Oyster gonad development level high-throughput precise phenotype analysis method based on U-Net deep learning and application of oyster gonad development level high-throughput precise phenotype analysis method

Abstract

The invention belongs to the technical field of oyster phenotype analysis, particularly relates to an oyster gonad development level high-throughput accurate phenotype analysis method based on U-Net deep learning, and effectively solves the problems of complicated operation, high subjectivity and low precision and throughput in existing oyster gonad development phenotype identification. The method comprises the following steps: acquiring an oyster gonad tissue slice image, constructing and training a U-Net deep learning segmentation model, automatically identifying and segmenting a gonad tissue region in the image, further accurately calculating the Gonadal Proportion (GP) of the gonad tissue in the total gonad area, and realizing high-throughput and accurate quantification of the gonad development level. According to the method, the gonad tissue and the connective tissue can be accurately distinguished, objective phenotypic data are provided, key technical support is provided for gonad development character genetic parameter estimation and molecular assisted breeding of large-scale groups, and the efficiency, accuracy and practicability of gonad development phenotypic analysis are improved.

Description

Technical Field

[0001] This invention belongs to the field of oyster phenotypic analysis technology, specifically relating to a precise phenotypic analysis method for quantifying the gonadal development level of the Pacific oyster. Background Technology

[0002] There are over 20,000 species of bivalves worldwide, mainly including oysters, scallops, clams, and mussels. These species occupy an important position in aquaculture due to their extremely high commercial and scientific value. Gonadal development in bivalves is one of the important reproductive traits, with some individuals within the same population exhibiting varying levels of gonadal development. Accurate quantification of gonadal development is crucial for research in shellfish reproductive biology, helping to predict the potential for genetic improvement. However, taking the Pacific oyster as an example, its gonads are diffusely distributed and have a complex tissue structure. Traditional methods, such as calculating the gonadal index (GSI) by gonadal weighing, have limitations such as being cumbersome, time-consuming, unable to accurately distinguish between gonadal tissue and connective tissue, and easily influenced by the subjective factors of researchers, making it difficult to achieve high-precision, high-throughput phenotypic identification of large-scale populations (e.g., hundreds to thousands of individuals).

[0003] With the development of computer vision technology, image segmentation methods have shown potential in the emerging phenotypic trait quantification in agriculture. As one of these technologies, the U-Net network can detect different features during image extraction and achieve good results in image segmentation. However, there is still no mature and accurate solution for the precise segmentation of the specific, fine structure of the oyster gonad. Developing a phenotypic analysis method that can stably, automatically, and objectively quantify the degree of gonadal development is crucial for achieving high-throughput genetic assessment of oyster reproductive traits and accelerating the molecular breeding process.

[0004] Based on this, the present invention is proposed. Summary of the Invention

[0005] This invention proposes a method for automatically segmenting and quantifying gonad proportion (GP) from gonad tissue slice images, effectively overcoming the shortcomings of existing technologies such as difficult operation and large errors, and providing an efficient and accurate method for oyster GP phenotypic analysis. This method utilizes U-net deep learning technology to automatically identify and segment gonad tissue regions from standard gonad tissue slice images, accurately calculating their proportion of the entire gonad sample region (including gonad tissue and connective tissue), thereby providing a novel and standardized quantitative indicator of gonadal development.

[0006] The technical solution of the present invention is as follows:

[0007] The technical solution of this invention is to provide a precise quantitative analysis method for oyster GP phenotype based on U-Net deep learning, comprising:

[0008] S1. Image Acquisition and Preprocessing: Digital microscopic images of paraffin-embedded sections of gonadal tissue after staining were acquired; the original images were standardized and preprocessed, including resizing, color space conversion and contrast enhancement, to form the model input dataset.

[0009] S2. Training Dataset Construction: Randomly select a portion of the sample images and manually and finely annotate them using an image annotation tool. Classify each pixel as "gonadal tissue" or "background (mainly connective tissue)" to generate the corresponding binary segmentation image. Divide the annotated image pairs into training and validation sets.

[0010] S3. Deep Learning Segmentation Model Construction and Training: A U-Net network model based on an encoder-decoder architecture is constructed. The encoder is used for downsampling to extract multi-scale high-level semantic features and contains a 4-level downsampling module, each level containing two 3x3 convolutional layers and a 2x2 max pooling layer. The decoder is used for upsampling to restore spatial resolution and contains a 4-level upsampling module, each level containing a 2x2 transposed convolutional or interpolated upsampling layer and two 3x3 convolutional layers. Feature fusion is performed between corresponding levels of the encoder and decoder through skip connections. The model uses the root mean square propagation (RMSprop) optimizer to optimize the model parameters using the training set and monitors the model performance using the validation set.

[0011] S4. Model Performance Evaluation: The trained model is evaluated using test set images. Its segmentation performance metrics at the pixel level are calculated, including Mean Intersection over Union (MIoU), Accuracy, Precision, and Recall, to ensure that the model meets the segmentation accuracy requirements. Among these, the higher the MIoU value, the better the model's prediction performance.

[0012] S5. Batch Precise Phenotypic Calculation: Input the batch of gonadal tissue slice images to be analyzed into the trained optimal model. The model automatically outputs the binary segmentation mask corresponding to each image. Based on the mask, the total pixel area (A_gonad) of the "gonadal tissue" category and the total pixel area (A_total) of the "gonadal tissue" and "background" categories in each image are calculated by image analysis algorithms. Finally, the gonadal proportion (GP) value of each sample is automatically calculated according to the formula GP (%) = (A_gonad / A_total) × 100%.

[0013] S6. Phenotypic Data Output and Application: The calculated GP values, the corresponding original images, and the segmentation results are associated, stored, and visualized to form a standardized phenotypic database;

[0014] Furthermore, in step S3, the loss function of the model is the binary cross-entropy loss function, with the following formula:

[0015] ,

[0016] in, The label represents sample i (positive class = 1 and negative class = 0). This represents the probability that sample i is predicted to be of the positive class.

[0017] Furthermore, in step S4, the model performance evaluation requires MIoU to be no less than 70% and Accuracy to be no less than 85%.

[0018] The beneficial effects of the technical solution provided in this application are as follows:

[0019] Precision and objectivity: This method, based on a deep learning-based segmentation model, can accurately distinguish between gonadal tissue and connective tissue. The GP index it provides is more refined and objective than traditional GSI or visual rating, reducing human error.

[0020] Standardization and reproducibility: This method establishes a standardized analytical process and quantitative indicators for gonadal development traits, making the research results from different batches and different laboratories comparable, which is conducive to data sharing and collaborative research.

[0021] Driver genetics research provides essential high-quality, high-dimensional phenotypic data for genetic analysis of reproductive traits in large populations (such as heritability estimation and GWAS), and is a key technological link connecting the genome and phenome.

[0022] With broad application prospects, this method is not only applicable to oysters, but its technical framework, after adaptation, can be extended to the reproductive phenotypic analysis of other aquatic economic species with complex glandular structures. Attached Figure Description

[0023] Figure 1 is a flowchart of the U-net method for gonadal proportion (GP) phenotype analysis.

[0024] Figure 2 shows the frequency distribution of the proportion of gonads (GP) in 680 Pacific oysters. Detailed Implementation Plan

[0025] The following embodiments illustrate the solution of the present invention. The core of the method of the present invention lies in using the U-Net model to learn and distinguish the general texture and morphological features of gonadal tissue and connective tissue. Although this embodiment uses the Pacific oyster as the implementation object, those skilled in the art will understand that the core steps and model architecture of this method, after being adapted with appropriate training data, can be applied to other bivalve animals with similar diffuse gonadal structures, such as clams and mussels; the following embodiments are only used to further illustrate the present invention and should not be construed as limiting the scope of protection of the present invention.

[0026] Traditional quantitative methods for gonadal development phenotypes in Pacific oysters are difficult to operate and have large errors. To solve these problems, the first embodiment of this application discloses a GP evaluation method based on the deep learning U-net algorithm to quantify gonadal development traits in Pacific oysters and estimate their genetic parameters.

[0027] Among them, GP showed a weak negative correlation with growth traits (approximately -0.08), while there were significant positive correlations among the various growth traits (phenotypic correlation 0.27-0.63; genetic correlation 0.48-0.81).

[0028] In this embodiment, the estimated heritability of gonadal development is at a medium-to-high level. Furthermore, phenotypic and genetic results show that independent selection can be implemented on gonadal development without affecting growth traits. The method of this embodiment includes:

[0029] S1. Image Acquisition and Preprocessing:

[0030] Sixty-eighty 12-month-old Pacific oysters were collected from the waters off Rongcheng (Shandong) for measurement of their growth and reproductive traits. Shell length and height were measured using calipers, and total weight and shell weight were measured using an analytical balance, with length and weight accuracy of 0.01 mm and 0.01 g, respectively. The oysters were dissected, and gonadal tissue was collected and preserved in paraformaldehyde solution for 24 hours, then transferred to 70% ethanol. The samples were subsequently dehydrated, embedded, sectioned, and stained to obtain gonadal sections. These sections were observed and photographed using an optical microscope, and HE-stained images were preserved. The images were then processed using image processing software, and a masking technique was applied to remove connective tissue from the entire image.

[0031] S2. Training dataset construction:

[0032] Use ImageJ software to adjust the grayscale threshold and calculate the gonadal area ratio: The total gonadal area includes gonadal tissue (area) and connective tissue (area). Figure 2U-net was used to convert the images into a clear gonadal development phenotype; 200 images were randomly selected as the training set and another 50 images were selected as the test set; Labelme was used for manual annotation to generate binary segmentation images, and the remaining images were validated using U-net; the GP value was calculated by dividing the area of ​​the gonadal region in the image by the total area.

[0033] S3. Construction and training of deep learning segmentation models:

[0034] The experimental framework consists of an encoding module and a decoding module: the encoding module is responsible for extracting high-level semantic features from the input image, processing the image through multiple convolutional layers, and then downsampling using max pooling; each convolutional layer uses a 3x3 kernel (filter) size to detect different features such as edges and complex structures in the image. The convolution stride is set to 1, meaning the filter moves one pixel at a time; padding is set to 0 (no padding) to ensure the output feature map size decreases after each convolution operation; downsampling uses max pooling with a 2x2 kernel size to reduce the spatial dimension of the feature map; the encoding module performs four rounds of downsampling, with two convolutional layers set before each round to extract high-level semantic information from the image; the decoding module uses convolution and upsampling operations corresponding to the encoding module, with upsampling achieved through 2x2 kernel size interpolation, filling missing pixels by estimating neighboring pixel values; four rounds of upsampling are performed to gradually restore the original image size, with two convolutional layers set before each round to obtain low-level semantic information lost during downsampling; feature fusion is achieved through channel concatenation, combining low-level and high-level feature information, allowing subsequent network layers to freely obtain local fine-grained details and global context information, thereby making better predictions based on the complexity of different regions in the image; 3x3 kernel size convolution operations are used for pixel classification and to generate prediction output; in order to efficiently train the model and solve the optimization problem in the experiment, this embodiment uses the root mean square propagation (RMSprop) algorithm;

[0035] Assume the input training dataset is:

[0036] S= ;

[0037] in, This represents the i-th input training image. This represents the segmentation and annotation image corresponding to the given image;

[0038] The objective function is as follows:

[0039] ;

[0040] Since this invention relates to the segmentation of gonadal images, therefore It was calculated using a loss function based on binary cross-entropy:

[0041] ;

[0042] in, The label represents sample i (positive class = 1 and negative class = 0). This represents the probability that sample i is predicted to be of the positive class.

[0043] S4. Model Performance Evaluation:

[0044] This dataset is used to evaluate the accuracy of the IoU (Intersection over Union) model. The performance of the pixel-level segmentation model is evaluated by accuracy, precision-recall and their respective averages, as shown in the following formula;

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] In this embodiment, the training set is based on an NVIDIA GeForce MX150 GPU for model training and evaluation, CUDA and cuDNN are used for GPU acceleration, and the best trained model is tested on the test set to obtain evaluation metrics.

[0054] In this example, the model achieved an average intersection-over-union (MIoU) of 74.76%, with higher scores for gonadal tissue segmentation (79.45%) and lower scores for connective tissue (70.06%). Accuracy reached 86.12%, demonstrating strong performance across both tissue types (gonads: 88.08%; connective tissue: 84.16%). Precision was 84.98%, exceptionally high for gonadal tissue (99.48%) but lower for connective tissue (70.49%). Furthermore, recall showed the opposite trend, at 79.79% for gonadal tissue and 99.14% for connective tissue. These results indicate robust training. Loading the trained model parameters into memory and inputting all gonadal slice images into the model, the total recognition time was 708.25 s, with an average recognition time per individual between 1.2 and 1.3 s.

[0055] S5. Batch Precise Phenotyping Calculation:

[0056] The 680 gonadal tissue slice images to be analyzed are input into the trained optimal model. The model automatically outputs a binary segmentation mask for each image. Based on the mask, the total pixel area (A_gonad) of the "gonadal tissue" category and the total pixel area (A_total) of the "gonadal tissue" and "background" categories in each image are calculated using image analysis algorithms. Finally, the GP value of each sample is automatically calculated according to the formula GP (%) = (A_gonad / A_total) × 100%.

[0057] S6. Output and Application of Phenotypic Data:

[0058] The calculated GP values, corresponding original images, and segmentation results are associated, stored, and visualized to form a standardized phenotypic database. The heritability (h) of gonadal development and growth-related traits is assessed using the linear mixture model (LAM) and linear paternal-maternal model (LSM) in ASReml software (3.0). 2 Phenotypic correlation (rp) and genetic correlation (rg) were estimated; all statistical analyses were performed using SPSS 26 statistical software (IBM, Armonk, NY, USA), and the significance level was uniformly set to P<0.05.

[0059] Furthermore, based on the high-precision GP phenotypic data of individual Pacific oysters obtained in batches through the above steps, this embodiment further verifies its direct application value in breeding. Using the standardized GP phenotypic database generated in S6, combined with the pedigree relationships between individuals, the results are shown in Table 1: When using LAM, the heritability estimate of GP is high (0.477 ± 0.094 and 0.576 ± 0.102); after applying LSM and considering the effects of sex and age, the heritability estimate is moderate (0.369 ± 0.065 and 0.386 ± 0.059), that is, when sex is included as a fixed effect, the heritability value increases, while the included age has no effect on the estimate.

[0060] Table 1. Heritability estimation (± standard deviation) of gonadal development (GP) trait using two models.

[0061]

[0062] The phenotypic and genetic correlations between GP and growth-related traits (TW, SH, and SL) are shown in Table 2. Phenotypic correlations were not significant, low, and close to 0, ranging from -0.01 ± 0.06 to 0.03 ± 0.05; genetic correlations showed the same trend, ranging from -0.08 ± 0.17 to 0.08 ± 0.18. In contrast, phenotypic and genetic correlations among growth traits were significant and positive. The phenotypic correlation between TW and SH was very high (0.63 ± 0.04), between TW and SL was moderate (0.48 ± 0.04), and between SH and SL was low (0.27 ± 0.05). On the other hand, the genetic correlation between TW and SH was very high (0.81 ± 0.06), between TW and SL was moderate (0.55 ± 0.12), and between SH and SL was also moderate (0.48 ± 0.14).

[0063] Table 2. Phenotypic correlations (above the diagonal) and genetic correlations (below the diagonal) between gonadal development (GP) and growth-related traits.

[0064]

[0065] This key finding demonstrates that it is feasible to implement independent or synergistic selection of gonadal development traits without affecting or even improving growth performance.

[0066] In breeding practice, this method provides a solution for high-throughput and accurate phenotypic identification of gonadal development traits in the Pacific oyster, specifically:

[0067] Live sampling, sectioning, image analysis, and GP calculation were performed on candidate parent populations. Individuals with superior gonadal development were directly selected as parents based on GP phenotypic values ​​to construct a high-fertility core breeding population.

[0068] In genetic evaluation models, using the precise GP values ​​obtained by this method as phenotypic inputs can more accurately estimate the individual's breeding value (EBV), thereby improving the accuracy of selection for reproductive traits.

[0069] Based on the conclusion that GP is not significantly related to the genetics of growth traits, breeders can develop an independent selection index for gonadal development traits, or combine it with growth traits to construct a comprehensive selection index, thereby achieving simultaneous genetic improvement of growth and reproductive performance.

[0070] In summary, the automated quantitative method for gonad ratio based on the U-Net model provided by this invention not only solves the problems of difficult operation and large errors of traditional methods, but also produces high-quality phenotypic data that directly serves breeding practices, providing key technical tools and decision-making basis for the precise genetic improvement of reproductive traits in Pacific oysters and similar species.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-throughput, precise phenotypic analysis method for oyster gonadal development based on U-Net deep learning, characterized in that, include: Digital microscopic images of paraffin sections stained from gonadal samples of Pacific oysters were obtained, and the original images were standardized and preprocessed to form the model input dataset. A portion of the sample images is selected, and each pixel is manually and precisely labeled using an image annotation tool. Each pixel is classified as either "gonadal tissue" or "background," generating a corresponding binary segmentation image. The labeled image pairs are then divided into training and validation sets. A U-Net network model based on an encoder-decoder architecture is constructed. The encoder is used for downsampling to extract multi-scale high-level semantic features and includes a multi-level downsampling module, with each level containing a convolutional layer and a pooling layer. The decoder is used for upsampling to restore spatial resolution and includes a multi-level upsampling module, with each level containing an upsampling layer and a convolutional layer. Feature fusion is performed between corresponding levels of the encoder and decoder through skip connections. The model is trained using the training set, and the model performance is monitored using a validation set. The trained model is evaluated using independent test set images, and its segmentation performance metrics at the pixel level are calculated to ensure that the model meets the preset segmentation accuracy requirements. Batch accurate phenotypic calculation: Input the batch of gonadal tissue slice images to be analyzed into the trained optimal model, and the model automatically outputs the binary segmentation mask corresponding to each image; based on the mask, calculate the total pixel area (A_gonad) of the "gonadal tissue" category and the total pixel area (A_total) of gonadal tissue (including gonadal tissue and background) in each image; automatically calculate the gonadal proportion (GP) value of each sample according to the formula GP (%) = (A_gonad / A_total) × 100%; Phenotypic data output and application: The calculated GP values ​​are associated and stored with the corresponding sample information to form a standardized phenotypic database.

2. The method for precise quantitative phenotypic analysis of oyster gonad ratio based on U-Net deep learning according to claim 1, characterized in that, The encoder of the U-Net model contains four downsampling modules, each of which contains two 3x3 convolutional layers and one 2x2 max pooling layer; the decoder contains four upsampling modules, each of which contains one 2x2 transposed convolutional or interpolated upsampling layer and two 3x3 convolutional layers. The model was trained using a binary cross-entropy loss function and a root mean square propagation (RMSprop) optimizer.

3. The method for precise quantitative phenotypic analysis of gonadal ratio in *Oryctosus longifolius* based on U-Net deep learning according to claim 1, characterized in that, The segmentation performance metrics include at least the Mean Intersection over Union (MIoU) and overall accuracy; the model performance evaluation requires MIoU to be no less than 70% and accuracy to be no less than 85%.

4. The method for precise quantitative phenotypic analysis of gonadal ratio in *Oryctosus longifolia* based on U-Net deep learning according to claim 1, characterized in that, The preprocessing operations include adjusting image size, color space conversion, and contrast enhancement; the "background" mainly refers to the connective tissue region in the gonadal tissue slice image.

5. The method for precise quantitative phenotypic analysis of gonadal ratio in *Oryctosus longifolia* based on U-Net deep learning according to claim 1, characterized in that, The loss function of the model is the binary cross-entropy loss function, and the calculation formula is as follows: ； in, The label represents sample i (positive class = 1 and negative class = 0). This represents the probability that sample i is predicted to be of the positive class.

6. The method according to claim 1, characterized in that, The GP values ​​in the standardized phenotypic database are used for subsequent genetic evaluation and breeding analysis.

7. The method according to any one of claims 1-6, characterized in that, The method is applicable to animals of the genus Crassostrea.

8. An electronic device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method for accurate quantitative phenotypic analysis of gonad ratio in Ostrea gigas based on U-Net deep learning as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the steps of the method for precise quantitative phenotypic analysis of gonad ratio in Ostrea gigas based on U-Net deep learning as described in any one of claims 1 to 7.

Images