Artificial intelligence-based animal embryo grading method and system

Abstract

The present application relates to the technical field of image recognition and processing, in particular to an animal embryo grading method and system based on artificial intelligence, which solves the problems of strong subjectivity, incomplete feature extraction and poor generalization of the prior art. The system includes image acquisition, preprocessing, feature extraction, model training and grading output. The method realizes grading through acquisition, preprocessing, feature extraction, model training and grading output. The present application accurately captures the characteristics of multiple species of embryos, improves the accuracy of grading, is efficient in processing, suitable for large-scale breeding, and improves the success rate of embryo transfer / hatching.

Description

Technical Field

[0001] This invention relates to the field of image recognition and processing technology, specifically to an artificial intelligence-based method and system for grading animal embryos. Background Technology

[0002] Animal breeding technology is a core support for improving the economic efficiency of the livestock industry and ensuring the supply of high-quality livestock products, while animal embryo quality assessment and grading are key links in determining the success rate of breeding superior livestock. With the development of large-scale farming and bio-breeding technology, higher demands are placed on the accuracy, objectivity, and efficiency of animal embryo grading. However, existing animal embryo grading technologies still have many problems that urgently need to be solved, which seriously restricts the improvement of breeding efficiency.

[0003] Currently, animal embryo grading primarily relies on manual assessment, with embryologists making subjective judgments based on morphological characteristics (such as the number, uniformity, and fragmentation rate of blastomeres) combined with their own experience. However, animal embryos exhibit significant species-specific differences: bovine / sheep embryos are relatively large (approximately 150 μm) and expand slowly in the blastocoel; mouse embryos are smaller (approximately 80 μm) and cleave rapidly; avian embryos develop in the form of a blastodisc and are not CO2-dependent. These differences make it difficult to standardize manual assessment criteria. Different embryologists have significantly different understandings of embryos from different species. Even within the same species, the consistency in grading among assessors with different experience is only 55%-70%, highlighting the issues of strong subjectivity and poor repeatability.

[0004] To overcome the limitations of manual grading, some studies have attempted to apply AI technology to animal embryo grading, but existing technologies still have the following obvious shortcomings.

[0005] First, it is mostly based on single-modal bright-field images, ignoring the complementary information of different imaging modalities of animal embryos. For example, polarized light images of bovine / sheep embryos can clearly show the inner cell mass structure, and infrared images of avian embryos can reflect metabolic activity. Single-modal data leads to incomplete feature extraction.

[0006] Secondly, image preprocessing methods lack species adaptability. Different animal embryo images have different types of noise (e.g., bovine embryo images are easily affected by impurities in the culture dish, producing salt-and-pepper noise, while mouse embryo images are easily affected by light, producing Gaussian noise) and degrees of distortion. General denoising and registration algorithms are difficult to effectively eliminate interference.

[0007] Third, the segmentation model is not good at recognizing small structures in animal embryos, such as blastomeres in early mouse embryos (only 10-15 μm in diameter) and the arrangement of trophoblast cells in avian embryonic discs. Traditional U-Net networks are prone to missed or false detections in segmentation.

[0008] Fourth, feature extraction only focuses on static morphology and ignores the species-specific dynamic developmental patterns of animal embryos. For example, the 2-8 cell stage of mouse embryos only requires 24 hours, while that of bovine embryos requires more than 48 hours. The lack of dynamic temporal features limits the accuracy of grading.

[0009] Fifth, the models have poor generalization ability. Most existing models are trained based on data from a single species, making it difficult to adapt to the differences in characteristics of different animal embryos. Furthermore, they lack confidence assessment and abnormality indication, resulting in high risks in clinical application.

[0010] Furthermore, existing technologies have not solved the engineering challenges of animal embryo grading, such as the rapid switching between embryos of multiple species and the real-time grading requirements in large-scale breeding scenarios. Therefore, developing an animal embryo grading method and system that integrates multimodal species-specific information, accurately captures dynamic and static characteristics, and possesses high generalization ability and reliability has become an urgent technical problem to be solved in this field. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and provide an artificial intelligence-based method and system for grading animal embryos. Through a combination of multi-dimensional creative technologies, this invention solves the technical problems of existing technologies, such as strong subjectivity, incomplete feature extraction, low grading accuracy, poor generalization ability, and lack of species adaptability, thereby achieving accurate, objective, and efficient grading of the quality of animal embryos of different species.

[0012] The technical solution adopted by this invention to solve its technical problem is: an artificial intelligence-based animal embryo grading system, comprising: an image acquisition module, used to acquire dynamic image sequences of different species of animal embryos during development through a multimodal imaging device, the dynamic image sequence including brightfield images and phase contrast images, optionally equipped with polarized light images or infrared images, the multimodal imaging device adjusting the objective lens magnification according to the animal species, the image acquisition module including a constant temperature culture unit adapted to different species; and a preprocessing module, used to preprocess the dynamic image sequence, including a registration unit that uses the SIFT feature point matching algorithm to align multimodal images at the same time point, a noise suppression unit that uses a combination of adaptive Gaussian filtering and median filtering to remove species-specific image noise, and a fusion unit that uses a species-adaptive weighted fusion algorithm to adaptively learn the weights assigned to different modal images through a deep learning model with the embryo key structure identification accuracy as the loss function, outputting standardized image data; The feature extraction module is used to segment embryonic regions in standardized image data and extract features. It includes an image segmentation unit that uses an improved U-Net network with dilated convolution and spatial attention mechanisms to improve the segmentation accuracy of animal embryonic microstructures; a feature computation unit that calculates static morphological features using morphological analysis algorithms and dynamic temporal features using temporal difference algorithms; a model training module that constructs and trains a hybrid deep learning model based on ResNet and Transformer. This hybrid deep learning model embeds channel attention and temporal attention mechanisms, focusing on the inner cell mass, trophoblast region, and species-specific features, and is trained using a species-diverse dataset; and a graded output module that inputs the combined static and dynamic features into the trained hybrid deep learning model and outputs animal embryo quality grades I-IV, confidence scores, and anomaly alerts. This graded output module has a built-in species-adaptive confidence threshold and outputs anomaly alerts via audible and visual alarms or interface pop-ups.

[0013] Specifically, the image acquisition module includes a bright-field camera (12-megapixel CMOS sensor), a phase-contrast microscope (with switchable objective magnification of 10× / 20× / 40×, adaptable to animal embryos of different sizes), a species-specific constant temperature incubator (cattle / sheep: 38℃±0.1℃, CO2 5%±0.1%; mice: 37℃±0.1℃, CO2 5%±0.1%; poultry: 38.5℃±0.1℃, no CO2), and an image transmission unit (using 5G or fiber optic transmission, with a rate ≥1Gbps).

[0014] Specifically, the decoder part of the improved U-Net network is equipped with dilated convolutional layers with a dilation rate of 2 and 4. The spatial attention mechanism enhances attention to microstructures by generating attention maps, wherein the microstructures include blastomeres or blastocysts of animal embryos.

[0015] Specifically, the static morphological characteristics include the number of blastomeres, uniformity, roundness, fragment area ratio, blastocoel expansion grade, inner cell mass morphological parameters, and trophoblast cell arrangement parameters; the dynamic temporal characteristics include the time interval between adjacent cleavages, morphological change rate, blastocoel expansion rate, and cleavage synchronicity.

[0016] Specifically, in the hybrid deep learning model, ResNet50 is used as a static feature encoder, Transformer encoder is used to capture dynamic temporal dependencies, and channel attention mechanism adaptively adjusts the weights of species-specific feature channels.

[0017] An AI-based animal embryo grading method, implemented using the aforementioned AI-based animal embryo grading system, includes the following steps: S1, acquiring dynamic image sequences of target species animal embryos from the pronuclear stage to the blastocyst stage using a multimodal imaging device, adjusting the acquisition interval according to the species: once every 1 hour for mice, once every 2-3 hours for cattle, and once every 3 hours for poultry, with an image resolution of 1024×1024 pixels and a bit depth of 16 bits; S2, preprocessing: registering multimodal images using the SIFT feature point matching algorithm, with a registration error ≤1 pixel; removing noise using an adaptive hybrid denoising algorithm, adaptively adjusting the Gaussian kernel of brightfield images from 3×3 to 7×7, and using 5×5 median filtering for phase difference images; obtaining standardized image data using a species-adaptive weighted fusion algorithm. S3. Feature Extraction: The embryonic region in the standardized image data is segmented using an improved U-Net network. Dice loss plus cross-entropy loss is used to solve class imbalance. Static morphological features and dynamic temporal features are extracted and standardized into a 256-dimensional static feature vector plus a 128-dimensional dynamic feature vector. S4. Model Training: A multi-species labeled dataset is constructed and divided into training, validation, and test sets in a 7:2:1 ratio. The model is optimized using cross-entropy loss plus focus loss, and the training rhythm is adjusted using a cosine annealing learning rate decay strategy to obtain a species-adapted hierarchical model. S5. Hierarchical Output: The joint features of the embryos to be graded are input into the species-adapted hierarchical model, and the quality level and confidence score are output. If the confidence score is lower than the species adaptation threshold, an anomaly is triggered and feature anomalies are output.

[0018] Specifically, in step S2, the weight learning process of the species adaptive weighted fusion algorithm enhances the structural information of polarized light images for animal embryos and optimizes the fusion of temperature-related features of infrared images for avian embryos.

[0019] Specifically, in step S3, the number of blastomeres in the static morphological features is counted through connected region analysis, the uniformity in the static morphological features is the coefficient of variation of the blastomer area, the roundness in the static morphological features is calculated by multiplying 4π by the blastomer area and then dividing by the square of the blastomer perimeter, and the blast synchronicity in the dynamic temporal features is the standard deviation of the division time of each blastomer in the same cell stage.

[0020] Specifically, in step S4, the temporal attention mechanism of the species-adapted hierarchical model assigns higher weight to the short-cycle dynamic features of mouse embryos and strengthens the temporal dependence capture of the long-cycle developmental features of bovine embryos.

[0021] Specifically, in step S5, the species-adaptive threshold is determined through clinical trials: 0.80-0.85 for bovine / sheep embryos, 0.78-0.82 for mouse embryos, and 0.75-0.80 for avian embryos.

[0022] The beneficial effects of this invention are as follows.

[0023] ① Adaptive fusion preprocessing of multimodal species eliminates image interference from different species, integrates complementary information, and lays the foundation for accurate classification.

[0024] ② An improved U-Net network, combined with a spatial attention mechanism, solves the problem of segmenting small structures in animal embryos with high segmentation accuracy.

[0025] ③ By combining static and dynamic feature extraction, the morphology and developmental patterns of embryos from different species can be fully captured, significantly improving feature recognition.

[0026] ④ The ResNet plus Transformer hybrid model embeds a dual attention mechanism, which enhances species-specific features and temporal dependence, resulting in high classification accuracy and excellent generalization ability.

[0027] ⑤ Species-adaptive confidence threshold and anomaly alert function reduce clinical application risks and improve system usability; ⑥ The system has fast processing speed to meet the needs of large-scale breeding.

[0028] This invention effectively solves many of the shortcomings of existing technologies, provides an efficient and precise technical tool for animal breeding, and has significant economic value and application prospects. Attached Figure Description

[0029] Figure 1 An architecture diagram of the artificial intelligence-based animal embryo grading system provided by this invention;

[0030] Figure 2 A flowchart of the artificial intelligence-based animal embryo grading method provided by the present invention. Detailed Implementation

[0031] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0032] like Figure 1 As shown, the artificial intelligence-based animal embryo grading system of the present invention includes an image acquisition module, a preprocessing module, a feature extraction module, a model training module, and a grading output module. Each module is connected through a data bus to achieve efficient data transmission and collaborative processing.

[0033] (1) Image acquisition module: used to acquire multimodal dynamic image sequences of animal embryos of different species. The core components include a bright-field camera (12-megapixel CMOS sensor, frame rate 15fps), a phase-contrast microscope (objective magnification 10× / 20× / 40× switchable, numerical aperture 0.5, adapted to animal embryos of different sizes: 10× for avian embryonic discs of about 200μm, 20× for bovine / pig embryos of about 150μm, and 40× for mouse / rat embryos of about 80μm), species-specific... The system includes a heterogeneous constant temperature incubator (cattle / sheep: 38℃±0.1℃, CO2 5%±0.1%, humidity 70%±5%; mice: 37℃±0.1℃, CO2 5%±0.1%, humidity 50%±5%; poultry: 38.5℃±0.1℃, humidity 60%±5%, no CO2 control), an LED cold light source (brightness 100-1000 lux, color temperature 3000-6000K adaptive adjustment), and an image transmission unit (5G or fiber optic transmission, rate ≥1Gbps). Optional polarized light imaging components (cattle / sheep embryos) or infrared imaging components (poultry embryos) can be added to achieve multimodal data acquisition.

[0034] (2) Preprocessing module: This module is used to standardize dynamic image sequences, including a registration unit, a noise suppression unit, and a fusion unit. The registration unit uses the SIFT feature point matching algorithm, which extracts feature points from multimodal images, uses K-nearest neighbor matching, and removes mismatched points using the RANSAC algorithm. It calculates the transformation matrix to achieve accurate alignment of images at the same time point, with a registration error of ≤1 pixel, thus solving the problem of spatial misalignment in multimodal images. The noise suppression unit uses an adaptive hybrid denoising algorithm, which adaptively adjusts the 3×3-7×7 Gaussian kernel for Gaussian noise in bright field images and uses 5×5 median filtering for salt-and-pepper noise in phase difference images. Combined with bilateral filtering, it preserves edge details and ensures effective noise suppression for embryo images of different species. The fusion unit uses a species-adaptive weighted fusion algorithm, which uses a deep learning model with the recognition accuracy of key embryonic structures (blastomeres, blastocoel, inner cell mass) as the loss function. It adaptively learns the weights of multimodal images, enhances the polarized light structure information of bovine / sheep embryos, optimizes the infrared metabolic features of avian embryos, and outputs standardized image data with comprehensive information.

[0035] (3) Feature Extraction Module: Used for embryo region segmentation and feature extraction, including an image segmentation unit and a feature calculation unit. The image segmentation unit adopts an improved U-Net network with dilated convolution and spatial attention mechanism. The decoder part adds dilated convolutional layers with a dilation rate of 2 and 4 to expand the receptive field without reducing the resolution. The spatial attention mechanism generates an attention map to enhance the attention to the microstructures of the animal embryo. The loss function adopts a weighted combination of Dice loss and cross-entropy loss to solve the problem of imbalance between embryo and background categories. The segmentation speed is ≥10 frames / second. The feature calculation unit calculates features through a preset algorithm (morphological analysis algorithm plus temporal difference algorithm): static morphological features include the number of blastomeres (connected region analysis count), uniformity (area variation coefficient), and roundness ( The data were calculated by multiplying 4π by the area of ​​the blastomeres and then dividing by the square of the blastomeres' circumference. Other features included: fragmentation area ratio (fragmentation area / total embryo area); blastocyst cavity expansion grade (grades 1-6); inner cell mass morphological parameters (area, roundness, compactness); and trophoblast cell arrangement parameters (intercellular space ratio, orderly arrangement). Dynamic temporal features included the time interval between adjacent cleavages (time difference from n-cell stage to n+1-cell stage), morphological change rate (rate of change of parameters at adjacent time points), blastocyst cavity expansion rate (area growth rate), and cleavage synchronicity (standard deviation of division time at the same stage). The feature data were standardized into a 256-dimensional static vector + a 128-dimensional dynamic vector.

[0036] (4) Model Training Module: Used to build and train hybrid deep learning models, including a dataset management unit, a model building unit, and a training optimization unit. The dataset management unit stores more than 10,000 multi-species labeled data (covering cattle, sheep, mice, and poultry), and supports data import, labeling, and 7:2:1 partitioning; the model building unit builds a hybrid ResNet50 and Transformer model, with ResNet50 as a static feature encoder and Transformer capturing dynamic temporal dependencies, embedding a channel attention mechanism (adaptively adjusting the weights of species-specific feature channels) and a temporal attention mechanism (strengthening the temporal correlation of dynamic features of embryos of different species); the training optimization unit uses the Adam optimizer, cross-entropy loss plus focus loss (weight parameter 2), and a cosine annealing learning rate decay strategy (initial learning rate 0.001), and saves the model with the highest accuracy on the validation set after 100 training rounds.

[0037] (5) Grading Output Module: Used for grading result output and anomaly prompts, including grading unit, confidence assessment unit and anomaly prompt unit. The grading unit calls the trained model and outputs quality grades I-IV (Grade I: Excellent, Grade II: Good, Grade III: Average, Grade IV: Poor). The confidence assessment unit calculates the confidence score based on the model's predicted probability entropy value. The lower the entropy value, the higher the confidence. The anomaly prompt unit has built-in species adaptive thresholds (0.80-0.85 for cattle / sheep, 0.78-0.82 for mice, and 0.75-0.80 for poultry). When the confidence is lower than the threshold, anomaly prompts and characteristic anomalies (such as excessive fragmentation rate and abnormal cleavage time interval) are output through audible and visual alarms and interface pop-ups. It also supports database storage and PDF / Excel export of grading results, confidence, and anomalies.

[0038] like Figure 2 As shown, the artificial intelligence-based animal embryo grading method of the present invention is implemented using the above-mentioned system and includes the following steps.

[0039] S1. Multimodal Dynamic Image Sequence Acquisition: Place the target species animal embryo culture dish into the corresponding constant temperature incubator, adjust the parameters of the multimodal imaging equipment (objective magnification, exposure time), start the acquisition program, and acquire dynamic image sequences at species-specific intervals (mouse once every 1 hour, covering the 2-8 cell stage to the blastocyst stage for 48 hours; cattle once every 2-3 hours, covering the pronuclear stage to the blastocyst stage for 7 days; birds once every 3 hours, covering the embryonic disc formation to the blastocyst stage for 96 hours). The acquired image sequences include bright field + phase contrast images, and polarized light / infrared images can be selected. The images are transmitted to the server in real time via 5G / fiber optic and stored in TIFF format.

[0040] S2. Multimodal image preprocessing.

[0041] ① Registration: SIFT algorithm is used to align multimodal images at the same time point, extract feature points, K-nearest neighbor matching, RANSAC to remove mismatched points, calculate the transformation matrix and adjust the phase difference image, and the registration error is ≤1 pixel.

[0042] ② Noise suppression: The noise type is identified by a noise detection algorithm, the bright field image is subjected to adaptive Gaussian filtering, the phase difference image is subjected to 5×5 median filtering, and finally bilateral filtering is used to preserve the edges. After denoising, PSNR≥35dB and SSIM≥0.92.

[0043] ③ Fusion: The species-adaptive weighted fusion algorithm is adopted, and the optimal weights are learned through the CNN model. After fusion, the image information entropy is ≥7.5 and the edge intensity is ≥20, resulting in standardized image data.

[0044] S3, Feature Extraction.

[0045] ①Embryo region segmentation: The standardized image is input into the improved U-Net network to segment the embryo region, eliminating interference from the edge of the culture dish and air bubbles, and the segmentation DSC≥0.94.

[0046] ② Static morphological feature extraction: Based on the segmentation results, 7 types of static features are calculated using morphological analysis algorithms and standardized into 256-dimensional vectors.

[0047] ③ Dynamic temporal feature extraction: Based on image sequences, four types of dynamic features are calculated using a temporal difference algorithm and standardized into 128-dimensional vectors.

[0048] ④ Joint feature construction: Combine static and dynamic feature vectors to form a 384-dimensional joint feature.

[0049] S4, Model Training.

[0050] ① Dataset construction: 12,000 images of embryos from multiple species were collected (3,000 each from cattle, sheep, mice, and birds), and the images were labeled by consensus voting from three embryologists with more than 5 years of experience.

[0051] ② Model construction: Build a hybrid ResNet50 and Transformer model, embedding channel attention and temporal attention mechanisms.

[0052] ③ Model training: Set batchsize=32, train for 100 rounds, validate every 5 rounds, save the best model, and the model accuracy on the validation set after training is ≥93%.

[0053] S5. Graded Output: Input the combined features of the embryos to be graded into the optimal model and output the quality grade and confidence score; if the confidence score is greater than or equal to the species adaptive threshold, output the result directly; if the confidence score is less than the threshold, trigger an anomaly prompt and output the feature anomaly items, prompting manual review.

[0054] Example 1: Specific implementation of bovine embryo grading.

[0055] I. Experimental equipment and data preparation.

[0056] Experimental equipment: Multimodal imaging equipment (bright-field camera MV-CA012-10GM, 12 megapixels; phase contrast microscope IX73, 20× objective lens; constant temperature incubator MINC-180, 38℃±0.1℃, CO2 5%±0.1%, humidity 70%±5%; LED cold light source LG-100, polarization light assembly); computing equipment (GPU server ThinkSystem SR860, 2 Intel Xeon Gold6248 processors, 4 NVIDIA Tesla V100 GPUs, 128GB memory); operating system Ubuntu 18.04 LTS, deep learning framework PyTorch 1.10.

[0057] Data Preparation: Data from 3000 Holstein dairy cow embryos from three large-scale dairy breeding centers were collected. All embryos were from in-vitro fertilization cycles and cultured to the blastocyst stage (7 days). For each embryo, 40 images were acquired using bright-field plus phase-contrast and polarized light imaging sequences, acquired at 2-3 hour intervals. Image resolution was 1024×1024 pixels, and bit depth was 16 bits. Labeling was performed by three experts with over 8 years of experience in bovine embryo grading, referencing the industry grading standards for bovine embryos: Grade I (uniform blastomeres, fragmentation rate <5%, fully expanded blastocoel, compact inner cell mass), Grade II (good uniformity, fragmentation rate 5%-10%, mostly expanded blastocoel), Grade III (moderate uniformity, fragmentation rate 10%-20%, partially expanded blastocoel), and Grade IV (poor uniformity, fragmentation rate >20%, poor blastocoel expansion). The embryos were divided into a training set of 2100 cases, a validation set of 600 cases, and a test set of 300 cases in a 7:2:1 ratio.

[0058] II. Specific implementation of the methods and steps.

[0059] S1. Image Acquisition: Place the bovine embryo culture dish in a constant temperature incubator, adjust the phase contrast microscope to 20× objective lens, the bright field camera exposure time to 15ms, and the polarization light component brightness to 800 lux. Acquire image sequences at 2.5-hour intervals, transmit them to the GPU server via optical fiber, and store them in the format of "bovine embryo_date_time_modality.tif". Monitor the incubator environmental parameters in real time to ensure stable temperature and CO2 concentration.

[0060] S2, Preprocessing.

[0061] ① Registration: The SIFT algorithm was implemented using the OpenCV library to extract feature points (threshold 0.03) from the bright field, phase difference, and polarized light images at the same time point. K-nearest neighbor (k=2) matching was used, and the RANSAC algorithm (1000 iterations, confidence level 0.99) was used to remove mismatched points. The affine transformation matrix was calculated, and the polarized light and phase difference images were adjusted. The average registration error was 0.6 pixels.

[0062] ② Noise Suppression: Bovine embryo images are prone to salt-and-pepper noise caused by impurities in the culture dish. The phase difference image is filtered with a 5×5 median filter. The bright field and polarized light images are adaptively selected with a 3×3-7×7 Gaussian kernel according to the noise variance. Finally, bilateral filtering (spatial sigma=50, grayscale sigma=50) is applied. After denoising, the average PSNR is 37.2dB and the average SSIM is 0.94.

[0063] ③ Fusion: A CNN model was constructed to learn fusion weights, with the accuracy of identifying the inner cell mass and blastocyst cavity of bovine embryos as the loss function. After 50 training rounds, the fusion weights were w1 (bright field) = 0.4, w2 (phase difference) = 0.3, and w3 (polarized light) = 0.3. The average information entropy of the fused image was 7.8, and the average edge intensity was 22.5.

[0064] S3, Feature Extraction.

[0065] ①Segmentation: The improved U-Net network was used as input to the fused image. The decoder had a hole ratio of 2 and 4. The loss function was Dice loss: cross-entropy loss = 1:1. After 100 training rounds, the average segmentation DSC was 0.95, and the inner cell mass of bovine embryos (approximately 20 μm in diameter) was successfully segmented.

[0066] ②Static characteristics: The number of blastomeres was counted by connected region analysis (area threshold 80 pixels), the uniformity was the area variation coefficient (average 0.12), the roundness was 0.85, the fragmentation area ratio was 8.3% on average, the blastocyst cavity expansion grade was 4.2 on average, the inner cell mass area was 1200 pixels on average, the compactness was 0.88 on average, and the trophoblast intercellular space ratio was 11.2% on average.

[0067] ③ Dynamic characteristics: The average time interval between adjacent cleavages is 12.5 hours, the average rate of morphological change is 0.08 / h, the average rate of blastocyst expansion is 150 pixels / h, and the average standard deviation of cleavage synchronicity is 1.8 hours; ④ Joint characteristics: spliced ​​into a 384-dimensional vector.

[0068] S4. Model Training: ResNet50 pre-trained weights fine-tuned, 6 layers of Transformer encoder (8 heads for multi-head attention), channel attention mechanism to enhance the channel weights of inner cell mass features, temporal attention mechanism to focus on the long-term dynamic features of bovine embryos, batch size=32, initial learning rate 0.001, cosine annealing decay, loss function cross-entropy loss + focus loss (alpha=0.25, gamma=2), training for 100 rounds.

[0069] S5. Graded Output: Input the model with 300 bovine embryos for the test set, and output the graded results and confidence levels. The bovine embryo threshold was set to 0.82. 27 cases had a confidence level <0.82, triggering an abnormality warning. Manual review revealed that 24 cases had excessively high fragmentation rates (>15%) or loose inner cell masses.

[0070] Example 2: Specific implementation of mouse embryo grading.

[0071] I. Experimental equipment and data preparation.

[0072] Experimental equipment: multimodal imaging equipment (bright field camera MV-CA012-10GM, 12 megapixels; phase contrast microscope IX73, 40× objective lens; constant temperature incubator MINC-180, 37℃±0.1℃, CO2 5%±0.1%, humidity 50%±5%; LED cold light source LG-100); computing equipment as in Example 1; software environment PyTorch 1.10, Python 3.8.

[0073] Data Preparation: Data from 3000 ICR mouse embryos were collected from an in vitro fertilization cycle at the blastocyst stage (48 hours). For each embryo, 30 bright-field and phase-contrast images were acquired at 1-hour intervals. Image resolution was 1024×1024 pixels with a bit depth of 16 bits. Labeling was performed by three experts with over 6 years of experience in mouse embryo research. Grading criteria were: Grade I (homogeneous blastomeres, fragmentation rate <3%, fully expanded blastocoel), Grade II (good homogeneity, fragmentation rate 3%-8%), Grade III (moderate homogeneity, fragmentation rate 8%-15%), and Grade IV (poor homogeneity, fragmentation rate >15%). The embryos were divided into a training set of 2100, a validation set of 600, and a test set of 300, using a 7:2:1 ratio.

[0074] II. Specific implementation of the methods and steps.

[0075] S1. Image Acquisition: The mouse embryo culture dish was placed in a constant temperature incubator. The phase contrast microscope was adjusted to a 40× objective lens, the bright field camera exposure time was 10ms, and the LED cold light source brightness was 600 lux. Image sequences were acquired at 1-hour intervals and transmitted to the GPU server via 5G. The storage format was "mouse embryo_number_time_modality.tif". Environmental parameters were monitored in real time to avoid temperature fluctuations affecting embryo development.

[0076] S2, Preprocessing.

[0077] ① Registration: SIFT algorithm extracts feature points of bright field and phase difference images, K-nearest neighbor matching, RANSAC removes mismatched points, and the average registration error is 0.5 pixels, which solves the problem of slight misalignment caused by high magnification of mouse embryo images.

[0078] ② Noise suppression: Mouse embryo images are susceptible to Gaussian noise due to illumination. A 3×3-5×5 Gaussian kernel was selected adaptively according to variance. A 5×5 median filter was used for phase difference images. Bilateral filtering preserved the edge details of the blastomere. After noise reduction, the average PSNR was 38.5dB and the average SSIM was 0.95.

[0079] ③ Fusion: Species-adaptive weighted fusion algorithm with learning weights w1 (bright field) = 0.5 and w2 (phase difference) = 0.5. The loss function is the recognition accuracy of mouse embryo blastomeres. The average information entropy of the fused image is 7.9 and the average edge intensity is 23.1, clearly presenting early blastomeres with a diameter of 10-15μm.

[0080] S3, Feature Extraction.

[0081] ①Segmentation: The improved U-Net network is optimized for the microstructure of mouse embryos, with a cavity rate of 2 and 4. The spatial attention mechanism enhances the focus on blastomeres. The loss function combines Dice loss and cross-entropy loss, and the average segmentation DSC is 0.96, successfully eliminating interference from small impurities at the edge of the culture dish.

[0082] ②Static characteristics: The number of blastomeres was counted by connected region analysis (area threshold 30 pixels), the uniformity was the area variation coefficient (average 0.09), the roundness was 0.88, the fragmentation area ratio was 6.2% on average, the blastocyst cavity expansion grade was 4.5 on average, the inner cell mass area was 800 pixels on average, the compactness was 0.90 on average, and the trophoblast intercellular space ratio was 9.8% on average.

[0083] ③ Dynamic characteristics: The average time interval between adjacent cleavages was 4.2 hours, the average rate of morphological change was 0.12 / h, the average rate of blastocyst expansion was 200 pixels / h, and the average standard deviation of cleavage synchronicity was 0.9 hours.

[0084] ④ Joint features: concatenated into a 384-dimensional vector.

[0085] S4. Model Training: In the hybrid model, the temporal attention mechanism assigns higher weight to the short-cycle dynamic features of mouse embryos, capturing the rapid changes in cleavage; with a batch size of 32 and 100 training rounds, the validation set accuracy is 95.8%, and the model can accurately identify the rapid cleavage process of mouse embryos in the 2-8 cell stage.

[0086] S5. Graded Output: The test set included 300 mouse embryos input into the model, with the mouse embryo threshold set at 0.80. 22 cases had a confidence level <0.80, and manual review revealed that 19 cases had poor cleavage synchronicity or excessive fragmentation rate.

[0087] Example 3: Specific implementation of avian embryo grading.

[0088] I. Experimental equipment and data preparation.

[0089] Experimental equipment: multimodal imaging equipment (bright field camera MV-CA012-10GM, 12 megapixels; phase contrast microscope IX73, 10× objective lens; constant temperature incubator MINC-180, 38.5℃±0.1℃, humidity 60%±5%; LED cold light source LG-100, infrared imaging component); computing equipment same as in Example 1; software environment PyTorch 1.10, Python 3.8.

[0090] Data Preparation: Data from 3000 chicken embryos from two poultry breeding bases were collected from artificial insemination cycles and cultured to the blastocyst stage (96 hours). For each embryo, 32 images each of bright field, phase contrast, and infrared sequences were collected, with a 3-hour interval between acquisitions. The image resolution was 1024×1024 pixels and the bit depth was 16 bits. Labeling was performed by three experts with over 7 years of experience in avian embryo grading. Grading criteria were: Grade I (intact embryonic disc, homogeneous blastomeres, fragmentation rate <5%, good blastocoel expansion), Grade II (relatively intact embryonic disc, good homogeneity, fragmentation rate 5%-10%), Grade III (incomplete embryonic disc, moderate homogeneity, fragmentation rate 10%-18%), and Grade IV (damaged embryonic disc, poor homogeneity, fragmentation rate >18%). The embryos were divided into a training set of 2100 cases, a validation set of 600 cases, and a test set of 300 cases in a 7:2:1 ratio.

[0091] II. Specific implementation of the methods and steps.

[0092] S1. Image Acquisition: Place the chicken embryo culture dish in a constant temperature incubator, adjust the phase contrast microscope to 10× objective lens, the bright field camera exposure time to 20ms, the infrared imaging component temperature sensitivity to ±0.1℃, and the LED cold light source brightness to 700 lux. Acquire image sequences at 3-hour intervals, transmit them to the GPU server via optical fiber, and store them in the format "chicken embryo_base_time_modality.tif". Monitor the humidity of the incubator to prevent the embryo disc from drying out.

[0093] S2, Preprocessing.

[0094] ① Registration: The SIFT algorithm aligns bright field, phase difference, and infrared images with an average registration error of 0.7 pixels, solving the problem of multimodal image misalignment caused by the large area of ​​the avian embryonic disc (approximately 200μm).

[0095] ② Noise suppression: Avian embryo images are prone to salt-and-pepper noise caused by feather debris. The phase difference and infrared images are filtered with a 5×5 median filter, and the bright field images are filtered with a 3×3-7×7 Gaussian kernel adaptively selected according to the variance. Bilateral filtering preserves the edge of the embryonic disc. After denoising, the average PSNR is 36.8dB and the average SSIM is 0.93.

[0096] ③ Fusion: Species adaptive weighted fusion algorithm, with learning weights w1 (bright field) = 0.3, w2 (phase difference) = 0.3, w3 (infrared) = 0.4, using the chicken embryo blastodisc recognition accuracy as the loss function, the fused image integrates the metabolic activity information of the infrared image, with an average information entropy of 7.7 and an average edge intensity of 21.8.

[0097] S3, Feature Extraction.

[0098] ①Segmentation: The improved U-Net network is optimized for large-sized avian blastodiscs with a void ratio of 2 and 4. The spatial attention mechanism enhances the focus on the edge of the blastodisc and the inner cell mass, and the average segmentation DSC is 0.94, successfully segmenting irregularly shaped avian blastodiscs.

[0099] ②Static characteristics: The number of blastomeres was counted by connected region analysis (area threshold 100 pixels), the uniformity was the area variation coefficient (average 0.13), the roundness was 0.83, the fragmentation area ratio was 9.5% on average, the blastocyst cavity expansion grade was 4.0 on average, the inner cell mass area was 1500 pixels on average, the compactness was 0.86 on average, and the trophoblast intercellular space ratio was 12.5% ​​on average.

[0100] ③ Dynamic characteristics: The average time interval between adjacent cleavages was 8.6 hours, the average rate of morphological change was 0.07 / h, the average rate of blastocyst expansion was 120 pixels / h, and the average standard deviation of cleavage synchronicity was 1.5 hours.

[0101] ④ Joint features: concatenated into a 384-dimensional vector.

[0102] S4. Model Training: In the hybrid model, the channel attention mechanism strengthens the channel weights of metabolic features in infrared images, and the temporal attention mechanism focuses on the long-term developmental patterns of avian embryos. The batch size is 32, and the training lasts for 100 rounds.

[0103] S5. Graded Output: The test set included 300 chicken embryos input into the model, with the threshold for poultry embryos set at 0.78. 31 cases had a confidence level <0.78, and manual review revealed that 26 cases had blastodisc damage or excessive fragmentation.

[0104] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The descriptions in the above embodiments and specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence based animal embryo grading system characterized in that, The application relates to an animal embryo quality grading method based on multi-species multi-modal image data, comprising the following steps: An image acquisition module is used for acquiring a dynamic image sequence of animal embryos in different species during development by a multi-modal imaging device, wherein the dynamic image sequence comprises bright field images and phase contrast images, and optional polarized light images or infrared images, the multi-modal imaging device adjusts an objective magnification according to the animal species, and the image acquisition module comprises a constant-temperature culture unit suitable for different species; A preprocessing module is used for preprocessing the dynamic image sequence, comprising a registration unit which adopts a SIFT feature point matching algorithm to align multi-modal images at the same time node, a noise suppression unit which adopts a method combining adaptive Gaussian filtering and median filtering to remove species-specific image noise, and a fusion unit which adopts a species adaptive weighted fusion algorithm, wherein the weight allocated to different modal images is adaptively learned by a deep learning model taking embryo key structure recognition accuracy as a loss function, and standardized image data is output; A feature extraction module is used for segmenting an embryo region in the standardized image data and extracting features, comprising an image segmentation unit which adopts an improved U-Net network containing a hollow convolution and a spatial attention mechanism to improve the segmentation accuracy of animal embryo microstructures, and a feature calculation unit which calculates static morphological features through a morphological analysis algorithm and calculates dynamic time sequence features through a time sequence difference algorithm; the static morphological features include blastomere number, uniformity, roundness, fragment area proportion, blastocyst cavity expansion level, inner cell mass morphological parameters and trophoblast cell arrangement parameters; the dynamic time sequence features include adjacent cleavage time interval, morphological change rate, blastocyst cavity expansion rate and cleavage synchrony; A model training module is used for constructing and training a hybrid deep learning model based on ResNet and Transformer, the hybrid deep learning model embeds channel attention mechanism and time sequence attention mechanism, focuses on the inner cell mass, trophoblast region and species-specific features, and is trained by using a species diversified data set; A hierarchical output module is used for inputting static and dynamic joint features into the trained hybrid deep learning model, outputting animal embryo I-IV grade quality grades, confidence scores and abnormal prompts, and the hierarchical output module is internally provided with a species adaptive confidence threshold, and abnormal prompts are output through sound-light alarm or interface pop-up window.

2. The artificial intelligence-based animal embryo grading system according to claim 1, characterized by: The image acquisition module comprises a bright field camera, a phase contrast microscope with switchable objective magnifications of 10x / 20x / 40x to adapt to different sizes of animal embryos, a species-specific constant-temperature incubator and an image transmission unit.

3. The artificial intelligence-based animal embryo grading system according to claim 1, characterized by: The decoder part of the improved U-Net network is provided with hollow convolution layers with hollow rates of 2 and 4, and the spatial attention mechanism strengthens the attention to microstructures by generating an attention map, wherein the microstructures include animal embryo blastomeres or blastocyst cavities.

4. The artificial intelligence-based animal embryo grading system according to claim 1, characterized by: In the hybrid deep learning model, ResNet50 is used as a static feature encoder, a Transformer encoder is used to capture dynamic time sequence dependence, and a channel attention mechanism is used to adaptively adjust the channel weight of species-specific features.

5. An artificial intelligence-based animal embryo grading method, which is implemented using the artificial intelligence-based animal embryo grading system according to any one of claims 1 to 4, characterized by, The application further discloses an animal embryo quality grading method based on multi-species multi-modal image data, comprising the following steps: S1, collecting dynamic image sequences of target species animal embryos from pronucleus stage to blastocyst stage by multi-modal imaging equipment, adjusting the collection interval according to the species, wherein the mouse is collected every 1 hour, the cow is collected every 2-3 hours, and the bird is collected every 3 hours, and the image resolution is 1024*1024 pixels and the bit depth is 16 bits; S2, preprocessing: using SIFT feature point matching algorithm to register multi-modal images, the registration error is less than or equal to 1 pixel; removing noise by adaptive mixed denoising algorithm, the Gaussian kernel of bright field image is 3*3-7*7 adaptive adjustment, and the phase difference image is 5*5 median filter; using species adaptive weighted fusion algorithm to obtain standardized image data; S3, feature extraction: segmenting the embryo region in the standardized image data by the improved U-Net network, using Dice loss plus cross-entropy loss to solve the class imbalance; extracting static morphological features and dynamic timing features, and standardizing them into 256-dimensional static feature vectors and 128-dimensional dynamic feature vectors; S4, model training: constructing a multi-species labeled data set, dividing the training set, validation set and test set according to the ratio of 7:2:1, using cross-entropy loss plus focus loss to optimize the model, and using cosine annealing learning rate decay strategy to adjust the training rhythm, to obtain a species-adapted hierarchical model; S5, hierarchical output: inputting the joint features of the embryo to be graded into the species-adapted hierarchical model, outputting the quality grade and confidence score, if the confidence is lower than the species adaptive threshold, triggering an abnormal prompt and outputting the feature abnormal item.

6. The artificial intelligence-based animal embryo grading method according to claim 5, characterized in that: In step S2, the weight learning process of the species adaptive weighted fusion algorithm strengthens the structural information of the animal embryo polarization light image, and optimizes the fusion of the temperature related features of the infrared image of the bird embryo.

7. The artificial intelligence-based animal embryo grading method of claim 5, wherein: In step S3, the number of blastomeres in the static morphological features is counted by connected region analysis, the uniformity in the static morphological features is the coefficient of variation of blastomere area, the circularity in the static morphological features is calculated by multiplying 4π by the area of the blastomere and then dividing by the square of the circumference of the blastomere, and the cleavage synchronization in the dynamic timing features is the standard deviation of the division time of each blastomere at the same cell stage.

8. The artificial intelligence-based animal embryo grading method of claim 5, wherein: In step S4, the timing attention mechanism of the species-adapted hierarchical model gives higher weight to the short-period dynamic features of mouse embryos and strengthens the timing dependence capture of long-period development features of cow embryos.

Images