AI-assisted high-throughput screening method for high-risk cardiotoxicity chemicals

Abstract

The invention is applicable to the technical field of chemical toxicity testing and computer image processing, and provides an AI-assisted high-throughput screening method for high-risk cardiotoxicity chemicals, which comprises the following steps: acquiring a Tg (myl7: eGFP) transgenic zebrafish embryo heart fluorescence image, video and average fluorescence intensity data subjected to chemical treatment; s2, performing instance segmentation of morphological recognition on the atrium and ventricular regions in the heart fluorescence image obtained in the step S1 by using a YOLOv8 model to obtain morphological information parameters of the atrium and ventricle, and quantitatively outputting major axis and minor axis of the atrium and ventricle and ventricular contraction and diastole area data in combination with OpenCV morphological operation; calculating a cardiac function evaluation index based on the data obtained in the step S1 and the step S2; and obtaining evaluation and screening results. The method has the advantages of high recognition speed, high accuracy, multiple evaluation indexes and low time and cost, the efficiency and accuracy of chemical cardiotoxicity evaluation can be improved, and the method can be used for early risk assessment and health early warning of chemicals.

Description

Technical Field

[0001] The present invention belongs to the technical field of chemical toxicity testing and computer image processing, and in particular to an AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity. Background Art

[0002] Cardiovascular disease is a type of chronic disease that poses a serious threat to human health and is the leading cause of death worldwide. A large number of studies have confirmed that chemical exposure can cause damage to the cardiovascular system and is one of the main causes of cardiovascular disease. Therefore, screening for high-risk chemicals with cardiotoxicity is of great significance.

[0003] Zebrafish is a widely used high-throughput model organism that shares 70% of protein-coding genes with humans. Its cardiac development process, cardiac function and heart disease characteristics are largely consistent with those of humans. Transgenic zebrafish whose cardiomyocytes express green fluorescent protein (myl7:eGFP) are widely used in the study of the cardiac toxicity of pollutants because their hearts emit fluorescence. Combined with a high-content cell imaging analysis system, the cardiac development phenotype can be observed in real time, clearly and dynamically. It can also achieve high-throughput screening of the cardiac toxicity of pollutants.

[0004] High-content cell imaging analysis systems can capture zebrafish heart images and heartbeat videos at high throughput, thereby generating a large amount of image information that can reflect the cardiotoxic effects of chemicals. However, manual processing of heart images and videos is subjective and has low throughput.

[0005] The prior art "Environ.Sci.Technol.2013,47,11302-11310" developed a method for high-content screening of chemicals that affect cardiovascular function based on Tg (fli1:eGFP) transgenic zebrafish. However, this method only focuses on the heart rate, circulation, body length, pericardial area, and intersegmental vessels of zebrafish, and can only automatically output results based on images for intersegmental vessels and body length, while other indicators are semi-automated quantitative analysis; the patent applications with publication numbers CN112418171A and CN111583207A identify the position and contour of the zebrafish heart, but the functions are relatively limited; the prior art "Comput.Biol.Med.2021,135,104565" uses a microscope to collect heart beating videos of wild zebrafish and zebrafish with cardiomyopathy mutations under bright field, and then uses the U-net deep learning model to segment the ventricles to obtain the long and short axes of the ventricles. and its contraction and diastolic areas, and calculate cardiac function indicators (ejection fraction (EF), ventricular short-axis contraction rate (FS), heart rate (HR), and stroke volume (SV)); the patent application with publication number CN111685740A uses a high-resolution microscope camera to shoot zebrafish heart movement video in bright field, obtains multiple frames of heart images, predicts the ventricular edge contour and the curvature of the pericardium through a convolutional neural network algorithm, obtains the long axis and short axis of the ventricle, and further calculates 8 cardiac function evaluation indicators (heart rate (HR), cardiac output (CO), stroke volume (SV), ejection fraction (EF), ventricular area (MA), arrhythmia degree (AF), ventricular diastolic contraction ratio (D:S), and pericardial edema degree (PEF)). The above technical solutions are difficult to achieve high-throughput testing of chemical cardiac toxicity and high-throughput acquisition of heart images. Moreover, in many cases, the ventricles and atria of zebrafish overlap, making it difficult to accurately identify the ventricles in bright field, resulting in large errors.

[0006] The patent application with publication number CN116385468A discloses a system based on the generation of zebrafish heart evaluation index image analysis software. The technical solution processes the input zebrafish heartbeat video image frame into a ventricular mask probability map, then calculates the ventricular mask area, and further processes the ventricular mask area frame by frame to obtain a one-dimensional heart rate signal, and finally outputs the heart rate value. However, the cardiac function evaluation indicators that can be obtained are very limited; the prior art "Adv. Sci. 2023, 10 (30), 2301136" uses cardiac fluorescent labeled transgenic zebrafish image data to train and develop a computer neural network ZVSegNet for automatic segmentation of zebrafish ventricles, and obtains multiple cardiac function evaluation indicators (ventricular end-diastolic area (EDA), ventricular end-systolic area (ESA), heart rate (HR), stroke volume (SV), ventricular short-axis contraction rate (FS), area change fraction (FAC), pericardial edema rate, long axis, short axis), and constructs HRNet through zebrafish heart beat video to realize quantitative analysis of heart rate.

[0007] Most of the above technical contents (except the prior art "Environ. Sci. Technol. 2013, 47, 11302-11310") mainly disclose the analysis and processing of zebrafish heart images and the acquisition of multiple evaluation indicators of cardiac function, but do not cover the development of high-throughput testing methods for chemical cardiotoxicity. Moreover, these studies cannot simultaneously segment the atria (At) and ventricles (Ve), resulting in limited accuracy of ventricular segmentation, and cannot perform RR interval, QT interval, QT C Quantitative analysis of the interval.

[0008] Therefore, there is an urgent need to propose a high-throughput screening method for the cardiotoxicity of chemicals based on zebrafish, which integrates high-throughput toxicity testing of chemical exposure, high-throughput acquisition of zebrafish heart images and average fluorescence intensity data under a fluorescence field, and automated measurement and analysis of more cardiac phenotypes and functional multi-evaluation indicators. It can be used for early risk assessment and health warning of chemicals, and also provide a basis for the development of environmentally friendly chemicals. Summary of the invention

[0009] The purpose of the embodiments of the present invention is to provide an AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity, aiming to solve the problems raised in the above-mentioned background technology.

[0010] The embodiment of the present invention is implemented as follows: an AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity comprises the following steps:

[0011] S1. Obtain fluorescent images, videos and average fluorescence intensity data of the heart of Tg(myl7:eGFP) transgenic zebrafish embryos treated with chemicals;

[0012] S2. Use the YOLOv8 model to perform instance segmentation of the atrium and ventricle regions in the cardiac fluorescence image obtained in step S1 to obtain the morphological information parameters of the atrium and ventricle, and combine the OpenCV morphological operation to quantitatively output the long axis and short axis of the atrium and ventricle, and the diastolic and systolic area data of the ventricle;

[0013] S3, calculating a cardiac function evaluation index based on the data obtained in step S1 and step S2;

[0014] S4. Obtain evaluation and screening results.

[0015] Preferably, in step S1, a high-content imaging analysis system is used to obtain fluorescent images and average fluorescence intensity data of the heart of Tg(myl7:eGFP) transgenic zebrafish embryos treated with chemicals, and a video is produced to establish an image dataset of the heart of zebrafish embryos exposed to chemicals.

[0016] Preferably, step S1 specifically includes the following process:

[0017] 1) Tg(myl7:eGFP) cardiac transgenic zebrafish were used for mating at a 1:1 ratio of male to female, and the fertilized eggs were treated with chemicals within 5 hours of fertilization and cultured in a constant temperature incubator at 28.5°C with a 14h light / 10h dark cycle;

[0018] 2) The dye solution was renewed every 24 h, and 210 μL of 5 mM N-phenylthiourea (PTU) was added 24 h after fertilization to inhibit pigment production, and cultured until 72 h after fertilization;

[0019] 3) Then transfer to a black transparent bottom 96-well plate, anesthetize the zebrafish larvae with tricaine (MS-222), fix the body position of the zebrafish larvae in the culture plate and centrifuge at 200 rpm for 2 minutes, adjust the ventral side to the same direction, and close the eyes;

[0020] 4) Use a high-content imaging analysis system to acquire images, perform layer scanning, select the FITC fluorescence field 488nm channel, use confocal mode under a 4x objective to acquire images, obtain average fluorescence intensity data, and make a video. Save the first 20 frames of the video as pictures to establish an image data set for the sample.

[0021] Preferably, step 4) specifically includes the following process:

[0022] 4.1) On-line: Run the high-content imaging analysis system, open the MetaXpress software, and load the black transparent-bottom 96-well plate containing chemically treated zebrafish embryos / larvae;

[0023] 4.2) Set the imaging program parameters: select 4X Plan Apo objective lens, the number of fields of view for each well is 2×2, and each field of view is close to each other; select laser autofocus and Z layer vertical shooting in the acquisition setting, laser autofocus is selected to focus on the bottom of the plate, and then offset according to the thickness of the bottom plate, select FITC fluorescence field 488nm channel, adjust the parameters to achieve an exposure time of 40ms; laser power 80%; focus adjustment mode Z-offset from W1 offset, turn on digital confocal, and shoot the selected wells;

[0024] 4.3) Make a movie: Make a heart beating video of the transgenic zebrafish. Select StreamAcquisition, set the shooting frame number to 250 frames, and use the FITC fluorescence channel for shooting. Select the Make movie option and select 100 frames to make a 15s video.

[0025] 4.4) The 250 frames of images taken by Stream Acquisition were calculated using Region Measurements to calculate the heart region and obtain the Average Intensity images and data within 15 seconds. The first 20 frames of each heart video were cut into images, and finally all the split heart images were integrated as an image dataset. The Labelme open source annotation tool was used to annotate the frame-by-frame image dataset to obtain the tissue contours of the atria and ventricles of individual zebrafish larvae.

[0026] Preferably, step S2 specifically includes the following process:

[0027] 1) Use the imgaug library to perform data augmentation on the image dataset through the image distortion algorithm: perform data augmentation on all data through different image augmentation algorithms according to random thresholds and random combinations to expand the scale and diversity of the training dataset, convert the annotated JSON file into a TXT file to generate the YOLO dataset format, and divide the image dataset into a training set, a validation set, and a test set in a ratio of 8:1:1;

[0028] 2) The position information of the atria and ventricles in the amplified data set is respectively passed into the established YOLOv8 instance segmentation model to obtain two models for segmenting the atria and ventricles. The YOLOv8 instance segmentation network model includes a Backbone network structure, a Neck network structure and a Head network structure, which are respectively used for feature extraction, feature fusion and prediction; wherein the Backbone network structure adopts CSPDarknet53, and extracts different levels of information of the image through multiple convolutional layers, residual blocks and C2f modules. The C2f module divides the feature map into two parts and performs convolution operations respectively, and then fuses them to reduce the amount of calculation and improve the feature expression ability; the Neck network structure includes a feature pyramid network (FPN) and a feature attention enhancement network (FAN). FPN fuses feature maps of different scales through top-down and lateral connections, and FAN further enhances the expression ability of the feature map, combines global and local information, and ensures that the target detection and segmentation tasks can handle targets of different scales; the Head network structure includes a detection head (Detection Head). The detection head and the segmentation head jointly implement the instance segmentation task. The detection head adopts the anchor-free method to generate the bounding box and category prediction of the target by predicting whether each feature point is the target center point and its offset, bounding box and category. The segmentation head starts from the feature map of Neck and gradually restores it to the same resolution as the input image through a series of convolution and upsampling operations, and finally generates a pixel-level segmentation mask for each detection box. The segmentation head accurately aligns the feature map with the detection box by using RoIAlign technology to avoid pixel-level deviation and improve the accuracy of the mask. The model uses CSPDarkNet53 as the backbone network, uses FPN to process feature maps of different scales, uses FAN to further enhance feature fusion, and further upsamples the feature map to the original image resolution. After the convolution layer, the features are gradually refined, thereby realizing the instance segmentation of zebrafish heart tissue in the image.

[0029] 3) Use the trained model to perform instance segmentation of the atrial and ventricular regions in the zebrafish heart images treated with chemicals to obtain the morphological information parameters of the atrial and ventricular regions: Training is based on the Pytorch deep learning framework, using NVIDIA's GeForce RTX 4060Ti GPU is used for calculation and CUDA12.1 library is used for acceleration. YOLOv8 instance segmentation models are trained for identifying and segmenting the atrium (At) and ventricle (Ve) regions respectively to avoid the problem of unsatisfactory segmentation results caused by overlapping of the two regions when a single model is used for segmentation. The trained model is evaluated to obtain the model's accuracy indicators, including the model's segmentation precision, recall rate and mAP50 score. The image is input into the trained model for prediction, the instance segmentation mask image predicted by the model is read, and the circumscribed circle and minimum circumscribed rectangle are made for the obtained mask image in combination with OpenCV. The diameter of the circumscribed circle is defined as the long axis of the atrium or ventricle, and the width of the minimum circumscribed rectangle is defined as the short axis. Finally, the long axis and short axis and area data of the atrium and ventricle in diastole and contraction are automatically output to an Excel file.

[0030] Preferably, in step S3, the cardiac function assessment indicators include heart rate (HR), ventricular end-diastolic volume (EDV), ventricular end-systolic volume (ESV), ventricular end-diastolic area (EDA), ventricular end-systolic area (ESA), stroke volume (SV), ventricular short-axis contraction rate (FS), ejection fraction (EF), fractional area change (FAC), cardiac output (CO), RR interval (the distance between adjacent peaks in the heart rate graph), QT interval (the distance between adjacent troughs in the heart rate graph) and QTc interval (heart rate-corrected QT interval), totaling 13 indicators.

[0031] Preferably, step S3 specifically includes the following steps: using the openpyxl library to compare the heart rate mean fluorescence intensity data in the heart rate Excel file, respectively calculating the heart rate (the number of peaks per unit time), calculating the RR interval (the distance between adjacent peaks in the heart rate graph), the QT interval (the distance between adjacent troughs in the heart rate graph) and the QTc interval (heart rate corrected QT interval) indicators; other indicators are calculated based on heart rate, major axis, minor axis and area.

[0032] An embodiment of the present invention provides an AI-assisted high-throughput screening method for chemicals with high-risk cardiotoxicity. A high-content imaging analysis system and a zebrafish model organism are used to conduct high-throughput testing of chemical cardiotoxicity, and AI-assisted evaluation of zebrafish heart images is used to achieve automatic output of chemical cardiac phenotype and functional evaluation indicators, thereby achieving high-throughput evaluation of chemical cardiotoxicity. The method established in the embodiment of the present invention integrates high-throughput toxicity testing of chemical exposure, high-throughput acquisition of zebrafish heart images and average fluorescence intensity data under a fluorescence field, and automated measurement and analysis of multiple evaluation indicators of cardiac phenotype and function. It has the advantages of fast recognition speed, high accuracy, covering many evaluation indicators, and low time and cost. It can greatly improve the efficiency and accuracy of chemical cardiotoxicity evaluation, and can also be used for early risk assessment and health warning of chemicals. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flow chart of an AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity provided in an embodiment of the present invention;

[0034] Figure 2 The original image of the implementation data and the enhanced image after enhancement provided by the embodiment of the present invention;

[0035] Figure 3 A schematic diagram of the YOLOv8 segmentation network structure model used in an embodiment of the present invention;

[0036] Figure 4 The result of predicting the fluorescence image of zebrafish larvae after the model training provided in the embodiment of the present invention;

[0037] Figure 5 The model provided in the embodiment of the present invention predicts the systolic and diastolic states of zebrafish heart and compares them. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0039] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0040] like Figure 1 FIG. 1 is a flow chart of an AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity provided by an embodiment of the present invention, which specifically includes the following steps:

[0041] S1. Obtaining fluorescent images of zebrafish embryos / larvae hearts treated with environmental pollutants:

[0042] 1) Chemical treatment and culture of transgenic zebrafish embryos:

[0043] Tg(myl7:eGFP) cardiac transgenic zebrafish embryos were exposed to chemicals within 5 hours of fertilization and then cultured; they were cultured in a constant temperature and light incubator (28.5°C, light / dark: 14 / 10h) until 72hpf, with water changed every 24 hours, and 210μL 5mmol / L PTU was added from 24hpf to inhibit pigment production;

[0044] 2) Pre-imaging processing:

[0045] The fixed position was to use MS-222 to anesthetize the zebrafish larvae in the 96-well plate, centrifuge at 200 rpm for 2 min, and adjust the ventral side to face the same direction with both eyes overlapped;

[0046] 3) Image data acquisition:

[0047] 3.1) On-line: Run the high-content imaging analysis system (ImageXpress Micro Confocal), open the MetaXpress software, and load the black transparent-bottom 96-well plate containing chemically treated zebrafish embryos / larvae;

[0048] 3.2) Set imaging program parameters:

[0049] Select 4X Plan Apo objective lens, shoot 2×2 fields of view per well, and make sure that each field of view is adjacent to the other. Select laser autofocus and vertical shooting of Z layer for acquisition settings. Select laser autofocus to focus on the bottom of the plate and then offset according to the thickness of the bottom plate. Select FITC fluorescence field 488nm channel, turn on digital confocal, and shoot the selected wells.

[0050] 3.3) Make a movie: Make a heart beating video of the transgenic zebrafish. Select StreamAcquisition, set the shooting frame number to 250 frames, and use the FITC fluorescence channel for shooting. Select the Make movie option and select 100 frames to make a 15s video.

[0051] 3.4) Use Region Measurements to calculate the heart region using the 250 frames of images taken by Stream Acquisition to obtain Average Intensity images and data within 15 seconds;

[0052] S2. Establish instance segmentation model of atrium and ventricle regions:

[0053] 1) Cut the first 20 frames of each video into pictures and integrate them as an image dataset;

[0054] 2) Preprocessing and enhancement of image datasets: The image datasets were preprocessed using the open source annotation tool Labelme to annotate the original image datasets and obtain the contours of the atria and ventricles of zebrafish larvae;

[0055] The embodiment of the present invention solves the problem of single environmental factors such as background elements and lighting conditions of artificially photographed images in the same batch and insufficient training data by means of data enhancement, thereby improving the robustness of the model for zebrafish instance segmentation and measurement. The data enhancement effect of the image dataset performed by the imgaug development library is as follows: Figure 2 As shown in the figure, 13 image processing algorithms including salt and pepper noise, Gaussian noise, Gaussian blur, rotation, translation, scaling, horizontal flip, vertical flip, gamma contrast adjustment, linear contrast adjustment, brightness transformation, hue and saturation transformation, and snowflake effect are applied. The three enhanced images randomly use some of the above 13 algorithms:

[0056] The addition ratio of salt and pepper noise is set to 0.01;

[0057] The noise intensity of Gaussian noise is set to 5;

[0058] The calculation formula for image blur is:

[0059] The calculation formula of Gaussian blur is: The standard deviation of blur level is set between 0 and 3;

[0060] The calculation formula for rotation is:

[0061]

[0062] Where x, y are the coordinates after rotation, x0, y0 are the coordinates before rotation, α is the rotation angle, x c is the horizontal coordinate of the rotation center, y c is the ordinate of the rotation center;

[0063] Random translation, percentage range is -20% to 20%;

[0064] The calculation formula for scaling is:

[0065] Where x, y are the coordinates after scaling, x0, y0 are the coordinates before scaling, S x is the horizontal scaling ratio, S y is the vertical scaling ratio;

[0066] The calculation formula for horizontal flipping is: [x, y, 1] = [W-x0, y0, 1],

[0067] The calculation formula for vertical flipping is: [x, y, 1] = [x0, H-y0, 1],

[0068] Where x, y are the coordinates after flipping, x0, y0 are the coordinates before flipping, H is the height of the image, and W is the width of the image;

[0069] Gamma contrast adjustment changes the brightness distribution of an image through nonlinear transformation. For the input pixel value I and the output pixel value O, the formula for gamma correction is:

[0070] O(x,y)=I γ (x,y),

[0071] Where I is the pixel value normalized to [0,1];

[0072] Linear contrast adjustment adjusts the contrast of an image through linear transformation, making the distribution range of image pixel values ​​wider or more compact, thereby enhancing or weakening the contrast. The calculation formula is:

[0073] O(x,y)=α(I(x,y)-μ)+μ,

[0074] Where I is the input pixel value, O is the output pixel value, α is the contrast factor, and μ is the average value of the image pixel value;

[0075] Brightness transformation is achieved by performing linear addition operation on each pixel value of the image. The calculation formula is:

[0076] O(x,y)=I(x,y)+Δ,

[0077] Where I is the input pixel value, O is the output pixel value, and Δ is the brightness adjustment value;

[0078] Hue and saturation transformations are achieved by converting images from RGB color space to HSV color space. In HSV color space, hue transformation changes the overall color of the image by adjusting the value of the hue channel, while saturation transformation changes the purity or intensity of the color by adjusting the value of the saturation channel. After adjustment, the image is converted back to RGB color space. Hue transformation is usually achieved through addition and modulus operations, while saturation transformation is usually achieved through multiplication. These transformations can flexibly adjust the color properties of the image without affecting its brightness and details.

[0079] The snowflake effect simulates the shape of snowflakes by randomly generating white dots, adjusting the size and number, blending them with the original image, and then applying motion blur to simulate the effect of snowflakes falling;

[0080] 3) Creation of YOLO file data: Classify the enhanced image dataset, 80% of which is used as training data, 10% as verification data, and 10% as test data. After the image is annotated using Labelme, the corresponding JSON label file is converted into a TXT text file to construct a dataset in the format required by the YOLO model.

[0081] 4) Build a YOLOv8 target instance segmentation network model and input the image dataset images and annotation information into the network for training:

[0082] The model loss functions are divided into four categories: 1) bounding box regression loss 2) confidence loss 3) classification loss 4) mask loss;

[0083] The bounding box regression loss is used to measure the difference between the predicted bounding box and the true bounding box, which is responsible for allowing the model to locate the target more accurately:

[0084] Bounding box regression loss function:

[0085] IoU is the intersection over union ratio between the predicted box and the real box, p(b,b g ) is the Euclidean distance between the center points of the predicted box and the real box, c is the diagonal length of the minimum closure area surrounding the two, α is the weight balancing factor, and v is a parameter that measures the consistency of the aspect ratio of the predicted box and the real box;

[0086] Confidence loss is used to measure the accuracy of the model's confidence prediction of the detected target, and is responsible for improving the accuracy of the model's judgment on whether the target exists:

[0087] Confidence loss function: L conf =-(ylog(p)+(1-y)log(1-p)),

[0088] y is the true label (1 if the target exists, 0 if the target does not exist), and p is the confidence of the model prediction;

[0089] The classification loss is used to measure the model's prediction accuracy for the detected target category and is responsible for improving the accuracy of target classification:

[0090] Classification loss function:

[0091] N is the number of categories, y i is the true category label, p i is the class probability predicted by the model;

[0092] The mask loss function includes Binary Cross-Entropy Loss and Dice Loss, which are used to measure the difference between the mask predicted by the model and the true mask, so that the model can accurately segment each pixel of the target and improve the accuracy of the instance segmentation mask:

[0093] Binary Cross-Entropy Loss:

[0094] Dice Loss:

[0095] Combining the above parts, the total loss function of YOLOv8 for instance segmentation can be expressed as:

[0096] L=λ CIoU L CIoU +λ conf L conf +λ cls L cls +λ mask (L mask +L Dice ),

[0097] Among them, λ CIoU , conf , cls , mask is the weight coefficient that balances the losses of each part. By optimizing these loss functions, the model can perform target detection and instance segmentation more accurately, thus performing well in the segmentation task of zebrafish larvae’s heart representation.

[0098] The model architecture of YOLOv8 for instance segmentation combines the requirements of target detection and instance segmentation. It realizes efficient feature extraction and prediction through the design of multi-layer networks. The backbone network (CSPDarknet) extracts low-level and high-level features of the image to obtain rich feature maps. This part mainly includes convolutional layers for preliminary feature extraction, and residual blocks for processing feature maps at different levels to capture complex patterns in the image.

[0099] After the backbone network, FPN further fuses features and fuses feature maps from different scales to better handle objects of different sizes. It includes top-down path enhancement and bottom-up path enhancement, which propagate information from high-level features to low-level features and from low-level features to high-level features, respectively, to strengthen multi-scale feature fusion.

[0100] The detection head and segmentation head are the core prediction parts of the model. The detection head is mainly composed of convolutional layers, which are used to further process feature maps and generate outputs related to target detection, including classification prediction, bounding box regression, and confidence prediction. The main tasks of these layers are to predict the target category at each anchor point, the position of the target bounding box, and the confidence of the target existence.

[0101] The segmentation head is responsible for instance segmentation. It restores the feature map size to the size of the original input image through the upsampling layer, and finally generates an instance segmentation mask to predict the category of each pixel.

[0102] The schematic diagram of the YOLOv8_seg network structure model is as follows Figure 3 As shown;

[0103] The training is based on the Pytorch deep learning framework, using NVIDIA's graphics processor GeForce RTX 4060Ti8G for calculations and CUDA 12.1 library for acceleration;

[0104] 5) Evaluate the overall indicators of the model: Evaluate the model obtained after training to obtain the accuracy, recall rate and mAP50 score of the model for heart classification and segmentation. The results are shown in Table 1:

[0105] Table 1 Overall average indicators of model training results

[0106]

[0107]

[0108] 6) Use the trained model to identify the fluorescent image of zebrafish larvae and obtain the instance segmentation mask image of the atrium and ventricle in the image:

[0109] After obtaining the segmentation mask images of the atria and ventricles, the areas of the atria and ventricles of the zebrafish larvae can be calculated by counting the number of pixels in each mask image and according to the scale provided by the image acquisition device, that is, the actual size represented by each pixel;

[0110] The long and short axes of the atria and ventricles are calculated through the segmentation masks of the atria and ventricles. The long axis is calculated by making a circumscribed circle for the mask and obtaining the diameter of the circumscribed circle, which is used as the long axis for subsequent calculations. The short axis is calculated by making a minimum circumscribed rectangle for the mask and the width of the rectangle is used as the length of the short axis. Then, the data results of the same batch are output one by one into the Excel file;

[0111] The trained model is used to predict the fluorescence images of zebrafish larvae. The results are as follows: Figure 4 As shown;

[0112] The trained model is used to predict the zebrafish heart systolic and diastolic state images. The results and comparisons are as follows: Figure 5 As shown;

[0113] S3. Calculation of cardiac function assessment indicators:

[0114] 1) Heart rate (HR): Read the measured heart rate mean fluorescence intensity data, find the peaks and troughs of the heart rate by screening the maximum and minimum values ​​within a fixed time period, calculate the number of peaks within the period, and calculate the heart rate value with 60 seconds as a unit time. Calculate the distance between adjacent peaks (RR interval) and the distance between troughs to obtain an average distance (QT interval), and obtain the correction value QTc;

[0115] 2) The calculation method of other cardiac function evaluation indicators is as follows:

[0116] Atrial or ventricular volume:

[0117] Ventricular end-diastolic volume: EDV = max (V);

[0118] End-systolic volume of the ventricle: ESV = min (V);

[0119] Stroke volume: SV = EDV-ESV;

[0120] Area change fraction: FAC = 100 (EDA-ESA) / EDA;

[0121] Ejection fraction: EF = 100 (EDV-ESV) / EDV;

[0122] Ventricular short axis shortening rate: FS = (D d -D s ) / D d ;

[0123] Cardiac output: CO = HR × SV;

[0124] QTc interval: QT C =QT / (RR^0.5);

[0125] Among them, D L : The length of the major axis, D s : length of the minor axis, EDA: end-diastolic area, ESA: end-systolic area, D d : Length of the ventricular minor axis at end-diastole, D s : the length of the short axis at the end of ventricular contraction, HR represents heart rate, RR represents RR interval, QT represents QT interval, QTc interval: QT interval corrected for heart rate;

[0126] S4. Obtain evaluation and screening results:

[0127] Origin and SPSS software were used to analyze the data sets and draw statistical graphs. One-way ANOVA LSD test was used to compare the cardiac morphology (long axis, short axis and area of ​​the atrium and ventricle) and functional evaluation indicators (EDV, ESV, EDA, ESA, EF, SV, FS, FAC, CO, HR, QT interval, RR interval, QTc interval) of zebrafish treated with different types and doses of chemicals. The characteristics of the cardiac developmental toxicity effects of chemicals were described, the dose-response / effect relationship of each indicator was analyzed, the sensitivity of different cardiac developmental toxicity indicators to chemical exposure was compared, high-throughput screening of chemicals with high risk of cardiac toxicity was carried out, and cardiac developmental toxicity was evaluated.

[0128] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity, characterized in that: The following steps are involved: S1. Obtain fluorescent images, videos and average fluorescence intensity data of the heart of Tg(myl7:eGFP) transgenic zebrafish embryos treated with chemicals; S2. Use the YOLOv8 model to perform instance segmentation of morphological recognition on the atrium and ventricle regions in the cardiac fluorescence image obtained in step S1, obtain morphological information parameters of the atrium and ventricle, and combine OpenCV morphological operations to quantitatively output the long axis and short axis of the atrium and ventricle, and the systolic and diastolic area data of the ventricle; S3, calculating a cardiac function evaluation index based on the data obtained in step S1 and step S2; S4. Obtain evaluation and screening results.

2. The AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity according to claim 1, characterized in that: In step S1, a high-content imaging analysis system is used to obtain fluorescent images and average fluorescence intensity data of the heart of Tg(myl7:eGFP) transgenic zebrafish embryos treated with chemicals, and a video is produced to establish an image dataset of the heart of zebrafish embryos exposed to chemicals.

3. The AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity according to claim 2, characterized in that: Step S1 specifically includes the following process: 1) Tg(myl7:eGFP) cardiac transgenic zebrafish were used for mating at a 1:1 ratio of male to female, and the fertilized eggs were treated with chemicals within 5 hours of fertilization and cultured in a constant temperature incubator at 28.5°C with a 14h light / 10h dark cycle; 2) The dye solution was renewed every 24 h, and 210 μL of 5 mM N-phenylthiourea was added 24 h after fertilization to inhibit pigment production, and cultured until 72 h after fertilization; 3) Then transfer to a black transparent bottom 96-well plate, anesthetize the zebrafish larvae with tricaine, centrifuge at 200 rpm for 2 min, adjust the plate so that the ventral side faces the same direction and the eyes overlap; 4) Use a high-content imaging analysis system to acquire images, perform layer scanning, select the FITC fluorescence field 488nm channel, use confocal mode under a 4x objective to acquire images, obtain average fluorescence intensity data, and make a video. Save the first 20 frames of the video as pictures to establish an image data set for the sample.

4. The AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity according to claim 3, characterized in that: Step 4) specifically includes the following process: 4.1) On-line: Run the high-content imaging analysis system, open the MetaXpress software, and load the black transparent-bottom 96-well plate containing chemically treated zebrafish embryos / larvae; 4.2) Set the imaging program parameters: select 4X Plan Apo objective lens, the number of fields of view for each well is 2×2, and each field of view is close to each other; select laser autofocus and Z layer vertical shooting in the acquisition setting, laser autofocus is selected to focus on the bottom of the plate, and then offset according to the thickness of the bottom plate, select FITC fluorescence field 488nm channel, adjust the parameters to achieve an exposure time of 40ms; laser power 80%; focus adjustment mode Z-offset from W1 offset, turn on digital confocal, and shoot the selected wells; 4.3) Make a movie: Make a heart beating video of the transgenic zebrafish. Select StreamAcquisition, set the shooting frame number to 250 frames, and use the FITC fluorescence channel for shooting. Select the Make movie option and select 100 frames to make a 15s video. 4.4) Use Region Measurements to calculate the heart region using the 250 frames of images taken by Stream Acquisition to obtain Average Intensity images and data within 15 seconds.

5. The AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity according to claim 1, characterized in that: Step S2 specifically includes the following process: 1) Use the imgaug library to perform data augmentation on the image dataset through the image distortion algorithm: All data are augmented by different image enhancement algorithms according to random thresholds and random combinations, and the image dataset is divided into training set, validation set and test set in a ratio of 8:1:1; 2) The position information of the atria and ventricles in the amplified data set is respectively passed into the established YOLOv8 instance segmentation model to obtain two models for segmenting the atria and ventricles. The YOLOv8 instance segmentation network model includes a Backbone network structure, a Neck network structure and a Head network structure, which are respectively used for feature extraction, feature fusion and prediction; wherein the Backbone network structure adopts CSPDarknet53, and extracts different levels of information of the image through multiple convolutional layers, residual blocks and C2f modules; the Neck network structure includes a feature pyramid network and a feature attention enhancement network, the feature pyramid network fuses feature maps of different scales through top-down and lateral connections, and the feature attention enhancement network further enhances the expression ability of the feature map, combines global and local information, and ensures that the target detection and segmentation tasks can handle targets of different scales; the Head network structure includes a detection head and a segmentation head, and the detection head and the segmentation head jointly realize the instance segmentation task; 3) Use the trained model to perform instance segmentation of morphological recognition of the atrial and ventricular regions in the zebrafish heart images treated with chemicals to obtain the morphological information parameters of the atrial and ventricular regions: train the YOLOv8 instance segmentation model for identifying and segmenting the atrial and ventricular regions respectively; for the segmentation mask images of the atrial and ventricular regions predicted by the model, use the circumscribed circle and minimum circumscribed rectangle methods in OpenCV to calculate the major axis and minor axis of the atrial / ventricular region, where the diameter of the circumscribed circle is defined as the major axis and the width of the minimum circumscribed rectangle is defined as the minor axis, and finally output the relevant data information.

6. The AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity according to claim 1, characterized in that: In step S3, the cardiac function assessment indicators include heart rate, ventricular end-diastolic volume, ventricular end-systolic volume, ventricular end-diastolic area, ventricular end-systolic area, stroke volume, ventricular short-axis contraction rate, ejection fraction, area change fraction, cardiac output, RR interval, QT interval and QTc interval, a total of 13 indicators.

7. The AI-assisted high-throughput screening method for chemicals with high risk of cardiotoxicity according to claim 6, characterized in that: The heart rate, RR interval, QT interval, and QTc interval are all calculated based on the mean fluorescence intensity data of the zebrafish heart; other indicators are calculated based on the heart rate, major axis, minor axis, and area.

Images