Deep learning-based zebrafish embryo chemical substance toxicity evaluation system and method

Abstract

The invention discloses a zebra fish embryo chemical substance toxicity evaluation system and method based on deep learning. The system comprises a data enhancement module, a data processing module and a data processing module, wherein the data enhancement module is used for executing staged differential enhancement; the development stage identification module is used for outputting a development stage classification result based on a YOLOv5m model; the toxicity end-point morphology recognition module is used for segmenting the multi-embryo instance based on a Mask R-CNN model and outputting a toxicity end-point morphology classification result; the toxicity judgment module is used for calculating a development delay index DDI according to a deviation between an actually measured morphological parameter of a development stage classification result and a standard development stage morphological parameter; receiving a toxicity end-point morphology classification result, and calculating an accumulated score TS; and classifying the toxicity risk levels of the chemical substances through weighted fusion of DDI and TS. Compared with the prior art, the method has the advantages that zebrafish embryonic development can be accurately identified, scientific quantitative evaluation of chemical substance toxicity is realized, and reliable support is provided for high-flux compound toxicity screening and environmental toxicity prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer technology, and in particular to a zebrafish embryo chemical substance toxicity evaluation system and method based on deep learning. BACKGROUND

[0002] In the field of environmental toxicology, zebrafish embryos show unique advantages in evaluating the toxicity of chemical substances due to their transparent embryonic structure and rapid development cycle. After exposure to chemical substances, the toxicity endpoints of zebrafish embryos have certain stability and repeatability. By analyzing these toxicity endpoints, it is helpful to determine the biological toxicity of chemical substances and explore the potential impact mechanisms of new chemical substances, thereby reducing their toxicity to organisms and the environment.

[0003] With the development of deep learning technology, computer vision models have been applied to automatically identify zebrafish embryo development stages and malformation phenotypes, improving the efficiency of toxicity screening. However, there are still certain limitations in practical applications: on the one hand, it is difficult to accurately distinguish morphological changes during the transition period of development stages, leading to quantitative bias of development delay degree, and since a single microscopic image often contains multiple embryo targets, traditional techniques cannot efficiently realize the simultaneous analysis of multiple targets and multiple types of toxicity endpoints; on the other hand, existing technologies often only identify a single development stage or a few malformations without establishing a collaborative mechanism for comprehensive analysis; in addition, existing research lacks standardized large-scale zebrafish embryo development stage databases and toxicity endpoint morphology databases, making it difficult to compare data under different experimental conditions and limiting the universality of the model.

[0004] After searching, Chinese patent publication No. CN119007210A discloses a zebrafish phenotype identification and quantitative analysis method based on YOLOv8. This method constructs a zebrafish embryo image dataset and uses the YOLOv8 model to fuse feature pyramids and attention mechanisms to identify and segment the morphologies of sample images treated with different compounds at different concentrations, obtaining the body surface morphology data information of the samples. This invention improves the recognition accuracy and efficiency of the contours of various tissues of zebrafish, but does not involve the identification of development stages, cannot capture the morphological change characteristics of key development stages, and cannot quantify the development delay toxicity endpoints, making the toxicity evaluation dimension of chemical substances single.

[0005] Therefore, how to accurately identify zebrafish embryos after exposure to chemical substances while efficiently evaluating the toxicity of chemical substances has become a technical problem to be solved. SUMMARY

[0006] The purpose of the present application is to overcome the defects of the prior art and provide a zebrafish embryo chemical substance toxicity evaluation system and method based on deep learning.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] According to a first aspect of the present invention, a deep learning-based system for assessing the chemical toxicity of zebrafish embryos is provided, the system comprising:

[0009] Data augmentation module: used to perform phased and differentiated augmentation processing on zebrafish embryo data;

[0010] Developmental stage identification module: Based on the YOLOv5m model, the data output by the data augmentation module is processed to identify the developmental stage of zebrafish embryos and output the corresponding developmental stage classification results;

[0011] Toxicity endpoint morphology recognition module: Based on the Mask R-CNN model, the data output by the data augmentation module is processed to perform instance segmentation of multiple embryo targets in a single image and output the toxicity endpoint morphology classification result of each embryo independently;

[0012] Toxicity determination module: Receives the deviation between the measured morphological parameters and the standard morphological parameters of the developmental stage classification results of each zebrafish embryo output by the developmental stage identification module, and calculates the developmental delay index (DDI).

[0013] Simultaneously, the toxic endpoint morphology classification results of each zebrafish embryo output by the toxic endpoint morphology recognition module are received, and the cumulative score TS is calculated.

[0014] The chemical substance toxicity risk level is classified by weighted fusion strategy based on the Development Delay Index (DDI) and the cumulative score of toxic endpoint morphology (TS).

[0015] As a preferred technical solution, the data augmentation strategy includes:

[0016] Enhanced geometric transformations, including vertical flip, horizontal flip, rotation, scaling, and distortion;

[0017] Local pixel perturbations include pixel multiplication, adjacent pixel multiplication, Gaussian noise, Gaussian blur, and contrast adjustment;

[0018] Color space enhancement, including single-channel enhancement and HSV space enhancement;

[0019] Structural feature enhancement, including superpixel enhancement and sharpening.

[0020] As a preferred technical solution, the phased differentiation enhancement includes:

[0021] For unhatched embryos, priority is given to geometric transformation enhancement and structural feature enhancement;

[0022] The local pixel perturbation and color space enhancement are preferentially performed on the post-hatching embryo;

[0023] As a preferred technical solution, the development delay index DDI calculation method is:

[0024] Obtain the measured morphological parameters P of the embryo m ={p m1 ,p m2 ,…,p mn};

[0025] Obtain the standard development stage morphological parameters P s ={p s1 ,p s2 ,…,p sn};

[0026] Calculate the weighted deviation according to the parameter type:

[0027]

[0028] The morphological parameters include at least two of the number of somites, embryo length, head-tail angle, otic vesicle length, yolk extension and yolk sac ratio, and chest fin height / width ratio.w i The weight coefficient is set according to the importance of each parameter in the development delay evaluation, and the sum of the weight coefficients is 1.

[0029] As a preferred technical solution, the cumulative scoring rule of the toxicity endpoint morphology is that the total score is 5 points, and one point is deducted for each occurrence of a toxicity endpoint morphology.

[0030] As a preferred technical solution, the zebrafish embryo chemical substance toxicity evaluation system based on deep learning has the characteristics that the chemical substance toxicity risk grade division step specifically includes:

[0031] The development delay index DDI and the toxicity endpoint morphology cumulative score TS are normalized;

[0032] According to the development delay degree and the toxicity endpoint morphology type, the weight is dynamically allocated;

[0033] Calculate the comprehensive toxicity score: w DDI ×DDI′+w TS ×TS′;

[0034] According to the comprehensive score, the risk grade is divided:

[0035] The comprehensive toxicity score is less than 0.3, and the chemical substance is determined to be low risk;

[0036] 0.3≤The comprehensive toxicity score is less than 0.7, and the chemical substance is determined to be medium risk;

[0037] If the comprehensive toxicity score is greater than or equal to 0.7, the chemical substance is determined as high risk;

[0038] wherein w DDI is the weight of the development delay index DDI, w TS is the weight of the toxicity endpoint morphology cumulative score TS, DDI' is the normalized development delay index, and TS' is the normalized toxicity endpoint morphology cumulative score.

[0039] As a preferred technical solution, the training data of the development recognition module comprises:

[0040] A standard typical morphology data set of embryos at different development stages is established.

[0041] The zebrafish embryo micrographs in the data set are standardized and cropped.

[0042] The standard typical morphology data set of embryos at different development stages comprises 11,700 zebrafish embryo micrographs at 6hpf, 12hpf, 18hpf, 24hpf, 36hpf, normal hatching, and death.

[0043] As a preferred technical solution, the standardization and cropping adopts 220x220 pixel cropping for the un-hatched zebrafish embryo micrographs and 600x600 pixel cropping for the hatched zebrafish embryo micrographs.

[0044] The training data of the toxicity endpoint morphology recognition module comprises:

[0045] A zebrafish embryo toxicity endpoint morphology data set after exposure to different chemical substances is established.

[0046] As a preferred technical solution, the zebrafish embryo toxicity endpoint morphology data set is labeled to obtain the label types and contours of the zebrafish embryos.

[0047] The zebrafish embryo toxicity endpoint morphology data set comprises 8,730 zebrafish embryo micrographs of eight types of morphologies, i.e., normal hatching, spinal curvature, pericardial edema, post-hatching necrosis, tail downward bending, tail upward bending, yolk extension deformity, and yolk sac edema.

[0048] As a preferred technical solution, the Mask R-CNN model converts the multi-classification task into multiple binary classification tasks, and each zebrafish embryo toxicity endpoint type corresponds to a binary classifier.

[0049] According to a second aspect of the present application, a zebrafish embryo chemical substance toxicity evaluation method based on the system is provided, and the method comprises:

[0050] User uploads zebrafish embryo image to cloud platform;

[0051] Cloud performs picture preprocessing and data enhancement;

[0052] YOLOv5m model-based identification of zebrafish embryo development stage;

[0053] Mask R-CNN model-based identification of zebrafish embryo toxicity endpoint morphology;

[0054] Output development stage and toxicity endpoint morphology type;

[0055] Determine the chemical toxicity risk level and automatically interpret the analysis results.

[0056] Compared with the prior art, the present application has the following advantages:

[0057] 1. The present application integrates YOLOv5m and Mask R-CNN deep learning models to realize the automatic identification of zebrafish embryo development stage and toxicity endpoint morphology, and improves the accuracy of the identification results, breaking through the efficiency limitation of traditional manual identification.

[0058] 2. The present application divides the toxicity risk level by the development delay index and the toxicity endpoint cumulative score, and realizes the scientific quantitative evaluation of chemical toxicity.

[0059] 3. The present application can accurately identify subtle development differences and deformity characteristics by constructing a standardized large-scale data set combined with a phased differentiated data enhancement strategy, and improves the adaptability of the model to different shooting scenes and embryo shapes.

[0060] 4. The present application supports cloud deployment, adapts to different experimental environments and user needs, reduces the threshold of cross-laboratory data sharing and model use, and has good universality and generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 It is a zebrafish embryo chemical toxicity evaluation system architecture diagram of the present application;

[0062] Figure 2 It is a standardized data set established by the present application, wherein (a) is a typical morphology data set of embryos at different development stages, and (b) is a zebrafish embryo toxicity endpoint morphology data set after exposure to different chemicals;

[0063] Figure 3 It is a data enhancement effect diagram of the zebrafish embryo morphology image of the present application;

[0064] Figure 4 It is the training and prediction process of the YOLOv5m model of the development stage identification module of the present application;

[0065] Figure 5 YOLOv5m model target detection training results of the present application;

[0066] Figure 6 Loss convergence process of the training set and the validation set of the YOLOv5m model in the present application in the training iteration;

[0067] Figure 7 Average precision of each class of the YOLOv5m model in the present application when IoU is 0.5;

[0068] Figure 8 Precision-Recall curve of the YOLOv5m model in the present application when IoU is 0.5;

[0069] Figure 9 Training and prediction process of the Mask R-CNN model for toxic endpoint morphology recognition module in the present application;

[0070] Figure 10 Toxic endpoint morphology categories recognized by the Mask R-CNN model in the present application;

[0071] Figure 11 Training sample number and accuracy rate of the Mask R-CNN model in the present application after 120 iterations, wherein Figure 11 (a) is the training sample number, Figure 11 (b) is the accuracy rate;

[0072] Figure 12 Training sample number and accuracy rate of the Mask R-CNN model in the present application after 120 iterations after the sample is expanded 12 times, wherein Figure 12 (a) is the training sample number, Figure 12 (b) is the accuracy rate;

[0073] Figure 13 Training sample number and accuracy rate of the Mask R-CNN model in the present application after 240 iterations after the sample is expanded 12 times, wherein Figure 13 (a) is the training sample number, Figure 13 (b) is the accuracy rate;

[0074] Figure 14 Accuracy rate of each toxic endpoint morphology category recognized by the Mask R-CNN model in the present application;

[0075] Figure 15 Cross-laboratory zebrafish larva morphology recognition results of the present application. DETAILED DESCRIPTION

[0076] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts should fall within the protection scope of the present application.

[0077] Embodiment 1

[0078] The zebrafish embryo chemical substance toxicity evaluation system and method based on deep learning of the present application integrates YOLOv5m and Mask R-CNN models, automatically identifies the developmental stage and toxicity endpoint morphology of the embryo through phased differential data enhancement, and realizes accurate identification of the zebrafish embryo and effective evaluation of the toxicity risk of the chemical substance by combining the developmental delay index and the cumulative score division.

[0079] As shown in Figure 1 The zebrafish embryo chemical substance toxicity evaluation system of the present application specifically includes:

[0080] A data enhancement module for performing phased differential enhancement strategies on zebrafish embryo data;

[0081] A developmental stage identification module based on the YOLOv5m model for identifying the developmental stage of the zebrafish embryo and outputting the corresponding developmental stage classification result;

[0082] A toxicity endpoint morphology identification module based on the Mask R-CNN model for instance segmentation of multiple embryo targets in a single image and independent output of the toxicity endpoint morphology classification result of each embryo;

[0083] A toxicity determination module for receiving the deviation of the measured morphological parameters of the developmental stage classification result of each zebrafish embryo output by the developmental stage identification module from the standard developmental stage morphological parameters to calculate the developmental delay index DDI;

[0084] Receiving the toxicity endpoint morphology classification result of each zebrafish embryo output by the toxicity endpoint morphology identification module, calculating the cumulative score TS;

[0085] According to the developmental delay index DDI and the cumulative score TS of the toxicity endpoint morphology, the toxicity risk grade of the chemical substance is divided through a weighted fusion strategy.

[0086] The present application can realize efficient identification of zebrafish embryos and effective evaluation of chemical toxicity, by integrating the YOLOv5m development stage identification and the Mask R-CNN multi-target toxicity endpoint morphology identification double model architecture, combining the phased data enhancement and the collaborative quantitative toxicity grade, without relying on manual microscopic observation, and improving the precision and efficiency of chemical toxicity screening.

[0087] As Figures 2-15 shown, the steps of the present application are specifically described:

[0088] Step S1: Establish a standardized embryo development stage typical morphology data set, including 6hpf, 12hpf, 18hpf, 24hpf, 36hpf, normal hatching, death, a total of 11,700 zebrafish embryo micrographs, as shown in Figure 2 (a); referring to the representative characteristics and periods of zebrafish embryo development stages, and considering the individual development time difference and experimental operation to select the above time points;

[0089] Establish a zebrafish embryo toxicity endpoint morphology data set after exposure to different chemicals, including 8,730 zebrafish embryo micrographs of eight types of morphologies such as normal hatching, spinal curvature, pericardial edema, post-hatching necrosis, tail downward bending, tail upward bending, yolk extension deformity, and yolk sac edema, as shown in Figure 2 (b); the selected chemicals include bisphenol A, nonylphenol, lead iodide, methylbenzoic acid, perfluorooctane sulfonic acid and cadmium chloride, etc. emerging pollutants, which have different chemical properties and environmental behaviors, and can induce the above typical deformities, providing representative indicators for toxicity evaluation;

[0090] Step S2: Preprocess the original image data, first eliminate the low-quality photos without embryos, embryos straightly attached to the microplate hole wall and development state cannot be defined;

[0091] Through the batch processing function of Photoshop, the un-hatched zebrafish embryo micrographs are cropped by 220x220 pixels; the hatched zebrafish embryo micrographs are cropped by 600x600 pixels;

[0092] Label the 7 types of development stages by ImageJ; label the embryo contour and category label by LabelMe tool, requiring to mark the outer contour and not overlap, and multiple deformities exist in the same fish, which need to be labeled completely, and export as JSON format and convert to COCO format for storage;

[0093] Step S3: as Figure 3As shown, geometric transformation enhancement and structural feature enhancement are mainly applied to un-hatched embryos to highlight subtle differences in embryo morphological structure; local pixel perturbation and color space enhancement are mainly applied to post-hatching embryos to adapt to image feature changes caused by post-hatching embryo pigmentation, motion state, etc.

[0094] The data enhancement method is specifically:

[0095] (1) The calculation formula of vertical flip is:

[0096] [x,y,1]=[x0,H-y0,1],

[0097] Where x, y are the coordinates after flipping, x0, y0 are the coordinates before flipping, and H is the image height.

[0098] (2) The calculation formula of horizontal flip is:

[0099] [x,y,1]=[W-x0,y0,1],

[0100] Where x, y are the coordinates after flipping, x0, y0 are the coordinates before flipping, and W is the image width.

[0101] (3) The calculation formula of scaling is:

[0102]

[0103] Where x, y are the coordinates after flipping, x0, y0 are the coordinates before flipping, S x is the horizontal scaling ratio, and S y is the vertical scaling ratio.

[0104] (4) The calculation formula of rotation is:

[0105]

[0106] Where x, y are the coordinates after flipping, x0, y0 are the coordinates before flipping, and a is the rotation angle, x c is the horizontal coordinate of the rotation center, and y c is the vertical coordinate of the rotation center.

[0107] (5) The calculation formula of image blur is:

[0108]

[0109] Where f(i+k,j+l) represents the pixel value of the original image at coordinates (i+k,j+l); h(k,l) represents the weight value of the blur convolution kernel at (k,l); n and m are respectively the dimension parameters of the convolution kernel in the horizontal and vertical directions.

[0110] (6) The calculation formula of Gaussian blur is:

[0111]

[0112] where G(k, l) is a two-dimensional Gaussian function, x and y are coordinates relative to the center of the Gaussian kernel, and σ is the standard deviation of the Gaussian function;

[0113] (7) The formula for calculating contrast is:

[0114] C =∑ δ δ(i,j) 2 P δ (i,j),

[0115] where δ(i,j) = |i-j| is the gray level difference between adjacent pixels; P δ (i,j) is the probability of pixel distribution with a gray level difference δ between adjacent pixels;

[0116] (8) The warping operation simulates morphological distortion caused by developmental state differences, swimming states, or shooting angles by applying non-linear geometric deformation to the image;

[0117] (9) The pixel multiplication changes the features of the image by multiplying all pixels in the image by a random coefficient;

[0118] (10) The adjacent pixel multiplication changes the features of the image by multiplying different pixels in the image by different random coefficients, and the algorithm ensures that adjacent pixels do not use the same coefficient;

[0119] (11) The image sharpening algorithm changes the features of the image by convolving the entire image with the Laplacian operator;

[0120] (12) The superpixel enhancement algorithm generates a certain number of superpixels based on the average value of the pixels in the original image, and randomly replaces individual pixels in the original image with superpixels;

[0121] (13) The single-channel enhancement algorithm multiplies a random coefficient and adds a random bias value to a certain channel in RGB, thereby changing the color features of the original image;

[0122] (14) The color space enhancement algorithm converts the original image to the HSV color space, adds a random value to the entire image, and finally displays the entire image matrix in the RGB color space;

[0123] Step S4: As shown in Figures 4-8 the development stage recognition, the model receives the preprocessed image, uses the YOLOv5m model for training and prediction, and the specific process is as follows:

[0124] Training parameter setting: divide the training set and test set by 8:2, use the established 7-class development stage data set as input, use YOLOv5m framework, and realize the rapid identification of embryonic development stage by learning the global and local features of the image;

[0125] Training process: the model is iterated for a total of 150 times, and starts to converge after the 43rd iteration. When the 150th iteration is reached, the training set loss is about 0.45, the validation set loss is about 0.49, and the total loss reaches a low level;

[0126] Performance evaluation: the model performance is evaluated by calculating the average precision mAP with an intersection over union IoU of 0.5 and the precision-recall PR curve. The average precision of the 6hpf, 36hpf, and normal hatching classes is 0.9688, 0.9524, and 0.9538, respectively. The overall average precision is 0.92, and the PR curve maintains high performance at most recall rates;

[0127] Prediction output: the model can output the development stage of the embryo in real time and the corresponding confidence, and the average recognition rate is 30 images per second;

[0128] As shown in Figures 9-15 the toxicity endpoint morphology recognition, the model receives the preprocessed image, uses the Mask R-CNN model for training and prediction, and the specific process is as follows:

[0129] Training parameter setting: divide the training set and test set by 7:3, package as TFRecord format, use the established 8-class toxicity endpoint data set as input, use Mask R-CNN framework, and use its pixel-level mask generation capability to realize fine identification of malformation morphology;

[0130] Training process: convert the multi-classification task into multiple binary classification tasks, each malformation type has an independent binary classifier, the initial learning rate is set to 0.0003, the batch size is set to 2, and the iteration is 150 times. The final training set loss is about 0.71, and the validation set loss is about 0.75;

[0131] Performance evaluation: The model's recognition accuracy for all types is higher than 75%, with the recognition accuracy of the tail up being higher than 90%; after expanding the sample size by 12 times, dividing the number of positive samples and negative samples according to 1:3, dividing the test set and training set according to the ratio of 2:8, and training for 120 epochs, the accuracy of all types except the tail down is higher than 80%, and the recognition accuracy of the tail up and pericardial effusion is higher than 90%; after training for 240 epochs, the accuracy of all types except the tail down is higher than 85%, the recognition accuracy of the hatching death, tail up and pericardial effusion is higher than 90%, and the training accuracy of each phenotype is significantly improved; in the cross-laboratory verification, the average accuracy of all classes on the public dataset is 0.99;

[0132] Prediction output: The model outputs the toxicity endpoint morphology type of each embryo, the corresponding confidence, and the prediction box and pixel-level mask wrapping the embryo contour, where different colors represent different types, supporting multi-target and multi-malformation simultaneous recognition;

[0133] Step S5: Obtain the measured morphological parameters of the embryo and the standard parameters, and calculate the development delay index DDI:

[0134] Obtain the measured morphological parameters P of the embryo m ={p m1 ,p m2 ,…,p mn};

[0135] Obtain the standard development stage morphological parameters P s ={p s1 ,p s2 ,…,p sn};

[0136] Calculate the weighted deviation according to the parameter type:

[0137]

[0138] Wherein the morphological parameters include at least two of the number of somites, embryo length, head-tail angle, ear vesicle length, yolk extension and yolk sac ratio, and chest fin height / width ratio; w i is a weight coefficient, which is set according to the importance of each parameter in the development delay evaluation, and the sum of the weight coefficients is 1;

[0139] According to the initial score of 5 points, deduct 1 point for each occurrence of a toxicity endpoint morphology, and calculate the cumulative score TS of the toxicity endpoint morphology;

[0140] According to the development delay index DDI and the cumulative score of the toxicity endpoint morphology, divide the chemical toxicity risk level:

[0141] The development delay index DDI and the toxicity endpoint morphology cumulative score TS are normalized;

[0142] The weights are dynamically allocated according to the development delay degree and the toxicity endpoint morphology type;

[0143] The comprehensive toxicity score is calculated: w DDI DDI' + w Ts TS';

[0144] The risk level is divided according to the comprehensive score:

[0145] The comprehensive toxicity score is less than 0.3, and the chemical substance is determined to be low risk;

[0146] The comprehensive toxicity score is 0.3 to less than 0.7, and the chemical substance is determined to be medium risk;

[0147] The comprehensive toxicity score is greater than or equal to 0.7, and the chemical substance is determined to be high risk;

[0148] Wherein, w DDI is the weight of the development delay index DDI, w TS is the weight of the toxicity endpoint morphology cumulative score TS, DDI' is the normalized development delay index, and TS' is the normalized toxicity endpoint morphology cumulative score.

[0149] Example 2:

[0150] This embodiment is a method of the zebrafish embryo chemical substance toxicity evaluation system based on the present application, and the method comprises:

[0151] The user uploads the zebrafish embryo image to the cloud platform;

[0152] The cloud performs picture preprocessing and data enhancement;

[0153] The zebrafish embryo development stage is identified based on the YOLOv5m model;

[0154] The zebrafish embryo toxicity endpoint morphology is identified based on the Mask R-CNN model;

[0155] The development stage and the toxicity endpoint morphology type are output;

[0156] The toxicity risk level of the chemical substance is determined according to the development delay index and the toxicity endpoint morphology cumulative score, and the analysis result is automatically interpreted.

[0157] The present application provides real-time data processing and toxicity analysis services by deploying a cloud platform, integrating target detection algorithms based on YOLOv5m and Mask R-CNN, and providing a user interface with high interactivity, which provides convenience for cross-laboratory verification.

[0158] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements shall be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A deep learning-based chemical toxicity assessment system for zebrafish embryos, characterized in that, The system includes: Data augmentation module: used to perform phased differential augmentation on zebrafish embryo data; Developmental stage identification module: Based on the YOLOv5m model, the data output by the data augmentation module is processed to identify the developmental stage of zebrafish embryos and output the corresponding developmental stage classification results; Toxicity endpoint morphology recognition module: Based on the Mask R-CNN model, the data output by the data augmentation module is processed to perform instance segmentation of multiple embryo targets in a single image and output the toxicity endpoint morphology classification result of each embryo independently; Toxicity determination module: Receives the developmental stage classification results of each zebrafish embryo output by the developmental stage identification module, and calculates the developmental delay index (DDI) based on the deviation between the measured morphological parameters and the standard developmental stage morphological parameters. Simultaneously, the toxic endpoint morphology classification results of each zebrafish embryo output by the toxic endpoint morphology recognition module are received, and the cumulative score TS is calculated. The chemical substance toxicity risk level is classified by weighted fusion strategy based on the Development Delay Index (DDI) and the cumulative score of toxic endpoint morphology (TS).

2. The zebrafish embryo chemical toxicity assessment system based on deep learning according to claim 1, characterized in that, The data augmentation strategies include: Enhanced geometric transformations, including vertical flip, horizontal flip, rotation, scaling, and distortion; Local pixel perturbations include pixel multiplication, adjacent pixel multiplication, Gaussian noise, Gaussian blur, and contrast adjustment; Color space enhancement, including single-channel enhancement and HSV space enhancement; Structural feature enhancement, including superpixel enhancement and sharpening.

3. The zebrafish embryo chemical toxicity assessment system based on deep learning according to claim 1, characterized in that, The developmental delay index (DDI) is calculated as follows: Obtain the measured morphological parameters P of the embryo m ={p m1 ,p m2 ,…,p mn }; Obtaining standard developmental stage morphological parameters P s ={p s1 ,p s2 ,…,p sn }; Calculate weighted bias by parameter type: The morphological parameters mentioned include at least two of the following: number of somites, embryo length, head-to-trunk angle, earwax length, yolk extension to yolk globus ratio, and pectoral fin height / width ratio; w i The weighting coefficients are set according to the importance of each parameter in the assessment of developmental delay, and the sum of the weighting coefficients is 1.

4. The zebrafish embryo chemical toxicity assessment system based on deep learning according to claim 1, characterized in that, The cumulative scoring rule for the toxic endpoint morphology is as follows: the total score is 5 points, and one point is deducted for each occurrence of a toxic endpoint morphology.

5. The zebrafish embryo chemical toxicity assessment system based on deep learning according to claim 4, characterized in that, The steps for classifying the toxicity risk level of the chemical substances specifically include: The Developmental Delay Index (DDI) and the cumulative morphological score (TS) of the toxic endpoint were normalized. Weights are dynamically assigned based on the degree of developmental delay and the morphological type of toxic endpoints. Calculate the overall toxicity score: w DDI ×DDI′+w TS ×TS′; Risk levels are determined based on overall scores: A chemical substance with a comprehensive toxicity score of <0.3 is classified as low-risk. If the overall toxicity score is 0.3 or less and < 0.7, the chemical substance is classified as medium risk. A chemical substance with a comprehensive toxicity score ≥0.7 is classified as high-risk. Among them, w DDI For the weights of the Developmental Delay Index (DDI), w Ts TS represents the weight of the cumulative morphological score for the toxic endpoint, DDI′ is the normalized developmental delay index, and TS′ is the normalized cumulative morphological score for the toxic endpoint.

6. The zebrafish embryo chemical toxicity assessment system based on deep learning according to claim 1, characterized in that, The training data for the developmental recognition module includes: Establish a standardized dataset of typical morphological features of embryos at different developmental stages; The zebrafish embryo micrographs in the dataset were standardized and cropped. The standardized dataset of typical morphological features of embryos at different developmental stages includes 11,700 zebrafish embryo micrographs at seven developmental stages: 6hpf, 12hpf, 18hpf, 24hpf, 36hpf, normal hatching, and death.

7. The zebrafish embryo chemical toxicity assessment system based on deep learning according to claim 6, characterized in that, The standardized cropping method uses 220×220 pixels for micrographs of unhatched zebrafish embryos and 600×600 pixels for micrographs of hatched zebrafish embryos.

8. The zebrafish embryo chemical toxicity assessment system based on deep learning according to claim 1, characterized in that, The training data for the toxic endpoint morphology recognition module includes: Establish a dataset of zebrafish embryo toxicity endpoint morphology after exposure to different chemical substances; The zebrafish embryo toxicity endpoint morphology dataset was labeled to obtain the label types and outlines of zebrafish embryos; The zebrafish embryo toxic endpoint morphology dataset includes 8,730 zebrafish embryo micrographs representing eight morphological categories: normal hatching, spinal curvature, pericardial edema, post-hatching necrosis, tail curvature downward, tail upward, yolk extension malformation, and yolk sac edema.

9. The zebrafish embryo chemical toxicity assessment system based on deep learning according to claim 1, characterized in that, The Mask R-CNN model transforms a multi-classification task into multiple binary classification tasks, with each zebrafish embryo toxicity endpoint type corresponding to a binary classifier.

10. A method for assessing the chemical toxicity of zebrafish embryos based on any one of the systems described in claims 1-9, characterized in that, The method includes: Users upload images of zebrafish embryos to the cloud platform; Perform image preprocessing and data augmentation in the cloud; Identifying the embryonic development stages of zebrafish based on the YOLOv5m model; Identification of zebrafish embryotoxic endpoint morphology based on Mask R-CNN model; Output developmental stage and morphological type of toxic endpoint; Determine the toxicity risk level of chemical substances and automatically interpret the analysis results.

Images