Skin contact poison rapid identification method and system based on AI image recognition

Abstract

The present application relates to the technical field of poison detection, and particularly relates to a skin contact poison rapid identification method and system based on AI image recognition, which comprises the following steps: based on a skin lesion area image collected by a mobile terminal, determining an image quality parameter, when the image quality parameter meets a preset collection threshold, executing a preprocessing strategy; extracting a morphological feature set of the skin lesion area, the morphological feature set comprising a color space component, a texture period parameter and an edge curvature feature of a blister structure; uploading the morphological feature set to a cloud AI model, and performing multi-label matching based on a poison knowledge base, the multi-label matching comprising: determining a candidate poison list according to a feature similarity coefficient, and determining an output priority based on a confidence score of each candidate poison; and generating a structured report comprising a poison type, a poisoning degree classification and first-aid suggestions according to a matching result. The present application improves the timeliness and accuracy of skin contact poison identification.

Description

Technical Field

[0001] This invention relates to the field of toxic substance detection technology, and in particular to a rapid identification method and system for skin contact toxic substances based on AI image recognition. Background Technology

[0002] Chemical poisoning through skin contact is widespread in industrial production, medical aesthetics, laboratory operations, and daily life. Upon skin contact, toxins can cause morphological damage such as discoloration, blisters, tissue necrosis, or specific distribution patterns. In clinical emergency care, rapid identification of the toxin is crucial for developing an antidote plan, preventing further absorption, and reducing the risk of systemic poisoning.

[0003] Currently, the identification of skin contact toxins mainly relies on morphological observation and laboratory chemical analysis. Morphological observation is based on empirical inferences made by clinicians according to the color, distribution, shape, and clinical symptoms of skin lesions; laboratory methods involve collecting skin swabs or blood samples and performing qualitative and quantitative analysis using high-precision equipment such as liquid chromatography-mass spectrometry (LC-MS) or gas chromatography-mass spectrometry (GC-MS).

[0004] However, existing technologies have significant drawbacks. First, morphological observation is highly dependent on the physician's personal experience, making it subjective. Different toxins initially present with highly similar skin lesions, and these lesion characteristics are easily affected by variables such as the patient's skin color, age, toxin concentration, and duration of exposure, resulting in poor robustness of manual identification and a high risk of misdiagnosis or missed diagnosis in complex environments. Second, while laboratory chemical analysis methods are highly accurate, the process is cumbersome—sample collection, transportation, and analysis can take hours or even days, failing to meet the real-time requirements of the "golden window" for emergency treatment. Third, high-end analytical equipment is bulky and expensive to maintain, making rapid screening at the poisoning site impossible, resulting in a significant time and space gap between initial treatment and subsequent detection.

[0005] This lag in identification directly leads to delays or deviations in treatment plans. In severe poisoning cases, if the type of poison cannot be accurately identified in the first instance, incorrect neutralization treatments may be implemented or the optimal decontamination time may be missed, allowing the poison to continue to enter the bloodstream through the skin, thereby causing irreversible and severe systemic damage such as acute kidney injury and multiple organ failure.

[0006] Therefore, there is an urgent need to develop a non-invasive identification method and system that can quickly, objectively, and accurately identify the characteristics of skin contact toxins without the need for large laboratory equipment, in order to achieve early and accurate diagnosis and rapid intervention in chemical poisoning. Summary of the Invention

[0007] To address this issue, the present invention provides a rapid identification method and system for skin contact toxins based on AI image recognition, which overcomes the problem that existing methods for effectively identifying the type of skin contact toxins without invasiveness can easily lead to missed opportunities for optimal treatment due to misjudgment or detection delays.

[0008] To achieve the above objectives, this invention provides a rapid identification method for skin contact toxins based on AI image recognition, comprising: Based on the images of skin lesion areas collected by the mobile device, image quality parameters are determined. When the image quality parameters meet the preset acquisition threshold, a preprocessing strategy is executed. The preprocessing strategy includes determining the kernel function size of the denoising algorithm based on the image noise level, determining the gain coefficient of image enhancement based on the color difference contrast between the lesion area and healthy skin, and determining the segmentation boundary of the lesion area based on the gradient distribution of skin texture. After preprocessing, a morphological feature set of the skin lesion area is extracted, which includes color space components, texture periodic parameters, and edge curvature features of the herpes structure. The morphological feature set is uploaded to the cloud AI model, and multi-label matching is performed based on the poison knowledge base. The multi-label matching includes: determining a list of candidate poisons based on feature similarity coefficients, and determining the output priority based on the confidence score of each candidate poison. Based on the matching results, a structured report is generated that includes the type of poison, the severity of poisoning, and emergency treatment recommendations. The severity of poisoning is determined based on the color shift of the lesion area, the edema range index, and the feature fusion weight.

[0009] Furthermore, the image quality parameters include illumination uniformity, focus sharpness, and motion blur coefficient; When the uniformity of illumination is lower than the preset illumination threshold, local histogram equalization is enabled for compensation. When the focus sharpness is lower than the preset sharpness threshold, the mobile device is triggered to refocus and resample. When the motion blur coefficient is greater than the preset blur threshold, the deconvolution restoration algorithm is executed, and the number of deconvolution iterations is determined based on the estimated length of the blur kernel.

[0010] Furthermore, the process of determining the segmentation boundaries of the lesion area includes: A lightweight segmentation network based on the U-Net architecture generates initial candidate regions based on the gradient change rate of skin texture. Based on the edge confidence map of the initial candidate regions, determine the seed points for the growth of connected components; The stopping criterion for region growth is determined based on the Euclidean distance of color difference between adjacent pixels and the texture difference. When the stopping criterion is triggered, a closed lesion region segmentation mask is output.

[0011] Furthermore, the color space components are jointly extracted from the three color spaces: RGB, HSV, and CIELAB. The texture periodicity parameters are determined based on the gray-level co-occurrence matrix, including energy value, contrast value, inverse difference moment value, and correlation coefficient; The edge curvature characteristics of the herpes structure are determined by calculating the curvature extreme values ​​at the boundary points of the lesion area, and are classified into unilocular herpes structures or multilocular fused herpes structures based on the distribution density of the curvature extreme values.

[0012] Furthermore, the feature similarity coefficient is calculated by mapping the extracted morphological feature set to a high-dimensional feature space and calculating the cosine similarity with each standard feature vector in the toxicology knowledge base. If the cosine similarity is greater than the first similarity threshold, the corresponding poison will be added to the candidate poison list. If the cosine similarity is between the second similarity threshold and the first similarity threshold, the auxiliary discrimination network is activated to determine whether to add it to the candidate list based on the morphological evolution time sequence characteristics of the lesion area. The auxiliary discrimination network is a time-series prediction model based on a long short-term memory network.

[0013] Furthermore, the confidence score is determined by weighting the overall confidence score of each candidate toxicant based on the feature similarity coefficient of the candidate toxicant, the epidemiological prior probability in the toxicant knowledge base, and the contact area index of the lesion area. The output priority is sorted from high to low according to the overall confidence level; When the difference between the highest overall confidence level and the second highest overall confidence level is less than the preset difference threshold, a list of toxic substances is output, and clinical symptom information is marked as needed.

[0014] Furthermore, the severity of poisoning is classified into three levels: mild, moderate, and severe. The chromaticity offset is the angle difference between the skin color vector of the diseased area and the healthy skin area; The edema extent index is the percentage of the area of ​​the raised portion of the lesion region. The feature fusion weights are adaptively adjusted according to the type of toxic substance. The weight of the chromaticity shift corresponding to corrosive toxic substances is higher than the weight of the edema range index, and the weight of the herpes structure feature corresponding to blistering toxic substances is higher than the weight of the chromaticity shift.

[0015] Furthermore, the first aid recommendations include decontamination procedure guidelines and tips on the use of antagonists; The decontamination operation guidelines determine the type of rinsing solution based on the acid-base properties of the identified toxic substance. The type of rinsing solution is a neutral buffer solution, a weakly acidic solution, or a weakly alkaline solution. The antagonist usage prompt is generated based on the identified poison-specific antidote, and the route of administration and dosage reference range of the antidote are output simultaneously.

[0016] Furthermore, the structured report also includes image acquisition timestamps, GPS location information, and anonymized thumbnails of the lesion area; The structured report is synchronized to the pre-bound medical institution terminal via the HL7 / FHIR medical information exchange standard protocol; When the identification result corresponds to a highly toxic substance, an emergency warning is automatically triggered. The recipients of the emergency warning include the nearest emergency medical center and the registered contact terminal.

[0017] This invention also provides a rapid skin contact toxicant identification system based on AI image recognition, comprising: The mobile acquisition module is configured to acquire images of skin lesion areas through a camera and perform quality assessment and preprocessing on the images. The preprocessing includes denoising, enhancement, and lesion area segmentation based on U-Net. The cloud-based AI analysis module is communicatively connected to the mobile acquisition module. The cloud-based AI analysis module contains a toxicology knowledge base and a multi-label classification model. The multi-label classification model is used to output a list of candidate toxicants and a comprehensive confidence level based on the received morphological feature set. The results output module, which is connected to the cloud AI analysis module, is configured to generate a structured report based on the comprehensive confidence level and match corresponding first aid recommendations according to the severity of poisoning. The data feedback module is connected to the mobile terminal acquisition module and the cloud AI analysis module, respectively. It is configured to collect the difference data between the actual diagnostic results and the model output, and trigger the incremental learning and update of the multi-label classification model based on the difference data.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention collects images of skin lesions on a mobile device and performs real-time analysis using a cloud-based AI model. It eliminates the need for blood or skin swab samples and large laboratory equipment such as liquid chromatography-mass spectrometry, enabling rapid identification of poison types at the scene of poisoning or in a clinic setting. From image acquisition to the generation of a structured report, it takes only tens of seconds to several minutes, significantly shortening the identification cycle and meeting the timeliness requirements of the "golden window" for emergency treatment. It also overcomes the technical shortcomings of existing laboratory analysis methods, which have excessively long cycles.

[0019] This invention extracts quantitative indicators such as color components, gray-level co-occurrence matrix texture parameters, and herpes edge curvature features of skin lesion areas in three color spaces (RGB, HSV, and CIELAB) and performs standardized comparisons with a toxicology knowledge base. This completely abandons the subjective judgment method that relies on the physician's personal experience in traditional morphological observation methods. By combining the cosine distance calculation of feature similarity coefficients, a dual-threshold candidate toxicology screening mechanism, and a weighted score of comprehensive confidence, the identification results are made repeatable and objective, thereby improving the reliability of toxicology identification.

[0020] This invention addresses the technical challenge of highly similar skin lesion appearances in the initial stages of exposure to different toxins. It employs a two-layer screening mechanism consisting of a first similarity threshold and a second similarity threshold. For ambiguous samples with similarity between the two thresholds, an auxiliary discrimination network based on a long short-term memory network is activated. The network utilizes the morphological evolution time-series characteristics of the lesion area for secondary discrimination. When the difference between the highest overall confidence score and the second highest confidence score is less than a preset difference threshold, a parallel toxin warning is output, avoiding misjudgments that may be caused by a single output, thereby further improving the reliability of toxin identification.

[0021] This invention adaptively adjusts the feature fusion weights according to the type of toxic substance being identified. For corrosive toxic substances, the weight of color shift is increased, and for vesicular toxic substances, the weight of herpes structural features is increased. This makes the grading of poisoning severity more consistent with the pathological damage characteristics of different toxic substances, thereby further improving the reliability of toxic substance identification.

[0022] This invention generates emergency advice that includes decontamination operation guidelines and antagonist usage tips simultaneously. It automatically determines the type of rinsing solution based on the acid-base properties of the poison and generates the administration route and dosage reference range based on the specific antidote. This forms a complete closed loop from poison identification to treatment plan, reducing the decision-making threshold for on-site emergency personnel or non-specialist doctors. It improves the reliability of poison identification and enhances the timeliness of identification.

[0023] This invention collects the difference data between actual clinical diagnostic results and model output by setting up a data feedback module, and triggers incremental learning and updates of the multi-label classification model based on the difference data. This enables the system to continuously enhance its ability to identify rare toxins and variant skin lesions as clinical cases accumulate, solving the problem that traditional static models are difficult to adapt to clinical diversity. It achieves self-evolution of identification performance, thereby further improving the reliability of toxin identification.

[0024] This invention automatically triggers an emergency alert when the identified poison belongs to the list of highly toxic substances, synchronizing a structured report to the nearest emergency center and registered contact terminals. It also achieves data exchange with pre-bound medical institution terminals via the HL7 / FHIR medical information exchange standard protocol, enabling the receiving hospital to know the type and severity of the poison in advance, allowing for better preparation of antidotes and emergency response. This creates a seamless connection from the scene to the pre-hospital stage, further enhancing the timeliness of poison identification. Attached Figure Description

[0025] Figure 1 This is a flowchart of a rapid identification method for skin contact toxins according to an embodiment of the present invention; Figure 2 This is a flowchart of mobile image acquisition and preprocessing according to an embodiment of the present invention; Figure 3 This is a module connection diagram of the rapid identification system for skin contact toxins according to an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0027] It should be noted that the data in this embodiment are all derived from a comprehensive analysis and evaluation of historical test data and corresponding historical test results from the three months prior to this test. Those skilled in the art will understand that the determination of the above-mentioned parameters for any single item in this invention can be achieved by selecting the value with the highest percentage based on the data distribution as the preset standard parameter, using weighted summation to obtain the value as the preset standard parameter, substituting each historical data point into a specific formula and using the value obtained from that formula as the preset standard parameter, or other selection methods, as long as the invention can clearly define different specific situations in the single-item judgment process through the obtained values.

[0028] Please see Figure 1 , Figure 2 As shown, the embodiment of the present invention provides a rapid identification method for skin contact toxins based on AI image recognition, comprising: Step S1: Based on the skin lesion area image acquired by the mobile device, determine the image quality parameters. When the image quality parameters meet the preset acquisition threshold, execute the preprocessing strategy. The preprocessing strategy includes determining the kernel function size of the denoising algorithm based on the image noise level, determining the gain coefficient of image enhancement based on the color difference contrast between the lesion area and healthy skin, and determining the segmentation boundary of the lesion area based on the gradient distribution of skin texture. If the image quality parameters do not meet the preset acquisition threshold, prompt the user to re-acquire the image until the quality requirements are met. Step S2: After preprocessing, extract the morphological feature set of the skin lesion area. The morphological feature set includes color space components, texture periodic parameters, and edge curvature features of the herpes structure. Step S3: Upload the morphological feature set to the cloud AI model and perform multi-label matching based on the toxic substance knowledge base. The multi-label matching includes: determining a list of candidate toxic substances based on the feature similarity coefficient and determining the output priority based on the confidence score of each candidate toxic substance. Step S4: Generate a structured report containing the type of poison, the degree of poisoning, and emergency treatment recommendations based on the matching results. The degree of poisoning is determined based on the color shift of the lesion area, the edema range index, and the feature fusion weight.

[0029] Specifically, this invention provides a rapid, objective, and non-invasive method for identifying the type of toxin in skin lesions resulting from skin contact with chemical toxins. The patient's skin lesion image to be identified is designated as the target image, and the skin lesion area within the target image is used as the core analysis object. The toxin knowledge base constructed in this invention is designed for various clinically confirmed cases of skin contact poisoning, including but not limited to acidic corrosive toxins, alkaline corrosive toxins, blistering toxins (such as mustard gas), organophosphorus pesticides, and irritant toxins. For a single lesion area, its multi-dimensional features, such as color, texture, and morphology, are acquired, and a feature vector is constructed. The feature vector is compared with the standard feature vector in the toxin knowledge base, and the most matching toxin type and corresponding emergency treatment plan are output.

[0030] This invention utilizes several clinical validation records. Each record documents at least one instance of identifying a case of skin contact poisoning, including image quality parameters, morphological feature sets, feature similarity coefficients, overall confidence levels, poisoning severity grading, and information reliability coefficients. Each record also includes a corresponding diagnostic marker, indicating whether the accuracy of the poison identification result aligns with clinical practice. It is understood that users can determine whether the accuracy of the poison identification result meets requirements based on the clinical diagnosis. For example, the criteria for accuracy may include, but are not limited to, whether the identified poison type is consistent with laboratory blood test or skin swab analysis results.

[0031] Specifically, the dynamic index of each toxic substance category is determined periodically based on the historical identification accuracy and image acquisition frequency, and the category of each toxic substance category is determined based on the dynamic index. Toxic substance categories with a dynamic index greater than the preset dynamic index are recorded as high-frequency contact toxic substances.

[0032] Specifically, this invention employs a one-cycle model optimization period, the duration of which can be determined by the user based on actual clinical needs. The higher the user's real-time requirements for toxicant identification accuracy, the shorter the model optimization period. One proposed model optimization period is 30 days. At the end of each optimization period, the dynamic index of each toxicant category is determined based on historical identification accuracy and image acquisition frequency, and the category of each toxicant category is determined based on the dynamic index. For a single toxicant category, if the dynamic index of that category is greater than a preset dynamic index, the category is classified as a high-frequency exposure toxicant; if the dynamic index is less than or equal to the preset dynamic index, the category is classified as a low-frequency exposure toxicant.

[0033] For a single poison category, the dynamic index is a weighted sum of historical identification accuracy and image acquisition frequency. The historical identification accuracy is the ratio of the number of times the poison category is correctly identified within a preset dynamic evaluation phase to the total number of times the poison category is identified. The image acquisition frequency is the number of skin lesion images corresponding to the poison category acquired within the preset dynamic evaluation phase. The end time of the preset dynamic evaluation phase is the end time of the current model optimization cycle. The duration of the preset dynamic evaluation phase can be determined by the user based on the actual work scenario; the higher the user's requirement for poison identification accuracy, the longer the duration of the preset dynamic evaluation phase. One possible value for the duration of the preset dynamic evaluation phase is ten times the duration of the model optimization cycle.

[0034] The value of the preset dynamic index can be determined by the user based on the actual work scenario. This embodiment of the invention provides a method for determining the preset dynamic index: the minimum value of the dynamic index of each frequently exposed toxin in clinical verification records that meets the user's requirements for the accuracy of toxin identification is recorded as the preset dynamic index.

[0035] Specifically, the image quality parameters include illumination uniformity, focus sharpness, and motion blur coefficient; When the uniformity of illumination is lower than the preset illumination threshold, local histogram equalization is enabled for compensation. When the focus sharpness is lower than the preset sharpness threshold, the mobile device is triggered to refocus and resample. When the motion blur coefficient is greater than the preset blur threshold, the deconvolution restoration algorithm is executed, and the number of deconvolution iterations is determined based on the estimated length of the blur kernel.

[0036] In this embodiment of the invention, the illumination uniformity is calculated as follows: the image is divided into several sub-blocks, and the average brightness of each sub-block is calculated. Illumination uniformity = standard deviation of sub-block brightness / global average brightness. The lower the illumination uniformity, the more likely there are local overexposed or underexposed areas in the image. The value of the preset illumination threshold can be determined by the user according to the actual working scenario. One preset illumination threshold value is provided, which is 0.15.

[0037] The focus sharpness is calculated by calculating the variance of the Laplacian gradient of the image; a higher sharpness value indicates more accurate focusing. The preset sharpness threshold is provided with a value of 100.

[0038] The motion blur coefficient is calculated by performing a Fourier transform on the image and calculating the ratio of low-frequency components to high-frequency components in the frequency domain. A larger ratio indicates more severe motion blur. The preset blur threshold is provided as a set value of 0.6.

[0039] In the deconvolution restoration algorithm, the estimated length of the blur kernel is determined by detecting the edge diffusion width in the motion direction. The number of deconvolution iterations is positively correlated with the estimated length of the blur kernel, specifically: iteration count = max(5, ceil(estimated length of blur kernel / 2)). This quality compensation mechanism ensures that the image entering the feature extraction stage has sufficient recognizability, avoiding feature extraction deviations caused by poor acquisition quality.

[0040] Specifically, the process of determining the segmentation boundaries of the lesion area includes: A lightweight segmentation network based on the U-Net architecture generates initial candidate regions based on the gradient change rate of skin texture. Based on the edge confidence map of the initial candidate regions, determine the seed points for the growth of connected components; The stopping criterion for region growth is determined based on the Euclidean distance of color difference between adjacent pixels and the texture difference. When the stopping criterion is triggered, a closed lesion region segmentation mask is output.

[0041] Specifically, the lightweight segmentation network of the U-Net architecture uses MobileNetV2 as the encoder backbone network to reduce the computational load on mobile devices. The gradient change rate of skin texture is calculated using the Sobel operator, and pixels with a gradient change rate greater than a preset gradient threshold constitute the boundaries of the initial candidate regions. The edge confidence map is generated by the softmax output layer of the segmentation network, and pixels with a confidence level higher than 0.7 are selected as seed points.

[0042] The formula for calculating the Euclidean distance of color difference between adjacent pixels is: In the formula, R, G, B are the color component values ​​of the pixel in the RGB space.

[0043] The formula for calculating the texture difference is: In the formula, fi,k is the k-th LBP texture feature value within the local window centered at pixel i.

[0044] The stopping criterion is as follows: growth stops when the Euclidean distance of the color difference between the pixel to be grown and the boundary pixel of the current region is greater than a preset color difference threshold (valued at 30) or the texture difference is greater than a preset texture difference threshold (valued at 0.4). Through this segmentation mechanism, lesion areas and healthy skin areas can be accurately separated, providing accurate regions of interest for subsequent feature extraction.

[0045] Specifically, the color space components are extracted jointly from the three color spaces: RGB, HSV, and CIELAB. The texture periodicity parameters are determined based on the gray-level co-occurrence matrix, including energy value, contrast value, inverse difference moment value, and correlation coefficient; The edge curvature characteristics of the herpes structure are determined by calculating the curvature extreme values ​​at the boundary points of the lesion area, and are classified into unilocular herpes structures or multilocular fused herpes structures based on the distribution density of the curvature extreme values.

[0046] Specifically, the RGB color space is used to extract basic color components, the HSV color space is used to extract hue and saturation features (which are robust to changes in lighting), and the CIELAB color space is used to extract color difference features consistent with human visual perception. The three spaces are jointly used to extract a 9-dimensional color feature vector.

[0047] The parameters for constructing the gray-level co-occurrence matrix are: distance d = 1 pixel, and orientation θ = 0°, 45°, 90°, and 135° (averaged). The calculation formulas for each parameter are as follows: Energy value This reflects the uniformity of the texture; Contrast value, This reflects the clarity of the texture; Inverse moment value, This reflects the local uniformity of the texture; Correlation coefficient This reflects the directionality of the texture.

[0048] The edge curvature feature is calculated as follows: sampling points are taken at equal intervals along the boundary of the lesion area, and the curvature value of each sampling point is calculated. The distribution density of curvature extrema is calculated as the number of curvature extrema divided by the total boundary length. If the distribution density is greater than 0.15, it is classified as a multilocular confluent herpes structure; if the distribution density is less than or equal to 0.15, it is classified as a unilocular herpes structure. This distinction is of significant clinical importance in differentiating between vesicant toxins (such as mustard gas causing multilocular confluent vesicles) and irritant toxins (causing unilocular vesicles).

[0049] Specifically, the feature similarity coefficient is calculated by mapping the extracted morphological feature set to a high-dimensional feature space and calculating the cosine similarity with each standard feature vector in the toxicology knowledge base. If the cosine similarity is greater than the first similarity threshold, the corresponding poison will be added to the candidate poison list. If the cosine similarity is between the second similarity threshold and the first similarity threshold, the auxiliary discrimination network is activated to determine whether to add it to the candidate list based on the morphological evolution time sequence characteristics of the lesion area. The auxiliary discrimination network is a time-series prediction model based on a long short-term memory network.

[0050] Specifically, the formula for calculating cosine similarity is: In the formula, F is the extracted morphological feature vector, and Sk is the standard feature vector of the kth poison in the poison knowledge base.

[0051] The first similarity threshold (valued at 0.85) and the second similarity threshold (valued at 0.65) are set as follows: based on clinical verification records, the average of each similarity coefficient in records that meet the identification accuracy rate greater than 95% is recorded as the first similarity threshold; the average of each similarity coefficient in records that meet the identification accuracy rate between 70% and 85% is recorded as the second similarity threshold.

[0052] The LSTM-assisted discriminant network takes as input multiple frames of skin lesion image features acquired at consecutive time points (time intervals of 2 hours, 6 hours, 12 hours, and 24 hours) and outputs as morphological evolution probabilities. This network is used to handle atypical early manifestations of skin lesions and assists in differential diagnosis by monitoring the evolution trend of lesions over time (such as whether erythema transforms into vesicles, whether vesicles merge, etc.).

[0053] Specifically, the confidence score is determined by weighting the overall confidence score of each candidate toxicant based on the feature similarity coefficient of the candidate toxicant, the epidemiological prior probability in the toxicant knowledge base, and the contact area index of the lesion area. The output priority is sorted from high to low according to the overall confidence level; When the difference between the highest overall confidence level and the second highest overall confidence level is less than the preset difference threshold, a list of toxic substances is output, and clinical symptom information is marked as needed.

[0054] Specifically, the formula for calculating the overall confidence level is: In the formula, C is the overall confidence level, α, β, and γ are weighting coefficients with values ​​of α=0.65, β=0.20, and γ=0.15. P is the prior probability of the k-th toxin in epidemiological statistics (adjusted for region and season), and A is the contact area index = min(lesion area / body surface area, 0.8).

[0055] The preset difference threshold is set to 0.12. When the difference between the highest confidence level and the second highest confidence level is less than this threshold, it indicates that the model's ability to distinguish between the two candidate toxins is similar, and clinicians need to supplement medical history or symptom information for a final judgment. This mechanism effectively reduces the risk of misdiagnosis that may result from a single output.

[0056] Specifically, the severity of poisoning is classified into three levels: mild, moderate, and severe. The chromaticity offset is the angle difference between the skin color vector of the diseased area and the healthy skin area; The edema extent index is the percentage of the area of ​​the raised portion of the lesion region. The feature fusion weights are adaptively adjusted according to the type of toxic substance. The weight of the chromaticity shift corresponding to corrosive toxic substances is higher than the weight of the edema range index, and the weight of the herpes structure feature corresponding to blistering toxic substances is higher than the weight of the chromaticity shift.

[0057] Specifically, the chromaticity offset is the angle between the CIELAB color difference vector of the lesion area and the healthy skin area, which is determined according to the color difference evaluation method specified by the International Commission on Illumination. The edema extent index is based on the three-dimensional reconstruction of the segmented mask, calculating the ratio of the pixel area of ​​the protruding portion to the total area of ​​the lesion area.

[0058] The rules for determining the severity of poisoning are as follows: Mild: Color shift <15° and edema range index <0.2; Severe: Color deviation >45° or edema range index >0.6, or signs of tissue necrosis. Moderate: Between mild and severe.

[0059] The adaptive adjustment mechanism for feature fusion weights is as follows: when the candidate toxicant is a corrosive toxicant (acid or alkali), the weight of chromatic offset is set to 0.6, the weight of edema range index is set to 0.3, and the weight of herpes feature is set to 0.1; when the candidate toxicant is a vesicular toxicant, the weight of chromatic offset is set to 0.2, the weight of edema range index is set to 0.3, and the weight of herpes feature is set to 0.5; other toxicants use a balanced weight (0.33, 0.33, 0.34).

[0060] Specifically, the first aid recommendations include decontamination procedure guidelines and tips on the use of antagonists; The decontamination operation guidelines determine the type of rinsing solution based on the acid-base properties of the identified toxic substance. The type of rinsing solution is a neutral buffer solution, a weakly acidic solution, or a weakly alkaline solution. The antagonist usage prompt is generated based on the identified poison-specific antidote, and the route of administration and dosage reference range of the antidote are output simultaneously.

[0061] Examples of matching rules for specific antidotes: Organophosphorus pesticides are matched with atropine and pralidoxime, with intravenous injection as the recommended route of administration; hydrofluoric acid is matched with calcium gluconate gel, with topical application as the recommended route of administration; mustard gas is matched with sodium thiosulfate, with topical wet compress as the recommended route of administration. Dosage reference ranges are output differently for adults / children and according to the severity of poisoning.

[0062] Specifically, the structured report also includes image acquisition timestamps, GPS location information, and anonymized thumbnails of the lesion area; The structured report is synchronized to the pre-bound medical institution terminal via the HL7 / FHIR medical information exchange standard protocol; When the identification result corresponds to a highly toxic substance, an emergency warning is automatically triggered. The recipients of the emergency warning include the nearest emergency medical center and the registered contact terminal.

[0063] Specifically, the list of highly toxic substances includes, but is not limited to: paraquat, cyanide, organophosphates (highly toxic), hydrofluoric acid (high concentration), and VX nerve agent. Emergency alerts include: the type of poison, the severity of poisoning, the approximate location of the patient (GPS coordinates), suggested first aid measures, and estimated arrival time at the hospital. This alert mechanism achieves seamless information flow from the scene to the pre-hospital stage, significantly improving the efficiency of emergency response to poisoning incidents.

[0064] Please see Figure 3 As shown, the embodiment of the present invention provides a rapid skin contact toxicant identification system based on AI image recognition, comprising: The mobile acquisition module is configured to acquire images of skin lesion areas through a camera and perform quality assessment and preprocessing on the images. The preprocessing includes denoising, enhancement, and lesion area segmentation based on U-Net. The cloud-based AI analysis module is communicatively connected to the mobile acquisition module. The cloud-based AI analysis module contains a toxicology knowledge base and a multi-label classification model. The multi-label classification model is used to output a list of candidate toxicants and a comprehensive confidence level based on the received morphological feature set. The results output module, which is connected to the cloud AI analysis module, is configured to generate a structured report based on the comprehensive confidence level and match corresponding first aid recommendations according to the severity of poisoning. The data feedback module is connected to the mobile terminal acquisition module and the cloud AI analysis module, respectively. It is configured to collect the difference data between the actual diagnostic results and the model output, and trigger the incremental learning and update of the multi-label classification model based on the difference data.

[0065] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in the details for the sake of brevity.

[0066] This invention is intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A rapid identification method for skin contact toxins based on AI image recognition, characterized in that, include: Based on the images of skin lesion areas collected by the mobile device, image quality parameters are determined. When the image quality parameters meet the preset acquisition threshold, a preprocessing strategy is executed. The preprocessing strategy includes determining the kernel function size of the denoising algorithm based on the image noise level, determining the gain coefficient of image enhancement based on the color difference contrast between the lesion area and healthy skin, and determining the segmentation boundary of the lesion area based on the gradient distribution of skin texture. After preprocessing, a morphological feature set of the skin lesion area is extracted, which includes color space components, texture periodic parameters, and edge curvature features of the herpes structure. The morphological feature set is uploaded to the cloud AI model, and multi-label matching is performed based on the poison knowledge base. The multi-label matching includes: determining a list of candidate poisons based on feature similarity coefficients, and determining the output priority based on the confidence score of each candidate poison. Based on the matching results, a structured report is generated that includes the type of poison, the severity of poisoning, and emergency treatment recommendations. The severity of poisoning is determined based on the color shift of the lesion area, the edema range index, and the feature fusion weight.

2. The rapid identification method for skin contact toxins based on AI image recognition according to claim 1, characterized in that, The image quality parameters include illumination uniformity, focus sharpness, and motion blur coefficient; When the uniformity of illumination is lower than the preset illumination threshold, local histogram equalization is enabled for compensation. When the focus sharpness is lower than the preset sharpness threshold, the mobile device is triggered to refocus and resample. When the motion blur coefficient is greater than the preset blur threshold, the deconvolution restoration algorithm is executed, and the number of deconvolution iterations is determined based on the estimated length of the blur kernel.

3. The rapid identification method for skin contact toxins based on AI image recognition according to claim 2, characterized in that, The process of determining the segmentation boundaries of the lesion area includes: A lightweight segmentation network based on the U-Net architecture generates initial candidate regions based on the gradient change rate of skin texture. Based on the edge confidence map of the initial candidate regions, determine the seed points for the growth of connected components; The stopping criterion for region growth is determined based on the Euclidean distance of color difference between adjacent pixels and the texture difference. When the stopping criterion is triggered, a closed lesion region segmentation mask is output.

4. The rapid identification method for skin contact toxins based on AI image recognition according to claim 3, characterized in that, The color space components are jointly extracted from the RGB, HSV, and CIELAB color spaces; The texture periodicity parameters are determined based on the gray-level co-occurrence matrix, including energy value, contrast value, inverse difference moment value, and correlation coefficient; The edge curvature characteristics of the herpes structure are determined by calculating the curvature extreme values ​​at the boundary points of the lesion area, and are classified into unilocular herpes structures or multilocular fused herpes structures based on the distribution density of the curvature extreme values.

5. The rapid identification method for skin contact toxins based on AI image recognition according to claim 4, characterized in that, The feature similarity coefficient is calculated by mapping the extracted morphological feature set to a high-dimensional feature space and calculating the cosine similarity with each standard feature vector in the toxicology knowledge base. If the cosine similarity is greater than the first similarity threshold, the corresponding poison will be added to the candidate poison list. If the cosine similarity is between the second similarity threshold and the first similarity threshold, the auxiliary discrimination network is activated to determine whether to add it to the candidate list based on the morphological evolution time sequence characteristics of the lesion area. The auxiliary discrimination network is a time-series prediction model based on a long short-term memory network.

6. The rapid identification method for skin contact toxins based on AI image recognition according to claim 5, characterized in that, The confidence score is determined by weighting the overall confidence score of each candidate toxicant based on the feature similarity coefficient of the candidate toxicant, the epidemiological prior probability in the toxicant knowledge base, and the contact area index of the lesion area. The output priority is sorted from high to low according to the overall confidence level; When the difference between the highest overall confidence level and the second highest overall confidence level is less than the preset difference threshold, a list of toxic substances is output, and clinical symptom information is marked as needed.

7. The rapid identification method for skin contact toxins based on AI image recognition according to claim 6, characterized in that, The severity of poisoning is classified into three levels: mild, moderate, and severe. The chromaticity offset is the angle difference between the skin color vector of the diseased area and the healthy skin area; The edema extent index is the percentage of the area of ​​the raised portion of the lesion region. The feature fusion weights are adaptively adjusted according to the type of toxic substance. The weight of the chromaticity shift corresponding to corrosive toxic substances is higher than the weight of the edema range index, and the weight of the herpes structure feature corresponding to blistering toxic substances is higher than the weight of the chromaticity shift.

8. The rapid identification method for skin contact toxins based on AI image recognition according to claim 7, characterized in that, The first aid recommendations include decontamination procedure guidelines and tips on the use of antagonists; The decontamination operation guidelines determine the type of rinsing solution based on the acid-base properties of the identified toxic substance. The type of rinsing solution is a neutral buffer solution, a weakly acidic solution, or a weakly alkaline solution. The antagonist usage prompt is generated based on the identified poison-specific antidote, and the route of administration and dosage reference range of the antidote are output simultaneously.

9. The rapid identification method for skin contact toxins based on AI image recognition according to claim 8, characterized in that, The structured report also includes image acquisition timestamps, GPS location information, and anonymized thumbnails of the lesion area; The structured report is synchronized to the pre-bound medical institution terminal via the HL7 / FHIR medical information exchange standard protocol; When the identification result corresponds to a highly toxic substance, an emergency warning is automatically triggered. The recipients of the emergency warning include the nearest emergency medical center and the registered contact terminal.

10. A system for rapid identification of skin contact toxins based on AI image recognition as described in any one of claims 1 to 9, characterized in that, include: The mobile acquisition module is configured to acquire images of skin lesion areas through a camera and perform quality assessment and preprocessing on the images. The preprocessing includes denoising, enhancement, and lesion area segmentation based on U-Net. The cloud-based AI analysis module is communicatively connected to the mobile acquisition module. The cloud-based AI analysis module contains a toxicology knowledge base and a multi-label classification model. The multi-label classification model is used to output a list of candidate toxicants and a comprehensive confidence level based on the received morphological feature set. The results output module, which is connected to the cloud AI analysis module, is configured to generate a structured report based on the comprehensive confidence level and match corresponding first aid recommendations according to the severity of poisoning. The data feedback module is connected to the mobile terminal acquisition module and the cloud AI analysis module, respectively. It is configured to collect the difference data between the actual diagnostic results and the model output, and trigger the incremental learning and update of the multi-label classification model based on the difference data.

Images