctor. In other words, the embodiments of this application do not have the direct purpose of obtaining a diagnosis of mushroom poisoning. Therefore, the embodiments of this application do not belong to the methods of diagnosing and treating diseases.

[0017] First, the implementation steps of the mushroom poisoning identification method provided in this application will be described in detail below with reference to the accompanying drawings.

[0018] This application provides a method for identifying mushroom poisoning, which can be applied to terminals, servers, or software running on either terminal or server. Terminals can be tablets, laptops, desktop computers, etc., but are not limited to these. Servers can be independent physical servers, server clusters or distributed systems composed of multiple physical servers, or cloud servers providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. Furthermore, a server can be a node server in a blockchain network, but is not limited to these. Blockchain is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0019] Reference Figure 1 , Figure 1 This is a flowchart of a mushroom poisoning identification method provided in this application, which may include the following steps S101-S103.

[0020] S101, Obtain the mushroom dataset.

[0021] It should be noted that the mushroom poisoning dataset includes multiple case samples, which can be obtained through a pre-defined database. Each case sample includes mushroom-related images, chief complaint text information, and biochemical test information. The mushroom-related images may include, but are not limited to, images of mushrooms and images of food residue; the chief complaint text information may include, but is not limited to, chief complaint information and intake history information; and the biochemical test information may include, but is not limited to, data on liver and kidney function, biochemical indicators, coagulation function, blood pressure, and heart rate.

[0022] In addition, each case sample has a corresponding label, which can include mushroom type, toxin type, poisoning type, and severity level. Mushroom type can be classified into Amanita, Mycorrhiza, and Agaricus types, etc.; toxin type can be classified into nephrotoxic and hepatotoxic types, etc.; poisoning type can be classified into gastrointestinal, hemolytic, and neurotoxic types, etc.; and severity level can be classified into mild, moderate, and severe types. It should be understood that each type is assigned a corresponding numerical value. For example, the severity level is assigned a value of 0 for mild, 1 for moderate, and 2 for severe, but this is not a limitation.

[0023] In this step, based on existing multimodal real cases, multimodal data samples are generated through automatic sampling and conditional combination based on clinical statistical laws and typical manifestations of mushroom poisoning. Specifically, this includes: (1) Sampling based on case statistical distribution: Random sampling is performed based on the historical case distribution probability for variables such as incubation period, symptom combination, and mushroom type to ensure that the simulated cases are consistent with the real clinical scenario. (2) Multimodal conditional combination: Samples that match each other are selected from the three modalities of image, text, and numerical data to form a complete case task. For example, “images of suspected Amanita mushrooms” are combined with “text of delayed nausea and vomiting symptoms” and “numerical indicators of liver function elevation trend” to form a consistent scenario. (3) Toxicological constraint mechanism: The toxicological characteristics of different mushroom categories are strictly referred to during the combination process (such as the delayed liver damage pattern of Amanita toxin and the acute neurological symptom pattern of filamentous fungi) to ensure that the simulated data does not violate medical laws. Accordingly, multiple case samples containing mushroom-related images, chief complaint text information, and biochemical test information are collected to obtain a mushroom dataset.

[0024] Optionally, after obtaining the poisonous mushroom dataset, preprocessing is performed. For image-based data, preprocessing methods may include, but are not limited to, format normalization (i.e., format conversion), size standardization, brightness and contrast correction, and automatic cropping to remove irrelevant background, remove background noise, and highlight suspicious tissue features of the poisonous mushroom morphology or vomit. For text-based data, preprocessing methods may include, but are not limited to, text segmentation, stop word filtering, medical terminology normalization, and extraction of key information related to the incubation period of poisoning, symptom spectrum, and intake history. For numerical-based data, preprocessing methods may include, but are not limited to, unit standardization, missing word completion, and Z-score normalization to eliminate scale differences caused by different detection devices. This preprocessing method can effectively improve the quality of the poisonous mushroom dataset, providing accurate data input for subsequent steps.

[0025] S102, using the poisonous mushroom dataset for model training, to obtain a poisonous mushroom poisoning identification model.

[0026] In this step, the pre-set model is trained and tested using a mushroom dataset to obtain a mushroom poisoning identification model, which is then used to identify samples of the cases to be tested.

[0027] Optionally, the type of mushroom poisoning identification model can be set according to the actual situation. For example, the mushroom poisoning identification model can be a multimodal large model MycoDiag-LLM, but it is not limited to this.

[0028] S103. The mushroom poisoning identification model is used to identify and process the test case samples to obtain the mushroom poisoning identification results.

[0029] It should be noted that the sample of the case to be tested includes the image of the mushroom to be tested, the text information of the chief complaint to be tested, and the biochemical test information to be tested. The definitions of these data are the same as those of the image of the mushroom to be tested, the text information of the chief complaint, and the biochemical test information, and will not be repeated here.

[0030] In this step, a sample of the patient to be tested is obtained, which includes the image associated with the mushroom to be tested, the patient's chief complaint text information, and the biochemical test information. This data is then input into the mushroom poisoning identification model, which outputs the mushroom poisoning identification results, such as mushroom type, toxin type, poisoning classification, and severity level.

[0031] Therefore, this application embodiment, by fully considering data such as images of suspected poisonous mushrooms or food residues, textual descriptions of clinical symptoms, and blood test indicators, achieves the training and processing of a mushroom poisoning identification model, and applies this mushroom poisoning identification model to real-time mushroom poisoning identification. This can comprehensively cover the symptom evolution and damage process in the process of mushroom poisoning, thereby effectively improving the identification accuracy of mushroom poisoning.

[0032] In some embodiments, before step S102, which uses the mushroom dataset for model training to obtain a mushroom poisoning identification model, the method further includes steps S201-S205, where the target case sample is any single case sample: S201, Perform diffusion generation processing on the mushroom association image of the target case sample to obtain at least one mushroom association generated image of the target case sample; S202, synthesize and generate the chief complaint text information of the target case sample to obtain at least one chief complaint text information of the target case sample; S203, Perform perturbation simulation processing on the biochemical test information of the target case sample to obtain at least one biochemical test generation information of the target case sample; S204, integrate at least one mushroom-associated image, at least one chief complaint text, and at least one biochemical test result of the target case sample to obtain at least one generated case sample of the target case sample; wherein, the generated case sample includes a single mushroom-associated image, a single chief complaint text, and a single biochemical test result, and the label of the generated case sample is the same as the label of the target case sample. S205, add all generated case samples as new case samples to the mushroom dataset.

[0033] Specifically, current sample data on mushroom poisoning is relatively scarce. To address this, this implementation introduces a generative data augmentation mechanism to expand the sample size associated with mushroom poisoning under limited sample conditions, thereby increasing the sample diversity of mushroom poisoning scenarios. This helps improve the training accuracy of the mushroom poisoning identification model. In this generative data augmentation mechanism, key modalities such as images, text, and numerical data are generatively augmented. For example, image samples with different lighting, viewing angles, and degrees of decay are generated; semantically equivalent symptom complaint texts are synthesized; and indicators such as liver function and coagulation are reasonably perturbed. This constructs samples representing multiple mushroom types and multiple poisoning stages, enriching the mushroom dataset.

[0034] More specifically, taking any case sample as the target case sample, the amplification operation performed on the target case sample includes: first, performing diffusion generation processing on the mushroom association image of the target case sample to obtain at least one mushroom association generated image of the target case sample; second, performing synthesis generation processing on the chief complaint text information of the target case sample to obtain at least one chief complaint text generated information of the target case sample; and third, performing perturbation simulation processing on the biochemical test information of the target case sample to obtain at least one biochemical test generated information of the target case sample. Then, integrating the at least one mushroom association generated image, at least one chief complaint text generated information, and at least one biochemical test generated information of the target case sample, the principle of integration is to randomly select a single mushroom association generated image, a single chief complaint text generated information, and a single biochemical test generated information and integrate them into a new sample, which is denoted as the generated case sample. The label of the generated case sample is the label of the target case sample. Through this integration processing, at least one generated case sample of the target case sample can be obtained. Finally, all generated case samples are added as new case samples to the mushroom dataset, thereby expanding the data sample in the mushroom dataset.

[0035] In some embodiments, step S201 above, which involves performing diffusion generation processing on the mushroom association image of the target case sample to obtain at least one mushroom association generated image of the target case sample, includes the following step S01: S01, using a diffusion model to perform at least one noise addition and reverse denoising process on the mushroom association image of the target case sample, to obtain at least one mushroom association generated image of the target case sample.

[0036] In this embodiment, for the image modality data in the target case sample, the following steps are taken: At least one operation of adding controlled Gaussian noise and reverse denoising is performed on the mushroom-associated image. This operation of adding controlled Gaussian noise and reverse denoising is the core generation mechanism of the diffusion model. Gaussian noise is gradually injected into the original mushroom-associated image through a forward diffusion process, and then a trained reverse diffusion network is used to gradually denoise and reconstruct the mushroom-associated image with injected noise. This generates a new image that varies in lighting, angle, and degree of putrefaction but still maintains the key structural information of the mushrooms, i.e., at least one mushroom-associated image of the target case sample. Therefore, this method can generate images simulating different lighting, angles, and degrees of putrefaction, effectively covering complex mushroom poisoning scenarios and thus improving the sample diversity of mushroom poisoning scenarios. The type of diffusion model is not specifically limited; for example, it can be a DDPM (Denoising Diffusion Probabilistic Models) model, but it is not limited to this.

[0037] Furthermore, the data enhancement processing for image modalities is mainly used to simulate mushroom image samples under different lighting environments, acquisition angles, degrees of decay, and background complexity. For a single mushroom-related image, the process of adding noise and reverse denoising specifically includes the following steps: The first step, forward noise addition: Based on a set time step, controllable Gaussian noise is gradually injected into the image, causing it to gradually degenerate into a noisy image while maintaining its contour structure. The noise intensity is controlled by a preset scheduling function, allowing for automatic simulation of the degree of decay, changes in illumination, or variations in imaging noise.

[0038] The second step is inverse denoising: A pre-trained backward diffusion network iteratively recovers a new image from an image with random noise or added noise. In each inverse step, a noise prediction network removes noise and restores the mushroom morphology, ensuring the final image is structurally similar to the original. Figure 1 While similar in style, they exhibit diversity in terms of lighting, background composition, shooting angle, and degree of decay.

[0039] The third step is morphological preservation constraints: In order to ensure the recognizability of key parts of the poisonous mushroom (such as cap morphology, gill structure, and color distribution), the generated image is screened by the Structure Similarity Index Measure (SSIM) and the results of poisonous mushroom region segmentation to ensure that the enhanced image does not produce structural distortion or loss of key features.

[0040] Specifically, for each generated image, firstly, the sharpness score of the generated image is calculated using the Laplacian variance sharpness index, i.e. , Indicates the clarity score. This represents the generated image. If the sharpness score of the generated image is less than a preset sharpness threshold, the generated image is considered blurry and is deleted. Otherwise, the structural similarity score between the generated image and the original image is calculated. If the structural similarity score is less than a preset similarity threshold, the key mushroom morphology is considered damaged, and the generated image is deleted. Otherwise, a pre-trained segmentation network is used to process the original and generated images to obtain the mushroom region segmentation area in the original image and the mushroom region segmentation area in the generated image. If the ratio of the mushroom region segmentation area in the generated image to the mushroom region segmentation area in the original image is less than a preset area ratio threshold, the main structure of the mushroom is considered damaged or incomplete, and the generated image is deleted.

[0041] In subsequent steps, the final generated images are written as new samples into the poisonous mushroom dataset to improve the robustness of the poisonous mushroom poisoning identification model under real and complex collection conditions.

[0042] In some embodiments, step S202 above involves synthesizing and generating the chief complaint text information of the target case sample to obtain at least one chief complaint text generation information of the target case sample, including the following step S02: S02, using a diffusion language model to generate at least one chief complaint text information of the target case sample, thereby obtaining at least one chief complaint text information of the target case sample.

[0043] In this embodiment, the data of the text modality in the target case sample includes: First, extracting key clinical information such as incubation period, symptom combination, and mode of ingestion from the chief complaint text information as conditions for the diffusion language model, aiming to maintain core medical semantics; then, based on these conditions, performing at least one generation process on the chief complaint text information through the diffusion language model to obtain at least one text with equivalent expression, i.e., at least one chief complaint text generation information of the target case sample. Thus, this method can synthesize semantically equivalent symptom chief complaint texts, effectively covering complex mushroom poisoning scenarios and improving the sample diversity of mushroom poisoning scenarios. For example, if the chief complaint text is "nausea, vomiting, and abdominal pain occurred 6 hours later," this method can generate chief complaint text generation information such as "significant nausea accompanied by repeated vomiting and upper abdominal pain occurred several hours after eating wild mushrooms." The type of diffusion language model is not specifically limited; it can be a continuous spatial diffusion language model (Continuous DLMs), but is not limited to this.

[0044] Furthermore, augmentation processing for text-based data is primarily used to generate chief complaint sentences, intake history descriptions, and symptom texts that are semantically consistent with the original case but expressed differently. For a single chief complaint text, the process of generating it in a single step specifically includes the following steps: The first step is to extract key clinical information: extract elements such as mushroom type, incubation period description, chief complaint symptoms, and accompanying symptoms from the original text, such as "nausea, vomiting, and abdominal pain occurred 6 hours after eating wild mushrooms".

[0045] The second step is conditional encoding: the above-mentioned key clinical elements are input into the diffusion language model as conditions, so that the model can maintain these medical semantics when generating text.

[0046] The third step is forward noise addition: the diffusion language model maps the original text to a latent space vector and injects noise into it, so that its expression deviates from the fixed pattern of the original text, providing a space for linguistic diversity for subsequent reconstruction.

[0047] The fourth step is reverse generation: Through the reverse diffusion process, the model gradually reconstructs the text from the noise vector, but can output generated text with different language styles, text formats, or description methods. For example, the original text is "Six hours after eating wild mushrooms, the patient experienced nausea, repeated vomiting, and abdominal pain." The generated text could be "Several hours after eating wild mushrooms, the patient experienced significant nausea, persistent vomiting, and significant upper abdominal pain." The core clinical semantics of the original and generated texts remain consistent, but the expression methods are diverse.

[0048] The fifth step is text checking: Consistency text checking will use at least one method, such as rule verification, keyword matching, and symptom logic checking, to ensure that the generated text does not lose key medical information such as mushroom category, incubation period, and symptom combination.

[0049] Specifically, firstly, the number of keywords matched by the generated text is counted from a preset keyword set. If the ratio of the number of keywords to the total number of keywords in the keyword set is less than a preset keyword coverage threshold, it indicates that the generated text is missing keywords related to mushroom poisoning, and the generated text is deleted; otherwise, the generated text is retained. Optionally, after keyword matching, a medical knowledge base is used to perform a logical consistency check on the retained generated text for the combination of mushroom type, incubation period, and symptoms. If a combination contradicts the toxicological rules, the generated text is directly deleted; otherwise, the generated text is retained.

[0050] In subsequent steps, the final generated text is written as a new sample into the poisonous mushroom dataset to improve the robustness of the poisonous mushroom poisoning identification model under real and complex collection conditions.

[0051] In some embodiments, step S203 above involves perturbation simulation processing of the biochemical test information of the target case sample to obtain at least one biochemical test generation information of the target case sample, including the following step S03: S03, within the perturbation range of the biochemical test information of the target case sample, perform at least one noise addition and condition correction process on the biochemical test information of the target case sample to obtain at least one biochemical test generation information of the target case sample.

[0052] In this embodiment, for the numerical modal data in the target case sample, the following steps are taken: First, the upper and lower limits of the biochemical test information are set based on the typical biochemical change range of the relevant mushroom toxins. Then, within this range, the biochemical test information is subjected to at least one small Gaussian noise perturbation, and conditional correction is performed on the perturbated biochemical test information in conjunction with toxicological trends (e.g., the delayed significant increase pattern of liver enzymes caused by amatoxins, and the trend of prolonged prothrombin time). This generates a new sample that conforms to the distribution of specific mushroom toxicological characteristics, i.e., at least one biochemical test result of the target case sample, such as the delayed significant increase pattern of liver function related to amatoxins. Therefore, this method can generate indicator data that are associated with mushroom poisoning but different from the original sample, thus effectively covering complex mushroom poisoning scenarios and improving the sample diversity of mushroom poisoning scenarios.

[0053] Specifically, augmentation processing for numerical modal data is mainly used to expand structured data such as liver and kidney function, biochemical indicators, and coagulation indicators, enabling the model to learn physiological change patterns of different mushroom types and poisoning stages. For single biochemical test information, the process of adding noise and conditional correction for a single test may include the following steps: The first step is range setting: Based on information such as toxicology literature and expert experience, a reasonable range of variation from normal values ​​to poisoning risk values ​​is set for each indicator (numerical vector) in the biochemical test information, and characteristic trend constraints are introduced for the toxicity of different mushrooms. For example, the delayed increase of alanine aminotransferase / aspartate aminotransferase (ALT / AST) is a typical manifestation of amatoxin poisoning, electrolyte disturbances caused by neurotoxic mushrooms, or small changes in biochemical indicators of gastrointestinal irritant mushrooms.

[0054] The second step is forward noise addition: Based on the above range, controllable Gaussian noise is applied to the numerical vectors of each item in the biochemical test information to generate samples with reasonable perturbation at different levels; at the same time, the changes in indicators under different poisoning stages are simulated by trend offset terms. For example, when liver enzymes are rising, the relevant indicators are added with an offset that conforms to the trend range, so that the generated data has both distribution diversity and consistency with the toxicological trend.

[0055] The third step is reverse denoising: the conditional diffusion model is used to reverse reconstruct the denoised numerical vector, and new index combinations are generated through multi-step denoising iterations, so that the generated results conform to medical statistical laws in the overall distribution and are consistent with the toxicity characteristics of the corresponding mushrooms.

[0056] The fourth step is the generation and verification process: The generated numerical samples undergo rigorous medical constraint screening, including checks on whether they meet toxicological trends, fall within medically safe ranges, and maintain reasonable correlations between key indicators. Samples that do not meet the constraints are automatically removed, ensuring that the final enhanced samples are authentic, reasonable, and accurately reflect the biochemical characteristics of different toxin types and stages of poisoning, thereby significantly improving the model's generalization ability and robustness on structured data.

[0057] Specifically, firstly, if at least one indicator in the generated numerical sample exceeds the corresponding numerical range, the generated numerical sample is considered to be inconsistent with medical safety standards, and is therefore deleted. Otherwise, the deviation metric between the generated numerical sample and the corresponding typical trend curve of toxic mushrooms is calculated. If the deviation metric is greater than a preset threshold, it indicates that the generated numerical sample does not conform to the existing toxicological trend, and is therefore deleted. Otherwise, the correlation coefficient (e.g., Pearson correlation coefficient) between the first target indicator and the second target indicator in the generated numerical sample is calculated. The first target indicator can be any indicator, and the second target indicator can be any indicator other than the first target indicator. If the correlation coefficient between the first target indicator and the second target indicator is less than a preset threshold, it indicates that the generated numerical sample significantly violates clinical statistical laws, and is therefore deleted from the generated numerical sample. Optionally, after detecting the correlation of indicators, in order to meet the characteristics of poisoning diagnosis, an indicator importance score is introduced, that is, indicators with an importance score lower than a preset importance threshold are deleted from the generated numerical samples. This aims to enable the model to explicitly establish a medical chain association from symptom description to changes in liver function and then to the suspected type of poisonous mushroom.

[0058] In subsequent steps, the final generated numerical samples are written as new samples into the poisonous mushroom dataset to improve the robustness of the poisonous mushroom poisoning identification model under real and complex collection conditions.

[0059] In summary, this application's embodiments introduce a multi-level screening mechanism in data augmentation processing of image, text, and numerical modalities. This mechanism retains only samples that meet constraints such as clarity, structural similarity, keyword completeness, toxicological trends, and relevance in terms of image morphology, symptom semantics, and biochemical indicators. This significantly reduces pseudo-sample noise caused by diffusion generation, ensuring that the multimodal data entering the training and inference stages has sufficient authenticity and medical interpretability, thereby improving sample quality.

[0060] In some implementations, refer to Figure 2 The aforementioned mushroom poisoning identification model includes: The image feature extraction module is used to extract features from the images associated with the poisonous mushrooms to be tested, thereby obtaining the image features of the poisonous mushrooms; The text feature extraction module is used to extract features from the chief complaint text information to obtain the chief complaint text features; The indicator feature extraction module is used to extract features from the biochemical test information to obtain biochemical test features. The fusion module is used to fuse the features of the poisonous mushroom image, the features of the chief complaint text, and the features of the biochemical test to obtain the target fused features; The multi-task decision layer is used for identification processing based on target fusion features to obtain the mushroom poisoning identification result.

[0061] It should be noted that the results of mushroom poisoning identification include at least one of the following: mushroom type, toxin type, poisoning classification, and severity level of poisoning.

[0062] In this embodiment, the mushroom poisoning identification model may include an image feature extraction module, a text feature extraction module, an indicator feature extraction module, a cross-modal attention module, and a multi-task decision layer. The image feature extraction module, text feature extraction module, and indicator feature extraction module are connected in parallel, and a fusion module and a multi-task decision layer are sequentially connected after each of these modules.

[0063] In the image modality, considering that images of suspected poisonous mushroom samples or food residues often exhibit characteristics such as local fragmentation, inconsistent levels of decay, and dim lighting, a local-global dual-stream coding path is constructed in the image feature extraction module. Specifically, the image feature extraction module can include an image mapping layer and a local-global dual-stream coding path. The local-global dual-stream coding path includes a global path and a local path, wherein the image mapping layer is used to perform coded mapping on the image to obtain image coded features, which are represented as follows: , Represents image coding features, Data representing the input image modality. This represents the operation of encoding and mapping an image. This encoding and mapping operation allows data from different image modalities to be projected more naturally into a unified semantic space, achieving stable alignment of cross-modal features. The global path is responsible for extracting overall morphological information such as cap morphology and gill structure from the image encoded features, corresponding to global feature extraction. The local path focuses on key details such as detritus in the image encoded features, corresponding to local feature extraction. The feature information obtained through these two paths constitutes the mushroom image association features. The global path can be implemented using models such as convolutional neural networks, and the local path can be implemented using models such as Vision-Transformer, but is not limited to these. Optionally, an adaptive illumination correction module can be added after the local-global dual-stream encoding path to improve the visibility and robustness of the mushroom image association features.

[0064] In the text modality, the text feature extraction module can include a text mapping layer and a text feature extraction layer. The text mapping layer is used to encode and map the text information of the subject complaint to obtain text encoding features, which can be represented as follows: , Indicates text encoding features, This represents data representing the input text modality. This represents the operation of encoding and mapping text. This encoding and mapping operation allows text modality data to be projected more naturally into a unified semantic space, achieving stable alignment of cross-modal features. To adapt to the characteristics of medical narratives about mushroom poisoning, a text feature extraction layer is used to extract features from the text encoding features, enabling it to explicitly capture the sequential logical relationship between the incubation period, symptom combination, and toxin damage mechanism, thereby improving the cross-modal correlation ability between text modality data and data from other modalities. The text feature extraction layer can be implemented using a medical BERT model, but is not limited to this.

[0065] In numerical mode, the indicator feature extraction module may include an indicator mapping layer and an indicator feature extraction layer. The indicator mapping layer is used to encode and map the biochemical test information to obtain indicator coding features, which can be represented as follows: , Indicates the characteristics of the indicator coding. This represents the input numerical modal data. This represents the operation of encoding and mapping numerical values. This encoding and mapping operation allows data from numerical modalities to be projected more naturally into a unified semantic space, achieving stable alignment of cross-modal features. The indicator feature extraction layer is used to introduce a trend-sensitive encoding mechanism based on toxicological knowledge. This mechanism extracts features from the encoded indicators, aiming to enhance the model's ability to identify the increasing or decreasing patterns of key indicators over time, such as automatically identifying the delayed increase trend of liver enzyme indicators in amatoxin-related cases. The indicator feature extraction layer can be implemented using a Trend-Aware Embedding mechanism, but is not limited to this.

[0066] By using the feature extraction methods described above for each modality of data, we can comprehensively locate the general features associated with mushroom poisoning and effectively capture the key features associated with mushroom poisoning hidden in different modalities of data, thereby helping to improve the accuracy of mushroom poisoning detection.

[0067] The fusion module is used to realize feature interaction and fusion between image modalities and text modalities, image modalities and numerical modalities, and text modalities and numerical modalities, so as to obtain the target fused features.

[0068] The multi-task decision layer can be set with four classification heads, corresponding to the four tasks of identifying poisonous mushroom type, identifying toxin type, identifying poisoning type, and identifying poisoning severity. By processing the target fusion features through the multi-task decision layer, the poisonous mushroom poisoning identification results including poisonous mushroom type, toxin type, poisoning type, and poisoning severity can be output.

[0069] Therefore, this implementation method uses different feature extraction modules to uncover key features associated with mushroom poisoning hidden in image, text, and numerical modalities, respectively. Based on this, it achieves feature interaction and fusion between image and text modalities, image and numerical modalities, and text and numerical modalities. Subsequently, based on the fusion results, multi-task decision-making is performed to obtain mushroom poisoning identification results including mushroom type, toxin type, poisoning classification, and severity of poisoning. This can effectively improve the identification accuracy of mushroom poisoning.

[0070] In some implementations, the fusion module fuses the mushroom image features, the chief complaint text features, and the biochemical test features to obtain the target fused features, including: The features of the poisonous mushroom image and the features of the main complaint text are fused to obtain the first fused feature; The image features and biochemical test features of poisonous mushrooms are fused together to obtain the second fused feature; The chief complaint text features and biochemical test features are fused together to obtain the third fused feature; The first fusion feature, the second fusion feature, and the third fusion feature are fused together to obtain the target fusion feature.

[0071] In this embodiment, firstly, using the mushroom image association feature as the query and the complaint text feature as the key and value, a multi-head attention mechanism is used to fuse the mushroom image association feature and the complaint text feature to obtain a first fused feature. Then, using the mushroom image association feature as the query and the biochemical test feature as the key and value, the mushroom image association feature and the biochemical test feature are fused to obtain a second fused feature. Finally, using the complaint text feature as the query and the biochemical test feature as the key and value, the complaint text feature and the biochemical test feature are fused to obtain a third fused feature. Then, to enhance the model's adaptive modeling of the importance of different modalities, the first, second, and third fused features are scored to obtain attention weights corresponding to each feature. This scoring process can be implemented using the Softmax function. Finally, based on the aforementioned attention weights, the first, second, and third fused features are weighted and summed to obtain the final target fused feature. Optionally, before weighted summation, the attention weights of each fusion feature are dynamically adjusted according to the clinical characteristics corresponding to the mushroom category. For example, in the amatoxins scenario, the attention weight of numerical modalities (such as liver function indicators) is increased, while in the gastrointestinal irritation scenario, the weight of text modalities is increased.

[0072] It is evident that when faced with challenges such as significant semantic differences between image, text, and numerical modalities, inconsistent data quality, and interference from redundant information, traditional neural network models struggle to directly achieve high-quality modality alignment and decision fusion, resulting in poor model performance and impacting the accuracy of mushroom poisoning identification.

[0073] To address this, this implementation introduces fine-grained modal interaction, employing a multi-head attention mechanism to achieve interaction and fusion between different modalities. Specifically, through the interaction and fusion between image and text modalities, the correspondence between the appearance features of poisonous mushrooms (such as the cap scales and annulus of Amanita mushrooms) and the chief complaint (such as "vomiting 6 hours after ingesting a red-ringed mushroom") is captured. Through the interaction and fusion between image and numerical modalities, the potential link between the morphology of poisonous mushrooms (such as the slimy cap of Mycorrhiza mushrooms) and biochemical indicators (such as elevated liver enzymes and abnormal coagulation) is captured. Through the interaction and fusion between text and numerical modalities, the pathological association between symptom descriptions (such as "jaundice and darkened urine") and liver and kidney function indicators (such as bilirubin and creatinine) is constructed. This pairwise fusion approach constructs a mapping relationship between local modalities, enabling the model to comprehensively understand the complex relationships between modalities, rather than relying on single-dimensional matching. Furthermore, the multi-head attention mechanism allows the model to pay different attention to different modalities during fusion. For example, when processing images and text, some heads focus on "the color of the poisonous mushroom" and "the color description" in the complaint, some heads focus on "the shape of the stem" and "the edible part", and some heads focus on "images of food residue" and "ingestion history".

[0074] Following fine-grained modal interactions, adaptive weight adjustment is introduced, i.e., scoring is performed using the Softmax function, and the target fusion features are further obtained based on these scores. This allows the model to dynamically learn the importance of each modality combination in different cases. For example, if the image is clear but the text is missing, the weights of the image modality and the numerical modality will dominate the decision, thereby enhancing the model's sensitivity to the associated features of mushroom poisoning, ensuring the accuracy of these associated features, and helping to improve the identification accuracy of mushroom poisoning.

[0075] In some implementations, step S102 above, which involves training a model using a poisonous mushroom dataset to obtain a poisonous mushroom poisoning identification model, includes the following steps S301-S302: S301, use a general dataset and a weakly labeled dataset to pre-train the preset model to obtain the pre-trained preset model; S302, using the poisonous mushroom dataset to train the pre-trained preset model to obtain a poisonous mushroom poisoning identification model.

[0076] In this embodiment, firstly, multiple sets of multimodal samples are randomly selected from the poisonous mushroom target dataset to construct a training task set. Tasks may include identifying poisonous mushroom types, identifying toxin types, identifying poisoning classifications, and identifying the severity level of poisoning. A single task is represented as... , Data representing image modalities, Data representing text modality, Data representing numerical modes, Labels are used to represent the model. The training task set is divided into a support set and a query set for each task, used for the fusion training of the model backbone and modal encoder. Simultaneously, an enhanced test set is constructed by randomly selecting sample data from simulated real-world mushroom poisoning cases and dividing it into a support set and a query set.

[0077] Then, training is performed. In terms of specific training strategies, this step adopts a two-stage mechanism of pre-training and domain fine-tuning.

[0078] In the pre-training phase, a large-scale general medical image-text corpus and a partially weakly labeled multimodal dataset related to mushroom poisoning are constructed. Based on this, the image feature extraction module, text feature extraction module, and indicator feature extraction module in the pre-defined model are pre-trained, aiming to initialize the image modality, text modality, and structured diagnostic indicators. Specifically, the image modality pre-training content covers common tasks in medical scenarios such as illumination changes, local defects, and tissue morphology discrimination; the text modality pre-training content includes medical terminology understanding, symptom description recognition, and extraction of language patterns related to intake history; and the numerical modality pre-training content learns the basic distribution characteristics of liver, kidney, biochemical, and coagulation indicators through a self-supervised trend prediction task. In this stage, the model does not directly perform the mushroom poisoning identification task, but instead learns cross-modal representation capabilities through a large number of general medical semantic tasks. This allows the model to master basic medical semantics, avoiding being limited to information about mushroom poisoning. The weakly labeled multimodal dataset allows the model to learn image features, text descriptions, and indicator patterns related to mushroom poisoning in advance, providing rich prior knowledge for the subsequent fine-tuning stage. This pre-training phase establishes a preliminary correspondence between visual features, linguistic features, and structured features in a unified semantic space, thus laying the foundation for fine-tuning for subsequent specific tasks.

[0079] In the domain fine-tuning phase, a few-sample task training is performed using a cleaned and labeled training task set. Specifically, each case is divided into a support set and a query set. The images, symptom text, and biochemical indicators in the support set jointly guide the model to establish more accurate cross-modal semantic associations in the task scenario. The query set is used to evaluate the model's ability to distinguish specific mushroom types, toxin categories, and severity of poisoning. Once the model's discriminative ability meets the preset requirements, a mushroom poisoning identification model is output. In this phase, the general capabilities of the model's pre-training are transformed into specific capabilities for mushroom poisoning identification, significantly improving the accuracy of distinguishing mushroom types, toxin categories, and severity of poisoning.

[0080] In some implementations, the mushroom poisoning identification results include the mushroom type and the severity level of poisoning; the loss function of the mushroom poisoning identification model during training includes a loss function for the mushroom type, a loss function for the severity level of poisoning, and a loss function for the case sample and the generated case sample.

[0081] In this embodiment, a composite loss function is provided during the training of the mushroom poisoning identification model, as follows: In the classification task, a multi-class Focal loss function is used for the mushroom type, as shown in the following formula (1): (1); In equation (1), This represents the loss function for mushroom types; Indicates the number of case samples; Indicates the number of types of poisonous mushrooms; Indicates the first The mushroom type in the true label of each case sample should be understood to be represented by a numerical value, for example, Amanita genus type is represented by 1, Mycorrhiza genus type is represented by 2, Agaricus genus type is represented by 3, but is not limited to this; The output of the model is the first The predicted probability of a sample of cases; Indicates the type balance coefficient; This represents the modulation factor. Here, the mushroom type identification task is a multi-class classification task, but it suffers from problems such as class imbalance (e.g., fewer samples of some highly toxic types and more samples of common types) and confusion of difficult-to-distinguish samples. To address this, this loss function, through a class balance coefficient and a modulation factor, ensures that the model does not ignore rare but dangerous mushroom types, improving the recall rate of minority classes. Simultaneously, it reduces the loss weight of easily distinguishable samples, forcing the model to focus on morphologically similar and difficult-to-distinguish samples, thus improving overall classification accuracy.

[0082] The normalized regression loss function is used to determine the severity level of poisoning, as shown in the following formula (2): (2); In equation (2), This represents the loss function based on the severity level of poisoning. The output of the model represents the first... The severity level of poisoning in each case sample; Indicates the first The severity level of poisoning in the true label of each case sample; This indicates the number of severity levels of poisoning, for example, the severity levels of poisoning include mild, moderate, and severe types. The value is 3. Here, the severity level of poisoning is a non-discrete ordered rank variable. Directly using the original data to calculate the regression loss has problems such as excessive numerical differences (for example, the numerical difference between severe and mild may cause the model to be oversensitive to high-level samples during training) and loss of relative rank relationships (for example, the transition from moderate to severe may be ignored, leading to inaccurate prediction of boundary samples). This loss function can effectively avoid the problem of gradient explosion or training oscillation caused by excessive numerical differences in the model, while prompting the model to pay more attention to the relationship between ranks and improve its prediction accuracy for boundary samples.

[0083] A consistency loss function is applied to the case sample and at least one generated case sample, as shown in the following formula (3): , (3); In equation (3), Indicates the first Loss of consistency among individual case samples; Indicates the first The number of case samples generated per case sample; Indicates the first Target fusion features of individual case samples; Indicates the first The first case sample The target fusion features of each generated case sample; This represents the loss function applied to both the original case samples and the generated case samples. The consistency loss function here is applied to the unified semantic space features resulting from the fusion of the three modalities via a dedicated encoder and cross-modal attention—the target fusion features—rather than constraining each individual modality separately. By limiting the distance between the target fusion features of the original case samples and the enhanced target fusion features of the generated case samples, reasonable perturbations in local features across different modalities are allowed while maintaining overall semantic consistency. This forces the model to learn the essential features of the data, helping to improve its accuracy in diagnosing mushroom poisoning.

[0084] In summary, the loss function of the above mushroom poisoning identification model during training satisfies the following formula (4): (4); In equation (4), The weights of the loss function are indicated by the severity level of the poisoning. The weights of the loss function for generating case samples and case samples are indicated. This represents the loss function used during training of the mushroom poisoning identification model.

[0085] Therefore, this implementation method uses a multi-class Focal loss function for mushroom types, a normalized regression loss function for poisoning severity levels, and a consistency loss function for case samples and at least one generated case sample. By weighting these three loss functions, a loss function suitable for multi-task scenarios involving poisonous mushrooms can be obtained, thereby improving the training accuracy and generalization ability of the model.

[0086] In addition, refer to Figure 3 This application provides a mushroom poisoning identification system, which may include: The acquisition module 401 is used to acquire the mushroom dataset; wherein, the mushroom dataset includes multiple case samples, each case sample includes mushroom-related images, chief complaint text information and biochemical test information, and each case sample has a corresponding label.

[0087] The first processing module 402 is used to train a model using a poisonous mushroom dataset to obtain a poisonous mushroom poisoning identification model. The second processing module 403 is used to identify and process the test case samples using the mushroom poisoning identification model to obtain the mushroom poisoning identification results; wherein, the test case samples include the mushroom-related image to be tested, the chief complaint text information to be tested, and the biochemical test information to be tested.

[0088] In some implementations, the mushroom poisoning model is integrated into a complete software and hardware system. Through multimodal data access, a graphical user interface, and edge deployment, real-time intelligent diagnosis and decision support for mushroom poisoning cases are achieved. The system platform as a whole includes a multimodal data acquisition module, a software interaction module, a diagnostic reasoning platform, and a log and security management module. These modules communicate with each other through standardized data interfaces.

[0089] The multimodal acquisition module corresponds to the acquisition module mentioned above. The multimodal acquisition module is used to complete the synchronous acquisition and standardized conversion of three types of data: images, text, and numerical data in clinical or field environments. Specifically, it includes: (1) Image acquisition submodule, which uses a high-definition visible light camera, mobile terminal, or portable imaging device to acquire images of suspected poisonous mushroom samples, food residue on the table, or poisonous mushroom fragments in vomit, and records the shooting time, location, and information of the acquisition personnel; it supports batch import of relevant images already existing in the hospital imaging system. (2) Text acquisition submodule, which allows doctors or nurses to enter the patient's basic information, intake history (intake time, intake amount, self-collected or purchased source, information of co-eaters), chief complaint, and physical examination results through the terminal interface. The text content is structured according to the preset field template during the input stage. (3) Numerical acquisition submodule, which connects with the laboratory information system or bedside rapid testing equipment to automatically acquire biochemical indicators such as blood routine, liver and kidney function, myocardial enzymes, coagulation function, and necessary blood gas analysis results, and adds patient number and timestamp to each record. (4) Data standardization submodule performs format verification and field mapping on the three types of data, and converts the original multi-source data into a unified data structure used by the system, providing standardized input for subsequent preprocessing and inference.

[0090] The software interaction module uses a graphical user interface developed with Qt Designer or similar tools to enable visual interaction between medical staff and the diagnostic system, such as... Figure 4 As shown, it specifically includes: (1) Patient management area, which supports the creation, retrieval and modification of patient information, and can quickly locate historical records by name, hospital number or outpatient number, and display past mushroom poisoning related medical information. (2) Data entry and examination area, which provides image upload area, text entry window and test result import panel, where users can complete the loading of suspected mushroom images, filling in symptoms and intake history and confirming blood test data step by step, and the area provides prompts for missing fields. (3) Model configuration and inference control area, which allows users to select model version, inference mode (only category recognition or simultaneous assessment of severity, etc.), running device (local / remote server), and provides a one-click start inference button. (4) Result display and report export area, which displays the inference results in graphic and text form, including suspected mushroom category, poisoning type, severity level, recommended monitoring level and transfer suggestions, and provides contribution prompts for key modalities; it supports automatically generating PDF or standardized electronic reports for easy printing and archiving.

[0091] The diagnostic inference platform is responsible for loading a lightweight mushroom poisoning identification model and completing online inference. It can be deployed on hospital servers or edge computing devices. It corresponds to the first and second processing modules mentioned above, and specifically includes: (1) Model service submodule, which encapsulates the trained multimodal large model through a service interface, provides standardized APIs, and supports batch inference and concurrent requests; the model is compressed before deployment through distillation, pruning and quantization to reduce the computation and storage overhead during inference. (2) Inference execution submodule, which receives standardized multimodal input from the interactive interface, completes data preprocessing, feature encoding, cross-modal fusion and decision output in sequence, and returns the inference results to the front-end interface; on resource-constrained edge devices, only the necessary sub-models can be loaded to ensure a second-level response. (3) Edge deployment submodule integrates the model and necessary operating environment into portable terminals, emergency vehicle workstations or local servers in primary hospitals. Even when there is no stable network or access to the central server, it can still independently complete the intelligent diagnosis of mushroom poisoning. At the same time, it reserves a remote update interface so that when the central terminal releases a new model or knowledge base, it can be upgraded online through a secure channel.

[0092] The logging and security management module is used to ensure system traceability and data security compliance, specifically including: (1) Inference log recording: For each inference request, the patient's anonymity identifier, input data summary, model version number, inference time, and output results are recorded to facilitate subsequent model effect evaluation and problem tracking. (2) Access control and audit control: The operation permissions of different users are restricted through account and role management mechanisms, and sensitive operations involving model configuration changes, data export, and system maintenance are audited and recorded. (3) Data privacy and transmission security: Patient data stored on the server and edge devices are encrypted and stored. An encrypted communication protocol is used between the client and the server to prevent data from being illegally stolen or tampered with during transmission.

[0093] Furthermore, in edge deployment practices, the system can integrate the lightweight mushroom poisoning identification model with the necessary operating environment into portable computing terminals or emergency vehicle workstations, performing inference through local computation and reducing reliance on central servers and network bandwidth. When the central terminal releases a new model or updates the mushroom knowledge base, the update package can be synchronized to edge devices through a secure channel, enabling continuous iterative upgrades of the system.

[0094] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0095] In summary, this application's embodiments, by fusing multimodal data such as images of suspected poisonous mushroom samples or food residues, symptom and intake history texts, and blood test results, construct a mushroom poisoning identification model in a unified semantic space. This achieves comprehensive perception and accurate inference of the condition of mushroom poisoning patients, overcoming the problems of insufficient information dimensions and low accuracy in traditional single-modal diagnostics. The introduction of a diffusion model for generative data augmentation significantly improves the model's generalization ability under conditions of limited real-world mushroom poisoning case samples and uneven distribution of mushroom types, effectively alleviating the diagnostic instability caused by scarce and imbalanced training data. By designing a cross-modal attention mechanism and a modal-aware encoder, the technical challenges of semantic alignment and feature fusion of heterogeneous image, text, and numerical data are solved, significantly enhancing the model's understanding and robustness to complex poisoning scenarios and organ damage evolution processes. The system's overall structure is lightweight and can be deployed on edge computing platforms, enabling real-time diagnosis and decision support in primary healthcare institutions, emergency scenes, and environments with limited network conditions. It possesses good clinical applicability, response speed, and engineering promotion value.

[0096] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0097] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated to the contrary, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional technology for an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0098] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several programs to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable programs for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, a program execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can retrieve and execute a program from or in conjunction with such a program execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit a program for use by or in conjunction with a program execution system, apparatus, or device.

[0100] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0101] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable program execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0102] In the foregoing description of this specification, the references to terms such as "one embodiment," "another embodiment," or "some embodiments," etc., indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0103] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0104] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for identifying mushroom poisoning, characterized in that, Includes the following steps: Obtain a mushroom dataset; wherein the mushroom dataset includes multiple case samples, each case sample includes mushroom-related images, chief complaint text information and biochemical test information, and each case sample has a corresponding label; The aforementioned mushroom dataset was used for model training to obtain a mushroom poisoning identification model; The mushroom poisoning identification model is used to identify and process the test case samples to obtain mushroom poisoning identification results; wherein, the test case samples include the image associated with the mushroom to be tested, the text information of the chief complaint to be tested, and the biochemical test information to be tested; the mushroom poisoning identification results include mushroom type, toxin type, poisoning classification, and poisoning severity level; Before using the mushroom dataset for model training to obtain the mushroom poisoning identification model, the method further includes: The mushroom association image of the target case sample is subjected to diffusion generation processing to obtain at least one mushroom association generated image of the target case sample; The step of performing diffusion generation processing on the mushroom association image of the target case sample to obtain at least one mushroom association generated image of the target case sample includes: The mushroom association image of the target case sample is processed by adding noise and reverse denoising at least once using a diffusion model to obtain at least one mushroom association image of the target case sample. The steps of adding noise and reverse denoising in a single step for each of the mushroom-related images include: Gaussian noise is gradually injected into the mushroom-associated image to gradually degrade the mushroom-associated image into a noisy image, and the noisy image is iteratively restored into a new image through a backdiffusion network, wherein the new image refers to the generated image; For each generated image, the sharpness score of the generated image is calculated using the Laplacian variance sharpness index. If the sharpness score is less than a preset sharpness threshold, the generated image is deleted; otherwise, the structural similarity score between the generated image and the original image is calculated. If the structural similarity score is less than a preset similarity threshold, the generated image is deleted; otherwise, a pre-trained segmentation network is used to process the original image and the generated image to obtain the mushroom region segmentation area of ​​the original image and the mushroom region segmentation area of ​​the generated image. If the ratio of the mushroom region segmentation area of ​​the generated image to the mushroom region segmentation area of ​​the original image is less than a preset area ratio threshold, the generated image is deleted; otherwise, the generated image is retained. The retained generated image is identified as the mushroom-associated generated image.

2. The method according to claim 1, characterized in that, Before using the mushroom dataset for model training to obtain the mushroom poisoning identification model, the method further includes: The chief complaint text information of the target case sample is synthesized and generated to obtain at least one chief complaint text information of the target case sample; The biochemical test information of the target case sample is subjected to perturbation simulation processing to obtain at least one biochemical test generation information of the target case sample; The at least one mushroom-associated image, at least one chief complaint text, and at least one biochemical test result of the target case sample are integrated to obtain at least one generated case sample of the target case sample; wherein, the generated case sample includes a single mushroom-associated image, a single chief complaint text, and a single biochemical test result, and the tags of the generated case sample are the same as those of the target case sample. Add all the generated case samples as new case samples to the mushroom dataset; The target case sample can be any one of the case samples.

3. The method according to claim 2, characterized in that, The process of synthesizing and generating the chief complaint text information of the target case sample to obtain at least one chief complaint text generation information of the target case sample includes: The chief complaint text information of the target case sample is generated at least once using a diffusion language model to obtain at least one chief complaint text information of the target case sample.

4. The method according to claim 2, characterized in that, The perturbation simulation processing of the biochemical test information of the target case sample to obtain at least one biochemical test generation information of the target case sample includes: Within the perturbation range of the biochemical test information of the target case sample, the biochemical test information of the target case sample is subjected to at least one noise addition and condition correction process to obtain at least one biochemical test generation information of the target case sample.

5. The method according to claim 1, characterized in that, The mushroom poisoning identification model includes: The image feature extraction module is used to extract features from the image associated with the poisonous mushroom to be tested, and obtain the image features of the poisonous mushroom; The text feature extraction module is used to extract features from the subject complaint text information to obtain the subject complaint text features; The indicator feature extraction module is used to extract features from the biochemical test information to be tested, and obtain biochemical test features; The fusion module is used to fuse the mushroom image features, the chief complaint text features, and the biochemical test features to obtain the target fused features; A multi-task decision layer is used to perform identification processing based on the target fusion features to obtain the mushroom poisoning identification result.

6. The method according to claim 5, characterized in that, The process of fusing the mushroom image features, the chief complaint text features, and the biochemical test features to obtain the target fused features includes: The mushroom image features and the main complaint text features are fused together to obtain the first fused feature; The mushroom image features and the biochemical test features are fused together to obtain a second fused feature; The chief complaint text features and the biochemical test features are fused together to obtain a third fused feature; The first fusion feature, the second fusion feature, and the third fusion feature are fused together to obtain the target fusion feature.

7. The method according to claim 1, characterized in that, The step of using the mushroom dataset for model training to obtain a mushroom poisoning identification model includes: The preset model is pre-trained using a general dataset and a weakly labeled dataset to obtain the pre-trained preset model; The pre-trained preset model is trained using the poisonous mushroom dataset to obtain the poisonous mushroom poisoning identification model.

8. The method according to claim 1, characterized in that, The mushroom poisoning identification result includes the mushroom type and the severity level of poisoning; the loss function of the mushroom poisoning identification model during training includes a loss function for the mushroom type, a loss function for the severity level of poisoning, and a loss function for the case sample and the generated case sample.

9. A mushroom poisoning identification system, characterized in that, include: An acquisition module is used to acquire a mushroom dataset; wherein, the mushroom dataset includes multiple case samples, each case sample includes mushroom-related images, chief complaint text information and biochemical test information, and each case sample has a corresponding label; The first processing module is used to perform model training processing using the poisonous mushroom dataset to obtain a poisonous mushroom poisoning identification model; The second processing module is used to identify and process the test case samples using the mushroom poisoning identification model to obtain mushroom poisoning identification results; wherein, the test case samples include the mushroom-related image to be tested, the chief complaint text information to be tested, and the biochemical test information to be tested; the mushroom poisoning identification results include mushroom type, toxin type, poisoning classification, and poisoning severity level; Prior to training the model using the mushroom dataset to obtain the mushroom poisoning identification model, the system further includes: The mushroom association image of the target case sample is subjected to diffusion generation processing to obtain at least one mushroom association generated image of the target case sample; The step of performing diffusion generation processing on the mushroom association image of the target case sample to obtain at least one mushroom association generated image of the target case sample includes: The mushroom association image of the target case sample is processed by adding noise and reverse denoising at least once using a diffusion model to obtain at least one mushroom association image of the target case sample. The steps of adding noise and reverse denoising in a single step for each of the mushroom-related images include: Gaussian noise is gradually injected into the mushroom-associated image to gradually degrade the mushroom-associated image into a noisy image, and the noisy image is iteratively restored into a new image through a backdiffusion network, wherein the new image refers to the generated image; For each generated image, the sharpness score of the generated image is calculated using the Laplacian variance sharpness index. If the sharpness score is less than a preset sharpness threshold, the generated image is deleted; otherwise, the structural similarity score between the generated image and the original image is calculated. If the structural similarity score is less than a preset similarity threshold, the generated image is deleted; otherwise, a pre-trained segmentation network is used to process the original image and the generated image to obtain the mushroom region segmentation area of ​​the original image and the mushroom region segmentation area of ​​the generated image. If the ratio of the mushroom region segmentation area of ​​the generated image to the mushroom region segmentation area of ​​the original image is less than a preset area ratio threshold, the generated image is deleted; otherwise, the generated image is retained. The retained generated image is identified as the mushroom-associated generated image.

Images