Construction method and detection method of chicken follicle ultrasonic image automatic detection model based on deep learning
By constructing an improved deep learning model, the problems of low identification efficiency and difficulty in localization in chicken follicle detection were solved, realizing the automatic identification and localization of chicken follicles in ultrasound images, and supporting non-destructive observation in the breeding process.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA AGRI UNIV
- Filing Date
- 2026-02-02
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies rely on manual observation for chicken follicle detection, which suffers from low identification efficiency, high subjectivity, difficulty in accurately locating follicle structures in complex ultrasound backgrounds, and a lack of automated detection models suitable for chicken follicles.
An automatic detection model for chicken follicle ultrasound images based on deep learning was constructed. By improving the YOLOv8 model and introducing Haar wavelet transform and spatial adaptive feature modulation modules, an anchor-free target detection method was adopted to achieve automatic identification and localization of follicle structures.
It enables accurate and automatic identification of chicken follicles in ultrasound images, reduces the burden of manual identification, improves the consistency and intuitiveness of identification, and is suitable for auxiliary identification of chicken follicle ultrasound images, supporting non-destructive observation in the breeding process.
Smart Images
Figure CN122023914A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of biological breeding and artificial intelligence image recognition technology, and in particular to a method and system for constructing an automatic detection model of chicken follicle ultrasound images based on deep learning. Background Technology
[0002] As an important livestock breed, the egg production performance of chickens directly affects genetic breeding efficiency and industrial economic benefits. Egg production is the core indicator for measuring egg production performance, and the number of follicles in the ovary and their hierarchical developmental status are key biological factors determining egg production potential. Therefore, obtaining accurate phenotypic information of individual ovarian follicles is of great significance for improving the precision of selection in poultry breeding.
[0003] However, current breeding practices still largely rely on slaughter testing to obtain ovarian and follicle data, which involves dissecting hens to observe follicle morphology and then using this information for sibling selection. This traditional method is not only destructive, costly, and labor-intensive, but it also cannot achieve continuous monitoring of the same breeding individual, significantly limiting the improvement of breeding efficiency.
[0004] In recent years, the development of ultrasound imaging technology has provided a non-invasive means for observing in vivo structures. Ultrasound imaging can display in vivo tissue structures in real time and has been widely used in the medical field for organ detection, lesion localization, and auxiliary diagnosis. With the maturity of intelligent image analysis technology, deep learning-based target detection methods have achieved automatic identification and quantitative measurement in various types of medical ultrasound images, significantly improving detection accuracy and stability.
[0005] While the combination of ultrasound and target detection technology has achieved good applications in the medical field, research on non-destructive in vivo detection of chicken follicles in poultry breeding remains very limited. Existing work largely focuses on mammalian tissue structure identification; for targets like chicken follicles—which are small, densely distributed, and exhibit significant morphological variations—mature detection models and complete technical procedures are still lacking. Therefore, developing an automated detection technology suitable for chicken follicle ultrasound images is of great significance for overcoming the bottlenecks of traditional slaughter measurement and achieving non-destructive and accurate assessment of ovarian development status. Summary of the Invention
[0006] This invention addresses the problems of low recognition efficiency, high subjectivity, and difficulty in accurately locating follicle structures in complex ultrasound backgrounds during chicken follicle ultrasound detection, which mainly relies on manual observation of ultrasound images. It proposes a deep learning-based automatic recognition technology for chicken follicle ultrasound images. This method constructs a deep learning-based automatic detection model for chicken follicle ultrasound images and applies it to the detection of chicken follicles. Without causing harm to the chickens, it can automatically identify and locate follicle structures in chicken follicle ultrasound images, clearly marking follicles from complex ultrasound backgrounds. This provides breeders with intuitive and objective follicle identification results, assisting them in conducting follicle observation and related breeding analysis.
[0007] To achieve the above objectives, the first technical solution of this application discloses a method for constructing an automatic detection model of chicken follicle ultrasound images based on deep learning, characterized by the following steps:
[0008] S1. Dataset Construction: Construct and label a dataset of ultrasound images of chicken follicles based on the abdominal ovarian region of chickens for model training, validation and testing;
[0009] S2. Feature Extraction and Model Training: An improved YOLOv8 model is used for feature extraction and training to obtain an automatic detection model for chicken follicle ultrasound images based on deep learning. The improvement is as follows: a downsampling module based on Haar wavelet transform is introduced into the downsampling structure of the YOLOv8 model backbone network; a spatial adaptive feature modulation module is introduced in the multi-scale feature fusion process of the neck network; and a bounding box-less target detection method is adopted in the detection head.
[0010] Furthermore, the construction and annotation of the chicken follicle ultrasound image dataset based on the ovarian region of the chicken abdominal cavity, as described in S1, includes four steps: ultrasound image acquisition, image processing, image annotation, and dataset partitioning.
[0011] Furthermore, the image processing involves extracting and filtering the acquired ultrasound images using a time-stratified sampling method.
[0012] Furthermore, the image annotation is performed by labeling each processed image individually using the LabelImg annotation tool.
[0013] Furthermore, the features extracted in the feature extraction step described in S3 are the high echogenicity at the center of the chicken follicle dataset, the weak echogenicity at the edges, and the spatial distribution of echo intensity from the center to the edges. These features are obtained by the improved downsampling module described in S2.
[0014] Furthermore, the training process described in S3 is as follows: the constructed and labeled chicken follicle ultrasound image dataset is input into the improved YOLOv8 model for supervised training. The labeled follicle target location information and category information are used as the true labels. The model parameters are optimized through classification loss and bounding box regression loss, and finally a deep learning-based automatic detection model for chicken follicle ultrasound images is obtained.
[0015] Furthermore, the deep learning-based automatic detection model for chicken follicle ultrasound images, obtained according to the above-mentioned method for constructing an automatic detection model for chicken follicle ultrasound images, includes the following modules:
[0016] Data input module: used to input ultrasound image data of chicken follicles in the abdominal cavity and ovarian region of chickens;
[0017] Feature extraction module: used to extract features from input data;
[0018] Data processing module: Used to process input features to obtain the identification result of whether the image under test is a chicken follicle.
[0019] The second technical solution of this application discloses an automatic detection method for chicken follicle ultrasound images. Using the above model, an ultrasound image of the abdominal cavity region of the target chicken is input, and after model processing, the identification result of whether the input image is a chicken follicle or whether chicken follicles exist is obtained.
[0020] And, a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor being used to execute the above-described model.
[0021] Beneficial Effects: This invention addresses the problems of unclear follicle boundary identification, easy missed detection of small follicles, and easy identification of adjacent follicles as a single target in ultrasound images when using conventional target detection models to process chicken follicle ultrasound images without targeted structural optimization. The improved automatic identification model described in this application can more accurately identify the central echo region and its edge structure of the follicle, and adjacent follicles can be labeled separately, resulting in clearer and more stable follicle localization in ultrasound images. Therefore, compared with the prior art, the beneficial effects of this invention are: achieving automatic identification of chicken follicles in ultrasound images, reducing the burden of manual identification; improving the intuitiveness and consistency of follicle identification; assisting breeders in follicle observation without damaging chickens; and providing a complete technical solution that improves the identification effect of follicles in complex ultrasound backgrounds, making it suitable for auxiliary identification scenarios of chicken follicle ultrasound images. Attached Figure Description
[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 A diagram illustrating the dataset annotation process;
[0024] Figure 2 For model improvement diagrams;
[0025] Figure 3 This is a comparison chart of the model recognition results. Detailed Implementation
[0026] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
[0027] The first embodiment of this application discloses a method for constructing an automatic detection model of chicken follicle ultrasound images based on deep learning in order to achieve the above-mentioned objective, comprising the following steps:
[0028] S1. Dataset Construction: Construct and label a dataset of ultrasound images of chicken follicles in the abdominal ovarian region of chickens for model training, validation and testing.
[0029] like Figure 1 The process of constructing and annotating a chicken follicle ultrasound image dataset based on the abdominal ovarian region of chickens is disclosed, which consists of four steps: ultrasound image acquisition, image processing, image annotation, and dataset partitioning.
[0030] (1) Ultrasound image acquisition
[0031] An ultrasound scanner was used to scan the ovarian region of the hen's abdominal cavity to obtain raw ultrasound images or video data containing follicular structures. The raw ultrasound images typically exhibit complex background echoes, blurred follicular boundaries, and low contrast.
[0032] (2) Image processing
[0033] The acquired ultrasound videos were frame extracted and filtered to remove invalid images, thus obtaining ultrasound image samples containing follicular structures.
[0034] (3) Image annotation
[0035] The image samples are manually labeled, with follicle structures categorized into uniform classes, thereby constructing a chicken follicle ultrasound image dataset. During this process, such as... Figure 1 As shown, the starting point of the data is the unprocessed raw ultrasound image (Original Video), and the result of the dataset construction is an ultrasound image sample that can clearly mark the location of follicle structures (Object detection dataset), and then the samples are divided into datasets.
[0036] S2. Feature Extraction and Model Training: An improved YOLOv8 model is used for feature extraction and training to obtain an automatic detection model for chicken follicle ultrasound images based on deep learning. The improvement is as follows: a downsampling module based on Haar wavelet transform is introduced into the downsampling structure of the YOLOv8 model backbone network; a spatial adaptive feature modulation module is introduced in the multi-scale feature fusion process of the neck network; and a bounding box-less target detection method is adopted in the detection head.
[0037] The purpose of the improvement is to address the imaging characteristics of chicken follicle ultrasound images, such as strong noise interference, low contrast, and unclear follicle boundaries, by making targeted improvements to the network structure. Specific improvements include... Figure 2 As shown.
[0038] exist Figure 2 In this study, the traditional downsampling structure is improved for the backbone network by introducing a downsampling module based on Haar wavelet transform. This module performs multi-subband decomposition on the feature map and reorganizes it along the channel dimension, reducing spatial resolution while preserving follicle edge and fine structural features, and minimizing noise interference with effective features. The extracted features include high echogenicity at the center of the chicken follicle dataset, weak echogenicity at the edges, and spatial distribution features of echo intensity varying from the center to the edges, thus obtaining feature representations that characterize follicle morphology and echo patterns.
[0039] For the neck network, the multi-scale feature fusion process is improved by introducing a spatial adaptive feature modulation module. In the upsampling and feature fusion stages, features at different scales are adaptively modulated according to the local spatial feature content to enhance the expressive ability of follicle boundary contours and morphological features.
[0040] For the detection head, an anchorless target detection method is adopted to predict the grid cells on feature maps of different scales and directly output the position information of the follicle target, thereby realizing the automatic identification of chicken follicles in ultrasound images.
[0041] In a further implementation, the training process described in S3 is as follows: the constructed and labeled chicken follicle ultrasound image dataset is input into the improved YOLOv8 model for supervised training, the labeled follicle target location information and category information are used as the real labels, the model parameters are optimized through classification loss and bounding box regression loss, and finally a deep learning-based automatic detection model for chicken follicle ultrasound images is obtained.
[0042] The model construction method will be explained in detail below through specific embodiments.
[0043] Example 1: Constructing a Deep Learning-Based Automatic Detection Model for Chicken Ovary Ultrasound Images
[0044] S1. Dataset Construction: Construct and label a dataset of ultrasound images of chicken follicles in the abdominal ovarian region of chickens for model training, validation and testing.
[0045] In this embodiment, the experimental subjects were 383 Yingde free-range chickens, and the ultrasound equipment was a DCU10 fully digital color Doppler ultrasound diagnostic instrument with a 6.5 MHz micro-convex probe. The ultrasound image acquisition process included the following steps:
[0046] (1) Fix the chickens flat on the experimental table;
[0047] (2) Remove downy hairs from the area where the chicken’s left thigh root meets the left side of the abdomen, and apply medical ultrasound coupling agent;
[0048] (3) Place the micro-convex probe vertically in the area and adjust the probe angle until a clear circular or elliptical follicle structure appears in the ultrasound display interface;
[0049] (4) Freeze the current ultrasound image and save the corresponding ultrasound video data.
[0050] (5) Image processing was performed on the acquired ultrasound videos. The time-layer sampling method was used to extract the original images at a rate of 2 frames per second, and redundant data at the beginning and end of the ultrasound videos were removed. Subsequently, the extracted images were filtered to remove images that did not contain follicle targets or had duplicate follicle locations. Finally, a dataset containing 1,927 high-quality chicken follicle ultrasound images was constructed.
[0051] (6) Manually annotate the constructed dataset. Use the LabelImg annotation tool to annotate each of the chicken follicle ultrasound images, uniformly label the follicle target as a single category label "follicle", and save the corresponding annotation information in a YOLO format text file, thereby constructing an annotated dataset for training the target detection model.
[0052] (7) During the dataset partitioning stage, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. To avoid data leakage caused by ultrasound images of the same chicken appearing in different subsets, this paper uses the individual numbers of the 383 chickens as the basis for partitioning. First, the chickens are grouped, and then all image samples of the corresponding individuals are allocated to the same subset, thereby ensuring that the training set, validation set, and test set are independent of each other at the individual level.
[0053] S2. Feature Extraction and Model Training: An improved YOLOv8 model is used for feature extraction and training to obtain an automatic detection model for chicken follicle ultrasound images based on deep learning. The improvement is as follows: a downsampling module based on Haar wavelet transform is introduced into the downsampling structure of the YOLOv8 model backbone network; a spatial adaptive feature modulation module is introduced in the multi-scale feature fusion process of the neck network; and a bounding box-less target detection method is adopted in the detection head.
[0054] Chicken follicle ultrasound images are input into a YOLOv8-based deep learning target detection network for feature extraction and recognition. The YOLOv8 network includes a backbone network, a neck network, and a detection head, which are used to perform feature extraction, multi-scale feature fusion, and follicle target recognition output, respectively.
[0055] To address the common characteristics of high noise, low contrast, blurred follicle boundaries, and dense distribution of small targets in chicken follicle ultrasound images, a targeted improvement to the YOLOv8 network structure was made to construct an automatic recognition model for chicken follicle ultrasound images. The model includes:
[0056] Data input module: used to input ultrasound image data of chicken follicles in the abdominal cavity and ovarian region of chickens;
[0057] Data processing module: used for feature extraction and processing of input data;
[0058] Data output module: Used to output the processed image and obtain the identification result of whether the image under test is a chicken follicle.
[0059] Through the above improvements, the constructed automatic follicle ultrasound image recognition model based on the improved YOLOv8 has a more stable follicle recognition effect under conditions of low contrast, blurred boundaries and dense distribution of small targets.
[0060] Example 2: Comparison of Model Recognition Results
[0061] One target chicken was selected, and its abdominal ovarian region was scanned with ultrasound to obtain four raw ultrasound images of the chicken follicles, as shown below. Figure 3The image shown is the Origin Image, and its labeling results are also published. Figure 3 -Labeling Result), then processed by the existing YOLO-HS model and the improved YOLOv8 model of this application respectively, to obtain a comparison chart of the processed results ( Figure 3 -YOLO-HS Result (Pre-existing YOLO-HS model processing) and ( Figure 3 -YOLOv8 Result (processed by the improved YOLOv8 model in this application) shows that the recognition results of the improved YOLOv8 model in this application are more accurate than the recognition results of the prior art. It can identify follicles with lower contrast and blurred boundaries, and has more stable follicle recognition results.
[0062] Therefore, through the identification results described in this application, breeders can visually identify the distribution and quantity of follicles in ultrasound images, thereby assisting them in observing and recording follicle structures. It should be noted that the identification results output by this invention are used to assist in the identification and observation of follicle structures and do not directly make judgments on the degree of follicle development or breeding decisions.
[0063] The above-described 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A method for constructing an automatic detection model for chicken follicle ultrasound images based on deep learning, characterized in that, Includes the following steps: S1. Dataset Construction: Construct and label a dataset of ultrasound images of chicken follicles based on the abdominal ovarian region of chickens for model training, validation and testing; S2. Feature Extraction and Model Training: An improved YOLOv8 model is used for feature extraction and training to obtain an automatic detection model for chicken follicle ultrasound images based on deep learning. The improvement is as follows: a downsampling module based on Haar wavelet transform is introduced into the downsampling structure of the YOLOv8 model backbone network; a spatial adaptive feature modulation module is introduced in the multi-scale feature fusion process of the neck network; and a bounding box-less target detection method is adopted in the detection head.
2. The method for constructing the model according to claim 1, characterized in that, The construction and annotation of the chicken follicle ultrasound image dataset based on the abdominal ovarian region of chickens, as described in S1, includes four steps: ultrasound image acquisition, image processing, image annotation, and dataset partitioning.
3. The method for constructing the model according to claim 2, characterized in that, The image processing involves extracting and filtering the acquired ultrasound images using a time-stratified sampling method.
4. The method for constructing the model according to claim 2, characterized in that, The image annotation is performed by labeling each processed image individually using the LabelImg annotation tool.
5. The method for constructing the model according to claim 1, characterized in that, The features extracted in the feature extraction step of S3 are the high echogenicity at the center of the chicken follicle dataset, the weak echogenicity at the edge, and the spatial distribution of echo intensity from the center to the edge. These features are obtained by the improved downsampling module described in S2.
6. The model construction method according to claim 1, characterized in that, The training process described in S3 is as follows: The constructed and labeled chicken follicle ultrasound image dataset is input into the improved YOLOv8 model for supervised training. The labeled follicle target location information and category information are used as the true labels. The model parameters are optimized through classification loss and bounding box regression loss, and finally, a deep learning-based automatic detection model for chicken follicle ultrasound images is obtained.
7. An automatic detection model for chicken follicle ultrasound images based on deep learning, obtained by the method for constructing an automatic detection model for chicken follicle ultrasound images based on deep learning according to claims 1-6, characterized in that, Includes the following modules: Data input module: used to input ultrasound image data of chicken follicles in the abdominal cavity and ovarian region of chickens; Data processing module: used for feature extraction and processing of input data; Data output module: Used to output the processed image and obtain the identification result of whether the image under test is a chicken follicle.
8. An automatic detection method for chicken follicle ultrasound images, characterized in that, Using the model described in claim 6, an ultrasound image of the abdominal cavity region of the target chicken is input, and after processing by the model, the identification result of whether the input image is a chicken follicle or whether a chicken follicle exists is obtained.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor being configured to execute the model of claim 7.