Intelligent marine environmental pollution early warning method based on biomarkers

Abstract

The application discloses an intelligent marine environment pollution early warning method based on biomarkers and belongs to the technical field of marine ecological toxicology, which comprises the following steps: establishing an image quality control standard, acquiring a marine mussel histopathology image, constructing a data set according to the standard, then performing data labeling and quality control, and then performing pretreatment; a model is constructed based on a YOLOv11 architecture, a multi-task segmentation function is realized with the image segmentation module as a core, and model training, verification and evaluation are performed; different feature pixel points in the histopathology image are extracted through the trained model, the shortest distance between different feature pixel points is calculated by using the Euclidean distance, and an average result image containing distance labels is generated; and the marine environment pollution early warning grade is divided by calculating a response index H. The application precisely analyzes the histopathology image by means of deep learning, realizes the identification and segmentation of different tissue structures of marine mussels, and provides an efficient technical means for marine environment monitoring.

Description

Technical Field

[0001] This invention belongs to the field of marine ecotoxicology technology, specifically relating to an intelligent marine environmental pollution early warning method based on biomarkers. Background Technology

[0002] marine mussels ( Mytilus galloprovincialis As a typical filter-feeding economic shellfish, it plays an important ecological role in nearshore ecosystems. Through its continuous filtering and feeding, it can efficiently accumulate toxic substances such as heavy metals and persistent organic pollutants from seawater, thus becoming a key indicator species for marine ecotoxicology research. This species possesses a rich system of biomarkers, covering everything from molecular-level gene expression regulation and enzyme activity responses to cellular-level DNA damage, and even pathological changes at the tissue and organ level. Among these, tissue pathological damage, as a key link between molecular responses and organismal phenotypes, provides irreplaceable morphological evidence for assessing the biological effects of environmental pollution by visually demonstrating abnormalities in biological tissue structure.

[0003] Histopathology studies pathological changes in biological tissues and cells to assess an organism's adaptation to and response to its environment. Histopathological damage, as a sensitive detection indicator, can serve as an early warning system. When an organism exhibits pathological damage to its tissues and cells, although it may appear to survive at the individual level, it struggles to cope with environmental changes. Studying pathological changes to assess an organism's stress response to pollution can reveal the health status of both the environment and the organism. Common histopathological analysis methods include qualitative, semi-quantitative, and quantitative analyses. In the histopathological assessment technology system, quantitative detection indicators have become a research focus due to their objectivity and reproducibility. The thickness of digestive gland ductal cells and gill filament epithelium, as stable quantitative parameters, have been shown to be significantly correlated with pollutant exposure doses, providing important basis for constructing dose-response models. Traditional detection methods rely on manual microscopic observation combined with image processing software for morphological parameter measurement, requiring manual tracing of tissue boundaries and geometric transformation to calculate average thickness, or direct measurement methods. However, to meet statistical significance requirements, a single detection requires processing hundreds of pathological slide images, leading to a significant increase in workload.

[0004] More importantly, manual interpretation has significant technical limitations: on the one hand, differences in operator experience may lead to boundary recognition errors; on the other hand, repetitive measurements of complex organizational structures are prone to visual fatigue, resulting in data deviation.

[0005] Based on the intelligent analysis architecture of deep learning, combined with the image processing ability of CNN, the system realizes the end-to-end automatic analysis of mussel tissue structure, including feature extraction, target segmentation and parameter calculation. Not only solves the long-existing efficiency bottleneck in the field of marine mussel histopathology quantitative analysis, but also lays a technical foundation for building an intelligent marine ecological toxicology detection platform, which has important application value for realizing the rapid evaluation and early warning of marine pollution biological effects. SUMMARY

[0006] In view of the deficiencies in the background art, the purpose of the present application is to provide an intelligent marine environmental pollution early warning method based on biomarkers, which builds a data set through image acquisition, quality control, labeling and preprocessing, then uses an improved YOLOv11 architecture to build a multi-task segmentation network, then calculates the average thickness by Euclidean distance, and monitors the pollution by H index classification; realizes the automation of tissue analysis, improves the measurement accuracy and efficiency, provides technical support for marine ecological toxicology research, and helps to discover pollution and protect the marine environment in a timely manner.

[0007] The technical scheme adopted by the present application is as follows:

[0008] An intelligent marine environmental pollution early warning method based on biomarkers, comprising the following steps:

[0009] S1) Establish image quality control standards, acquire marine mussel histopathology images, build a data set according to the established image quality control standards, then data labeling and quality control, and then format conversion and division to preprocess the data for model training;

[0010] S2) Build a model based on YOLOv11 architecture, realize multi-task segmentation function with image segmentation module as the core, and perform model training, verification and evaluation to realize the recognition and segmentation of different structures in marine mussel histopathology images;

[0011] S3) Extract different feature pixel points in the histopathology image through the trained model, then calculate the shortest distance between different feature pixel points by Euclidean distance, and generate a result image containing distance labels after averaging;

[0012] S4) Divide the marine environmental pollution early warning level according to the value of response index H, and monitor the marine environmental pollution.

[0013] Preferably, in step S1), the image quality control standards are established, the marine mussel histopathology images are acquired, the data set is built according to the established image quality control standards, then the data is labeled and quality controlled, and then the data is preprocessed through format conversion and division, the specific process is as follows:

[0014] (1) Data preparation

[0015] Sa1) Establish image quality control standards, the standards are as follows: (1) Tissue morphological integrity: the tissue section is complete with no folding or damage; (2) Staining quality specification: uniform staining, clear contrast between cell nucleus and cytoplasm; (3) Optical imaging accuracy: accurate image focusing, no blurred areas;

[0016] Sa2) Image acquisition and dataset construction: collect Mytilus galloprovincialis gill and digestive gland tissue samples and perform quality screening, use staining method to make histopathological slides, use high-resolution digital pathology scanning system to obtain histopathological images that meet the standards, and construct a dataset that covers morphological variations of different individuals and tissues;

[0017] Sa3) Data annotation: use professional annotation tool Labelme to annotate target structures in histopathological images, including: for gill filament tissue, select the outer wall and inner wall of gill filament cells as features and annotate; for digestive gland tissue, select the inner wall and outer wall of lumen cells as features and annotate in layers; all annotations are saved in JSON format files;

[0018] Sa4) Data quality control: randomly select annotated histopathological images, calculate , the formula is , where A and B are the sets of annotation results of two researchers, represents the intersection of the two annotation results, and represent the absolute value of the total number of two annotation results; assuming the calculation result reaches 0.92±0.03, indicating that it meets the quality requirements;

[0019] (2) Data preprocessing

[0020] Sb1) Data format conversion: convert JSON format annotated information into YOLO format TXT file, each TXT file corresponds to one histopathological image, and histopathological images and corresponding TXT files are used as a dataset together;

[0021] Sb2) Dataset division: divide the dataset into training set and test set to ensure balanced data distribution.

[0022] Preferably, in step S2), the model is built based on YOLOv11 architecture, the image segmentation module is used as the core to realize multi-task segmentation function, and the model is trained, verified and evaluated, the specific steps are as follows:

[0023] Model Construction: The YOLOv11 architecture recognition and segmentation engine is adopted. This model extracts image features through its efficient backbone network and fuses them through multi-scale features in the Neck part. Finally, the segmentation head outputs binary segmentation masks for gill filament epithelial cells and digestive gland lumen tissue. The segmentation head is designed with a dual-branch structure: branch A is for gill filament cell segmentation and branch B is for digestive gland lumen cell layer segmentation. The role of YOLOv11 in this stage is to provide high-precision, pixel-level target contours for subsequent geometric parameter calculations.

[0024] Model Training: First, the prepared and preprocessed training and validation sets are input into the model. The parameters are set to train for 300 epochs with 16 samples per batch. The Stochastic Gradient Descent (SGD) optimizer is used to update the model by iteratively calculating the parameter gradients, as shown in the following formula:

[0025]

[0026] In the formula It is the first The model parameter vector after the next iteration. It's the learning rate. It is a loss function For parameter vectors gradient, It is the first The model parameter vector after the next iteration;

[0027] Then, Cosine annealing learning rate scheduling is used, where the maximum learning rate is set to 0.01 and the minimum learning rate is set to... ;

[0028] Finally, based on The metric is to stop training when the model shows no significant improvement after more than 100 training rounds.

[0029] Among them, adopt As a key indicator of model performance, it is calculated and obtained The average value within the interval is obtained The calculation formula is as follows:

[0030]

[0031] In the formula It is the average precision when the confidence threshold is t%. The value range is 50-95, which means that the average accuracy is calculated at each confidence level within the confidence threshold range of 50% to 95%.

[0032] Model verification and model evaluation: after training is completed, the trained model is verified on the validation set that does not participate in training, and the performance on the data is evaluated;

[0033] The calculation formula of the model accuracy is as follows:

[0034]

[0035] In the formula, the number of correctly predicted samples refers to the number of samples successfully predicted by the model on the test data set; the total number of samples is the total number of samples in the test data set;

[0036] The calculation formula of the recall rate is as follows:

[0037]

[0038] In the formula, the number of true positives is the number of actual positive samples successfully predicted by the model as positive; the number of false negatives is the number of samples that are actually positive but are incorrectly predicted by the model as negative.

[0039] Preferably, in step S3), different feature pixel points in the histopathology image are extracted by the trained model, that is, the outer wall pixel points and inner wall pixel points of the gill filament cells are extracted, the outer wall pixel points and inner wall pixel points of the glandular tissue lumen cells are extracted, and then the shortest distance between the outer wall pixel points and the inner wall pixel points, i.e. the thickness, is calculated by using the Euclidean distance, and the average thickness is calculated to generate a result image containing thickness labels. The specific steps are as follows:

[0040] Sc1) Digestive gland lumen cell and gill filament cell segmentation: using the above trained model to segment the digestive gland lumen cell and gill filament cell in the histopathology image, and extracting the pixel points constituting the outer wall and inner wall of the digestive gland lumen cell or gill filament cell, and using the pixel points to define the internal structure and external boundary of the cell, i.e. the outer wall pixel points and the inner wall pixel points;

[0041] Sc2) Euclidean distance calculation: Euclidean distance d is used for calculation. For the inner wall pixel point and the outer wall pixel point, the inner wall pixel point is assumed to be and the outer wall pixel point is The distance between each inner wall and outer wall is calculated, and the shortest distance between the inner wall pixel point and the outer wall pixel point is taken, and the calculation formula is:

[0042]

[0043] Where d is the straight-line distance between the inner wall pixel point and the outer wall pixel point, is the coordinate difference value of the inner wall pixel point and the outer wall pixel point in the horizontal direction, is the coordinate difference value of the inner wall pixel point and the outer wall pixel point in the vertical direction;

[0044] The minimum distance among the multiple Euclidean distances from the same inner wall pixel point to different outer wall pixel points is the thickness value, and the calculation formula is: ;

[0045] Sc3) Thickness average value calculation: average the thickness values of all positions, and the calculation formula is: , wherein N represents the number of thickness values of all positions, represents the thickness value of the i-th position, mark the result on the segmented cell, and record the number of recognized cells at the same time;

[0046] Sc4) Statistical result generation: generate a batch of labeled result images to show the quantitative results of cell thickness analysis.

[0047] Preferably, in the step S4), the response index H of the cell thickness is calculated, and the marine environmental pollution warning level is divided according to the value of the response index H to monitor the marine environmental pollution, and the specific steps are as follows:

[0048] The response index H of the cell thickness is calculated by comparing the cell thickness of the test sea area and the control sea area, and the calculation formula is as follows:

[0049]

[0050] , wherein T1 represents the thickness of the test sea area, and T0 represents the thickness of the control sea area;

[0051] According to the value of H, four warning levels are divided: H≤10% represents warning A level, 10%<H≤20% represents warning B level, 20%<H≤30% represents warning C level, and 30%<H represents warning D level.

[0052] Compared with the prior art, the present application proposes an intelligent marine environmental pollution warning method based on biomarkers, and the advantages of the method are:

[0053] 1. The present application establishes an intelligent identification and quantitative analysis method for marine mussel tissue pathology image based on deep learning through systematic experimental verification, breaks through the limitation of traditional manual experience, and provides a scientific and standardized technical path for tissue pathology analysis.

[0054] 2. The present application constructs a deep learning model suitable for marine mussel tissue pathology image analysis, realizes the automatic and accurate measurement of the thickness of the digestive gland lumen and the thickness of the gill filament epithelium relying on the improved YOLOv11 double-branch segmentation network, and avoids manual measurement error.

[0055] ​3. Deep learning model combined with traditional artificial detection method, verified the labeling consistency through Dice coefficient (0.92±0.03), combined with mAP 50-95 The index guarantees the performance of the model, and significantly improves the stability and reliability of the quantitative analysis of the marine mussel histopathology;

[0056] The intelligent identification method of histopathology image based on deep learning can greatly improve the analysis throughput and detection efficiency compared with artificial detection, and can quickly generate thickness labeled images and pollution early warning results, thereby providing efficient technical support for marine environment monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The flow framework diagram of the present application is shown in the figure;

[0058] Figure 2 The marine mussel histopathology slide preparation and image acquisition process diagram of the present application is shown in the figure; A, marine mussel tissue extraction and fixation; B, dehydration and transparency; C, paraffin embedding; D, tissue section; E, spreading and baking; F, hematoxylin and eosin staining; G, microscopic observation and image acquisition;

[0059] Figure 3 The cell thickness batch identification process diagram is shown in the figure; H, digestive gland histopathology image; I, using YOLOv11 to segment the digestive gland cells in the original image, P0 represents the outer wall, and P1 represents the inner wall; J, calculating the Euclidean distance between each outer wall and inner wall point, taking the inner wall point closest to each outer wall point as the cell thickness at the current outer wall position, and taking the average value as the average cell thickness; K, output result image, blue area represents cell area, label is cell category and current cell thickness;

[0060] Figure 4 The control group image of histopathological changes induced by sulfamethoxazole is shown in the figure. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be further described in detail below with reference to the drawings in the embodiments of the present application. It should be noted that the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0062] In order to make the purposes, technical solutions and advantages of the present application clearer, the embodiments of the present application will be further described in detail below with reference to the drawings of the specification. In order to more clearly understand the above purposes, features and advantages of the present application, the advantages of the present application will be further illustrated by comparing embodiments with reference to the drawings and specific embodiments.

[0063] The application proposes an intelligent marine environment pollution early warning method based on biomarkers, as shown in the accompanying Figure 1 The steps of the method are described in detail as follows:

[0064] S1) Establish image quality control standards, obtain marine mussel histopathology images, construct a data set according to the established image quality control standards, then data annotation and quality control, and then format conversion and division to preprocess the data, and prepare for model training;

[0065] Specifically, in the step S1), the image quality control standards are established, the marine mussel histopathology images are obtained, the data set is constructed according to the established image quality control standards, then the data is annotated and quality controlled, and then the data is preprocessed through format conversion and division, and the specific process is as follows:

[0066] (1) Data preparation

[0067] Sa1) Establish image quality control standards, the standards are as follows: (1) integrity of tissue morphology:

[0068] The tissue section is complete without folding or damage; (2) standardization of staining quality: uniform staining, clear contrast between cell nucleus and cytoplasm, and ensure that the cell morphological characteristics can be accurately identified; (3) optical imaging accuracy: accurate image focusing without blurred areas; to ensure that the image quality meets the requirements of deep learning model training;

[0069] Sa2) Image acquisition and construction of data set: collect the gill and digestive gland tissue samples of marine mussels and perform quality screening, use H&E staining method to make histopathology slides, and use high-resolution digital pathology scanning system to obtain histopathology images (including tissue microscopic morphology, cell distribution and staining characteristics), as shown in the accompanying Figure 2 Through standardized staining and digital scanning, the transformation of marine mussel pathological samples from "physical section" to "digital asset" is realized, providing a standardized and repeatedly called basic data carrier for subsequent use;

[0070] Based on the above standards, 200 samples are quality screened, 1000 histopathology images meeting the standards are obtained, and a data set is constructed, which covers morphological variations of different individuals and tissues;

[0071] Sa3) Data annotation: completed by two marine biological pathologists with more than 5 years of experience; professional annotation tool Labelme was used to accurately annotate the target structure in the histopathology image, including: for gill filament tissue, the outer wall and inner wall of gill filament cells were selected as features and annotated; for digestive gland tissue, the inner wall and outer wall of the lumen cells were selected as features and annotated in layers; all annotations were saved in JSON format files, including polygon vertex coordinates and class labels;

[0072] Sa4) Data quality control: 100 labeled histopathology images were randomly selected, and , the formula is , wherein A and B are the sets of annotation results of the two researchers, represents the intersection of the two annotation results, and represent the absolute values of the total number of two annotation results respectively; assuming the calculation result reaches 0.92±0.03 (mean±standard deviation), indicating that it meets the quality requirements of histopathology images;

[0073] (2) Data preprocessing

[0074] Sb1) Data format conversion: convert the polygon vertex information annotated in JSON format into TXT files in YOLO format, each TXT file corresponds to a histopathology image, the file contains the number of target categories and normalized bounding box coordinates, the histopathology image and the corresponding TXT file are used as a data set together;

[0075] Sb2) Data set division: divide the data set into training set (80%) and test set (20%), ensure the balance of data distribution.

[0076] S2) Based on the YOLOv11 architecture, the model is constructed, the image segmentation module is used as the core to realize the multi-task segmentation function, the model training, verification and evaluation are carried out, and the recognition and segmentation of different structures in the marine mussel histopathology image are realized;

[0077] Specifically, in the step S2), based on the YOLOv11 architecture, the model is constructed, the image segmentation module is used as the core to realize the multi-task segmentation function, the model training, verification and evaluation are carried out, and the specific steps are as follows:

[0078] Model construction: YOLOv11 architecture recognition and segmentation engine is adopted; the model uses its efficient Backbone network as the feature extractor of the model, gradually extracts features from the input image, and finally outputs the binary segmentation mask of the siphon epithelial cells and the digestive gland lumen tissue through the multi-scale feature fusion of the Neck part and the segmentation head, wherein the segmentation head part is designed as a double-branch structure: branch A siphon cell segmentation and branch B digestive gland lumen cell hierarchical segmentation, this double-branch design enables the model to simultaneously process two different but related segmentation tasks, breaking the limitations of single-task segmentation and improving the analysis capability of complex tissue images (such as marine mussel histopathology images); YOLOv11 plays a role in this stage to provide high-precision, pixel-level target outlines for subsequent geometric parameter calculation;

[0079] Model training: first, input the training set and validation set after data preparation and preprocessing into the model; the parameter setting is to train for 300 rounds, and the number of samples per batch is 16; the number of rounds determines the number of learning times of the model on the data set, and the batch affects the memory utilization and parameter update frequency; the stochastic gradient descent (SGD) optimizer is used to update the model by iterative calculation of parameter gradients, as follows:

[0080]

[0081] wherein is the model parameter vector after the i-th iteration, is the learning rate, is the loss function , the gradient of the parameter vector , the model parameter vector after the i-th iteration ;

[0082] Then, the Cosine annealing learning rate scheduling is adopted, which is to make the learning rate change periodically according to the cosine function, so that the learning rate converges faster in the early stage and the parameter is adjusted more finely in the later stage, avoiding overfitting or slow convergence, wherein the maximum value of the learning rate is set to 0.01 and the minimum value of the learning rate is set to ;

[0083] Finally, based on the validation set index, when the model does not improve significantly after more than 100 rounds of training in the training process, the training is stopped; this strategy can prevent the model from overfitting and save computing resources, thereby realizing the recognition and segmentation of the features of the digestive gland lumen cells and the siphon cells in the marine mussel histopathology images;

[0084] wherein, ​​The mean Average Precision between 50% to 95% (mAP The average value in the interval is obtained , and the calculation formula is as follows:

[0085]

[0086] In the formula, AP(t) is the average precision when the confidence threshold is t%; The value range of mAP is 50-95, which represents the average precision at each confidence level within the confidence threshold range from 50% to 95%;

[0087] Model verification and model evaluation: after training is completed, the trained model is verified on the validation set that did not participate in training to evaluate its performance on the data; by analyzing the prediction results of the model on the validation set, the accuracy of the recognition and segmentation of digestive gland lumen cells and gill filament cells is evaluated; this step is crucial for confirming the generalization ability and accuracy of the model, and provides a foundation for further optimization and application;

[0088] The accuracy of the trained model reached 88.8%, the recall rate reached 93.1%, and the mAP 50 (B) was 0.869, and the mAP 50-95 (B) was 0.63, which further confirmed the excellent performance of the model in the pixel-level segmentation task;

[0089] The calculation formula of the model accuracy is as follows:

[0090]

[0091] In the formula, the number of correctly predicted samples refers to the number of samples successfully predicted by the model on the test data set; the total number of samples is the total number of samples in the test data set;

[0092] The calculation formula of the recall rate is as follows:

[0093]

[0094] In the formula, the number of true positives is the number of actual positive samples successfully predicted by the model as positive; the number of false negatives is the number of samples that are actually positive but are incorrectly predicted by the model as negative.

[0095] S3) Extract different feature pixels in the histopathology image through the trained model, and then calculate the shortest distance between different feature pixels using the Euclidean distance to generate a result image containing distance labels after averaging; ​

[0096] Specifically, in step S3), different feature pixels in the histopathological image are extracted using a trained model, namely, the outer and inner wall pixels of gill filament cells, and the inner and outer wall pixels of digestive gland lumen cells. Then, the shortest distance between the outer and inner wall pixels, i.e., the thickness, is calculated using Euclidean distance. The average thickness is then used to generate a result image with thickness annotations, as shown in the attached figure. Figure 3 As shown, the specific steps are as follows:

[0097] Sc1) Segmentation of digestive gland lumen cells and gill filament cells: The trained model above is used to accurately segment the digestive gland lumen and gill filament cells in the histopathological image. At the same time, the pixels that constitute the outer and inner walls of the digestive gland lumen or gill filament cells are extracted. The internal structure and external boundary of the cell are defined by the pixels, namely the outer wall pixels and the inner wall pixels.

[0098] Sc2) Euclidean distance measurement calculation: Based on the obtained inner wall pixels and outer wall pixels, calculate the distance between each inner wall pixel and outer wall pixel, and take the shortest distance between the inner wall pixels and outer wall pixels, which is the cell thickness at the location of the inner wall pixels and outer wall pixels.

[0099] The calculation uses Euclidean distance d. For inner wall pixels and outer wall pixels, it is assumed that the inner wall pixel is... and outer wall pixels Calculate the distance between each inner and outer wall, and take the shortest distance between the outer wall pixels and the inner wall pixels. The calculation formula is:

[0100]

[0101] Where d is the straight-line distance between the outer wall pixels and the inner wall pixels. This represents the difference in horizontal coordinates between the outer and inner wall pixels. This represents the vertical coordinate difference between the outer wall pixels and the inner wall pixels.

[0102] The thickness value is calculated by selecting the minimum Euclidean distance from the same inner wall pixel to different outer wall pixels. The formula is as follows: ;

[0103] Sc3) Calculation of average thickness: The thickness values ​​at all locations are averaged using the following formula: Where N represents the number of thickness values ​​at all locations, Indicates the first The thickness value at each location is used to label the results on the segmented cells, and the number of cells identified is recorded at the same time.

[0104] S4) Statistical result generation: generate the result image of batch annotation, and show the quantitative result of cell thickness analysis.

[0105] S4) By calculating the response index H, the marine environmental pollution early warning level is divided according to the value, and the marine environmental pollution is monitored.

[0106] Specifically, in step S4), by calculating the response index H of the cell thickness, the marine environmental pollution early warning level is divided according to the value, and the marine environmental pollution is monitored, and the specific steps are as follows:

[0107] By comparing the thickness of the marine mussel tissue in the test sea area (which may be polluted) and the control sea area (which is assumed to be non-polluted and stable environment), the tissue thickness response index H is calculated, and the calculation formula is as follows:

[0108]

[0109] Wherein, T1 represents the tissue thickness of the test sea area, and T0 represents the tissue thickness of the control sea area;

[0110] According to the value of H, it is divided into four early warning levels: H≤10% represents early warning A level, 10%<H≤20% represents early warning B level, 20%<H≤30% represents early warning C level, and 30%<H represents early warning D level;

[0111] The pollution degree of the related sea area is characterized by calculating the tissue thickness response index. Under normal circumstances, the tissue thickness of the marine mussel remains in a relatively stable range, and thickening (such as tissue swelling, inflammatory reaction, etc.) and thinning (such as tissue necrosis, abnormal nutrition, etc.) all suggest that the marine mussel may be under stress of pollutants; if the tissue thickness response index of the related sea area is greater than 10%, it proves that there is a stress source in the sea area, and environmental protection should be strengthened.

[0112] The specific experiments of the present application are described in detail in combination with the technical solutions and the drawings as follows:

[0113] (1) Comparative Example 1 (recognition accuracy)

[0114] ① Use the histopathological images of marine mussels accumulated by the research team, randomly select 20 histopathological images of digestive gland and gill tissue from them to form a test data set;

[0115] ② The test data set is annotated by using the method, manual recognition (unfamiliar) and manual recognition (familiar) respectively, and the accuracy of the three recognition methods is judged by experienced researchers;

[0116] ③Result analysis: From Table 1, it can be seen that the recognition effect of the present method and artificial recognition (proficient) is better, which is significantly higher than artificial recognition (unfamiliar); the gill filament structure is relatively simple, and the gill filament recognition effect is as follows: artificial recognition (proficient) > the present method > artificial recognition (unfamiliar). The digestive gland lumen structure is relatively complex, and the digestive gland lumen recognition effect is as follows: the present method > artificial recognition (proficient) > artificial recognition (unfamiliar). It can be seen that the present method reduces the weight of experience in the recognition of key tissues of marine mussels, and is more conducive to the promotion of the quantitative detection method of marine mussel histopathology.

[0117] (2) Comparative Example 2 (measurement time)

[0118] ① The test data set construction method is the same as step ① in Comparative Example 1, only the digestive gland data set is constructed;

[0119] ② The thicknesses of the digestive gland lumen and gill filament epithelial cells in the test data set are measured by the present method, artificial recognition (geometric conversion method) and artificial recognition (direct measurement method) respectively, and the measurement time is compared;

[0120] ③ Main steps of artificial recognition (geometric conversion method): first, outline the digestive gland lumen profile, measure the relevant parameters using the image processing software ImageJ, and then convert the image, and then calculate the thickness of the digestive gland lumen using the formula;

[0121] ④ Main steps of artificial recognition (direct measurement method): randomly select five positions on each digestive gland lumen, measure them using the image processing software ImageJ, and take the average value as the average thickness of the digestive gland lumen;

[0122] ⑤ Result analysis: From Table 2, the interpretation times of the three methods are as follows: the present method (5 min) < direct measurement method (45 min) < geometric conversion method (55 min). The present method greatly saves the interpretation time, improves the detection throughput and efficiency.

[0123] (3) Comparative Example 3 (effect of sulfamethoxazole stress on the thicknesses of the digestive gland lumen and gill filament of marine mussels)

[0124] ① Laboratory acclimation of marine mussels: collect marine mussels and place them in laboratory acclimation, and the specific culture conditions are as follows: salinity 32.10±0.21‰, dissolved oxygen 82.71±3.36%, pH 8.12±0.22, temperature 18.30±0.27℃, light cycle 12 h darkness: 12 h light. Replace the artificial seawater every day, and feed the feed algae (isochrysis galbana), and after the marine mussels are stable, start the stress experiment;

[0125] ② Sulfamethoxazole stress experiment setting: The stress experiment was set up with two experimental conditions, namely, the control group (only adding artificial seawater) and the sulfamethoxazole stress group (50 μg / L); the stress experiment lasted for 6 days, and the artificial seawater, feed and sulfamethoxazole working solution were added every day to ensure the stability of the stress system; the sulfamethoxazole stress concentration was selected based on the environmental related concentration reported in the literature to simulate the pollution in the actual sea area;

[0126] ③ Tissue sampling: After the stress experiment, 6 mussels were randomly collected from each treatment group as biological repeats, and the standard H&E staining method was used to prepare the histopathological slides, and the high-resolution digital pathology scanning system was used to obtain the histopathological images; the slide samples were observed under an ordinary optical microscope (400x magnification), and the random sampling strategy was used to collect the histopathological images, ensuring that not less than 50 high-quality microscopic images were obtained for each sample;

[0127] ④ The deep learning image analysis method and the direct measurement method (same as step ④ of Comparative Experiment 2) were used to quantitatively analyze the marine mussel histopathology images, and the tissue thickness response index was calculated;

[0128] ⑤ The results show that after sulfamethoxazole stress, the digestive gland and gill of marine mussels show obvious histopathological response, the digestive gland lumen shows thickness increase, and the gill epithelium shows thickness decrease, i.e. the quantitative index of marine mussel histopathology can be used to represent the biological effect of marine pollution; as shown in the accompanying Figure 4 The results of the digestive gland lumen thickness and gill epithelium thickness obtained by the method and the traditional method are consistent, and there is no significant difference, indicating that the analysis results of the method are accurate, stable and greatly improve the detection efficiency; the gill tissue thickness response index H = 22.04% (warning C level), and the digestive gland tissue thickness response index H = 14.78% (warning B level), indicating that the environmental pollutants in the test sea area have caused obvious histopathological damage to marine mussels, and protective measures should be taken.

[0129] In summary, the present application provides a deep learning-based intelligent quantitative analysis method for marine mussel histopathology, which significantly improves the analysis throughput and detection efficiency through automatic image recognition and precise measurement technology; the present application provides an efficient, stable and reliable implementation scheme for the biological effect evaluation of marine environmental pollution, and provides important technical support for marine ecotoxicology research and environmental monitoring field.

[0130] Table 1 Comparison of recognition effects of different methods

[0131]

[0132] Table 2 Comparison of recognition time of different methods

[0133]

[0134] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the preferred embodiments by those of skill in the art once they have the benefit of the present disclosure. Therefore, the appended claims are intended to encompass within their scope all possible variations and modifications of the preferred embodiments.

[0135] It is apparent that those skilled in the art can, without departing from the spirit and scope of the application, make various changes and modifications of the application. Thus, the present application is intended to encompass all such changes and modifications within the scope of the application claimed.​

Claims

1. A biomarker-based intelligent marine environment pollution early warning method, characterized in that, Comprise: S1) Establish image quality control standards, obtain marine mussel histopathology images, construct a data set according to the established image quality control standards, then data annotation and quality control, and then format conversion and division to preprocess the data, and prepare for model training; S2) Based on the YOLOv11 architecture, the model is constructed, the image segmentation module is used as the core to realize the multi-task segmentation function, the model is trained, verified and evaluated, and the recognition and segmentation of different structures in the marine mussel histopathology images are realized; S3) Through the trained model, different feature pixel points in the histopathology image are extracted, and the shortest distance between different feature pixel points is calculated by using the Euclidean distance, and the average is calculated to generate a result image containing distance annotation; Through the trained model, different feature pixel points in the histopathology image are extracted, that is, the outer wall pixel points and inner wall pixel points of the gill filament cells are extracted, the outer wall pixel points and inner wall pixel points of the lumen cells of the digestive gland tissue are extracted, and then the shortest distance between the outer wall pixel points and the inner wall pixel points is calculated by using the Euclidean distance, that is, the thickness, and the thickness is averaged to generate a result image containing thickness annotation, the specific steps are as follows: Sc1) Digestive gland lumen cell and gill filament cell segmentation: using the above trained model to segment the digestive gland lumen cell and gill filament cell in the histopathology image, and at the same time extracting the pixel points constituting the outer wall and inner wall of the digestive gland lumen cell or gill filament cell, using the pixel points to define the internal structure and external boundary of the cell, that is, the outer wall pixel points and the inner wall pixel points; Sc2) Euclidean distance calculation: Euclidean distance d is used for calculation, for inner wall pixel points and outer wall pixel points, assuming that the inner wall pixel point is and the outer wall pixel point is , the distance between each inner wall and outer wall is calculated, and the shortest distance between the inner wall pixel point and the outer wall pixel point is taken, and the calculation formula is: ； wherein d is the straight-line distance between the inner-wall pixel point and the outer-wall pixel point, is the coordinate difference in the horizontal direction between the inner-wall pixel point and the outer-wall pixel point, is the coordinate difference in the vertical direction between the inner-wall pixel point and the outer-wall pixel point. The minimum distance of the multiple Euclidean distances from the same inner wall pixel point to different outer wall pixel points is the thickness value, and the calculation formula is: ; Sc3) Thickness average calculation: average the thickness values of all positions, the calculation formula is: average thickness = ∑ (thickness value of position i) / N where N represents the number of thickness values of all positions, represents the thickness value of the i-th position, mark the result on the segmented cell, and record the number of recognized cells at the same time; ​ Sc4) Statistical result generation: generate a batch of annotated result images to show the quantitative results of cell thickness analysis; S4) By calculating the response index H, the marine environmental pollution early warning level is divided according to the value, and the marine environmental pollution situation is monitored. 2.The biomarker-based intelligent marine environment pollution early warning method according to claim 1, characterized in that, In the step S1), the image quality control standards are established, the marine mussel histopathology images are obtained, the data set is constructed according to the established image quality control standards, then the data is annotated and quality controlled, and then the data is preprocessed by format conversion and division, the specific process is as follows: (1) Data preparation Sa1) Establish image quality control standards, the standards are as follows: (1) Tissue morphology integrity: the tissue section is complete, without folding or damage; (2) Staining quality specification: uniform staining, clear contrast between cell nucleus and cytoplasm; (3) Optical imaging accuracy: accurate image focusing, no blurred area; Sa2) Image acquisition and data set construction: collect gill and digestive gland tissue samples of marine mussels and perform quality screening, use staining method to make histopathology slides, use high-resolution digital pathology scanning system to obtain histopathology images meeting the standards, and construct a data set covering morphological variations of different individuals and tissues; Sa3) Data annotation: use professional annotation tool Labelme to annotate the target structures in the histopathology images, including: for gill filament tissue, select the outer wall and inner wall of gill filament cells as features and annotate; for digestive gland tissue, select the inner wall and outer wall of lumen cells as features and annotate in layers; all annotations are saved in JSON format file; Sa4) Data quality control: randomly selected labeled histopathology images, calculated , the formula is , wherein, is the set of labeling results of two researchers respectively, represents the intersection of two labeling results, and respectively represent the absolute value of the total number of two labeling results; it is assumed The calculation result of reaches 0.92±0.03, indicating that the quality requirement is met; (2) Data preprocessing Sb1) data format conversion: convert the information labeled in JSON format into a TXT file in YOLO format, each TXT file corresponds to one histopathology image, and the histopathology image and the corresponding TXT file are taken together as a data set; Sb2) data set division: divide the data set into a training set and a test set to ensure the balance of data distribution. 3.The biomarker-based intelligent marine environment pollution early warning method according to claim 1, characterized in that, In step S2), a model is constructed based on the YOLOv11 architecture, a multi-task segmentation function is realized with the image segmentation module as the core, and model training, verification and evaluation are performed. The specific steps are as follows: Model construction: YOLOv11 architecture recognition and segmentation engine is adopted; the model extracts image features through its efficient Backbone network, and finally outputs binary segmentation masks of gill epithelial cells and digestive gland lumen tissues through multi-scale feature fusion of the Neck part and segmentation head, wherein the segmentation head part is designed as a double-branch structure: branch A gill cell segmentation and branch B digestive gland lumen cell hierarchical segmentation; the role of YOLOv11 at this stage is to provide high-precision, pixel-level target contours for subsequent geometric parameter calculation; Model training: first, input the training set and validation set after data preparation and preprocessing into the model, set the parameters to train for 300 rounds, the number of samples per batch is 16, use the stochastic gradient descent (SGD) optimizer, update the model by iterative calculation of parameter gradients, and the formula is as follows: ； In the formula It is the first The model parameter vector after the next iteration. It's the learning rate. It is a loss function For parameter vectors gradient, It is the first The model parameter vector after the next iteration; Then, the Cosine annealing learning rate schedule is adopted, wherein the maximum value of the learning rate is set to 0.01, and the minimum value of the learning rate is set to 1 e-5 ; Finally, based on mAP 50-95 an indicator that training is stopped when the model does not improve significantly after more than 100 rounds of training during the training process; wherein the use of mAP 50-95 As a key indicator of model performance, the mean absolute error (MAE) is calculated as the average of the absolute differences between the predicted and actual values t The mean value over the interval is obtained as mAP 50-95 The formula is as follows: ； In the formula AP(t) is the average precision when the confidence threshold is t%; t The value range of is 50-95, which represents the average precision at each confidence value calculated in the range of confidence threshold from 50% to 95%. Model verification and model evaluation: after training is completed, the trained model is verified on the validation set that did not participate in training to evaluate its performance on the data; The calculation formula of the model accuracy is as follows: ； In the formula, the number of samples predicted correctly refers to the number of samples successfully predicted by the model on the test data set; the total number of samples refers to the total number of samples in the test data set; The calculation formula of the recall rate is as follows: ； In the formula, the number of true positives is the number of actual positive samples successfully predicted as positive by the model; the number of false negatives is the number of samples that are actually positive but are incorrectly predicted as negative by the model. 4.The biomarker-based intelligent marine environment pollution early warning method according to claim 1, characterized in that, In step S4), the response index H of cell thickness is calculated, and the marine environmental pollution warning level is divided according to the value, and the marine environmental pollution is monitored. The specific steps are as follows: By comparing the cell thickness of the test sea area and the control sea area, the response index H of cell thickness is calculated, and the formula is as follows: ； Wherein, T1 represents the thickness of the test sea area, and T0 represents the thickness of the control sea area; According to the value of H, it is divided into 4 warning levels: H≤10% represents warning level A, 10%<H≤20% represents warning level B, 20%<H≤30% represents warning level C, and 30%<H represents warning level D.

Images