Animal body fat detection method and animal body fat detection model generation method and device

Abstract

The invention belongs to the technical field of B-mode ultrasound diagnosis, and particularly relates to an animal body fat detection method and an animal body fat detection model generation method and device.The detection method comprises the steps that animal body fat B-mode ultrasound images are collected, and data preprocessing is conducted; the B ultrasonic image is converted into an image of a CT / MRI style; according to the converted CT / MRI style image, scanning lines with clear boundary lines of a fat layer and a muscle layer are obtained; calculating the pixel number of the upper and lower edges of the fat layer; and the actual fat thickness is calculated according to the conversion ratio of the detection device. The generation method of the detection model comprises the steps of data preparation, model building, model training and model deployment. According to the method, the B-mode ultrasound image is converted into the CT or MRI image through the cyclic generative adversarial network technology, speckle noise and artifacts in the B-mode ultrasound image are effectively removed, the boundary definition of a fat layer and a muscle layer is remarkably improved, and the fat thickness measurement error is greatly reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of B-ultrasound diagnosis, and particularly relates to a method for detecting animal body fat, a method for generating an animal body fat detection model, and a device therefor. Background Art

[0002] Ultrasonic measurement of animal body fat is a non-invasive detection technology widely used in the livestock industry. It mainly analyzes the thickness of the fat layer and the muscle layer through B-mode ultrasound images (B-ultrasound images) to evaluate the health status and feeding quality of animals. However, the traditional ultrasonic measurement technology has the following problems in practical applications: I. Image quality problems: 1) Speckle noise, artifacts and other interferences are common in B-ultrasound images, which lead to blurred boundaries between the fat layer and the muscle layer. Especially for animals with a thin fat layer or a small body size, the degree of boundary blur is higher, increasing the difficulty of segmentation and thickness measurement.

[0003] 2) The resolution of B-ultrasound is relatively low, which limits the accurate recognition of complex structures and small target areas.

[0004] II. Limitations of traditional algorithms: 1) Traditional image processing methods (such as filtering or threshold-based segmentation algorithms) perform poorly when processing B-ultrasound images with more noise, are vulnerable to artifact interference, and result in unstable measurement results or large errors.

[0005] 2) Existing methods lack effective extraction and enhancement of complex texture features and boundary information in images. Summary of the Invention

[0006] To solve the problems of noise interference and low resolution in traditional B-ultrasound images, the present invention provides a method for detecting animal body fat, a method for generating an animal body fat detection model, and a device therefor. By using the unsupervised image conversion ability of the cyclic adversarial generation network, the B-ultrasound image is converted into a high-quality image with a CT / MRI style, enhancing the boundary clarity of the fat layer and the muscle layer while retaining the original information of the ultrasound image. Through this conversion, the clarity and contrast of the image are significantly improved, thereby improving the measurement accuracy of body fat thickness.

[0007] To achieve the above object, the present invention adopts the following technical solutions: A method for detecting animal body fat based on a cyclic adversarial generation network model, the specific steps are as follows: Step 1: Collect B-ultrasound images of animal body fat and perform data preprocessing; Step 2: Use the trained animal body fat detection model based on the cyclic adversarial generation network to convert the collected B-ultrasound image into an image with a CT / MRI style; Step 3: Obtain scan lines with clear boundaries between the fat layer and the muscle layer based on the converted CT / MRI-style images; Step 4: Identify the upper and lower edges of the fat layer based on the demarcation point information on the scan lines, and calculate the number of pixels of the upper and lower edges of the fat layer; Step 5: Calculate the actual fat thickness based on the number of pixels of the upper and lower edges of the fat layer and the conversion ratio of the detection device.

[0008] Furthermore, a method for generating an animal body fat detection model includes: Data preparation: Mainly collect target modality images and B-ultrasound image data, and preprocess the collected images and data; Model construction: Build a cyclic adversarial generation network model through the design of a generator, a discriminator, and a loss function; Model training: Train and debug the model through a pre-training stage, unsupervised training, model evaluation, and iterative optimization; When the model meets the expectations after training and optimization, complete the model deployment.

[0009] Furthermore, the target modality images are CT or MRI images; the B-ultrasound image data should include obvious structures of the fat and muscle layers.

[0010] Furthermore, in the data preparation step, the preprocessing is to perform size and grayscale processing on all images.

[0011] Furthermore, the generator design uses a convolutional neural network based on residual blocks.

[0012] Furthermore, the discriminator is designed as a PatchGAN discriminator.

[0013] Furthermore, the loss function design includes generative adversarial loss, cycle consistency loss, and identity loss.

[0014] Furthermore, the loss function design is specifically as follows: a. Generative adversarial loss: Through adversarial training, make the generated image approximate the style of the target modality to achieve image style conversion; b. Cycle consistency loss: Ensure that the B-ultrasound image remains consistent with the original B-ultrasound image after two generations of conversion; c. Identity loss: Increase the loss of converting an image of the same modality into itself to ensure that the style remains stable during the conversion process.

[0015] Furthermore, the model training step is as follows: (1) Pre-training stage: First, perform pre-training on a small-scale B-ultrasound and CT / MRI data, initially evaluate the image conversion effect, and adjust the model parameters; (2) Unsupervised training: Utilize the unsupervised feature of the cyclic adversarial generative network and adopt the generative adversarial loss and cyclic consistency loss for training. First, convert the B-ultrasound image into a CT / MRI-style image, and then convert it back from the CT / MRI-style image to the B-ultrasound-style image. (3) Model evaluation: Compare the generated CT / MRI-style image with the real CT / MRI image, evaluate the clarity of the image conversion, the edge effect, and the accuracy of the boundary between fat and muscle, and confirm whether the image quality meets the expectations. (4) Iterative optimization: When the quality of the image conversion does not meet the expectations, gradually adjust the learning rates of the generator and discriminator, and dynamically adjust the loss weights during the training process to balance the conversion effect of the image and the accuracy of the structure preservation.

[0016] Further, an animal body fat detection device executes the above animal body fat detection method.

[0017] The present invention provides a technical solution that can not only improve the quality of B-ultrasound images, enhance the boundary clarity, but also achieve high-precision fat thickness measurement. By using the cyclic adversarial generative network technology, the B-ultrasound image is converted into a CT or MRI image, effectively removing the speckle noise and artifacts in the B-ultrasound image, significantly enhancing the boundary clarity between the fat layer and the muscle layer, making the boundary between the fat layer and the muscle layer clearer, and greatly reducing the error of fat thickness measurement. At the same time, this solution should have a low equipment cost and a low usage threshold to meet the needs of users of different scales in the livestock industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 : Flow chart of the animal body fat measurement method of the present invention; Figure 2 : Flow chart for generating the animal body fat detection model of the present invention; Figure 3 : Principle block diagram of the body fat detection device of the present invention DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The technical solution of the present invention will be described in detail below with reference to the drawings and specific embodiments. It should be understood that the specific features in the embodiments of the present application and the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application. Without conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other.

[0020] The term "and / or" in this article is merely a description of the association relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this article generally represents an "or" relationship between the associated objects before and after.

[0021] The present invention utilizes the unsupervised image conversion ability of the cyclic adversarial generation network to convert B-ultrasound images into high-quality images with CT / MRI styles, enhancing the boundary clarity between the fat layer and the muscle layer while retaining the original information of the ultrasound images. Through this conversion, the clarity and contrast of the images are significantly improved, thereby enhancing the measurement accuracy of the body fat thickness.

[0022] Specifically, in the application of body fat detection, the cyclic adversarial generation network significantly improves the visualization effect of images by converting low-quality B-ultrasound images into clearer versions. Especially in the case of a thinner fat layer, the converted images can clearly present the hierarchical structure of the fat layer. In the subsequent image segmentation and thickness calculation processes, it can effectively reduce the measurement errors caused by image blurring, reduce the uncertainty in measurement, and thus improve the accuracy and reliability of body fat thickness measurement.

[0023] The present invention discloses a method for detecting animal body fat, a method for generating an animal body fat detection model, and a device. By applying the cyclic adversarial generation network technology in animal body fat measurement, the quality of B-ultrasound images is improved and accurate body fat measurement is achieved.

[0024] The method for detecting animal body fat provided by the present invention specifically includes the following steps: Step 1: Collect animal body fat images through the B-ultrasound probe of the body fat detection device and perform data preprocessing to achieve data standardization.

[0025] Step 2: Use the trained animal body fat detection model based on the cyclic adversarial generation network to convert the real-time collected B-ultrasound images into CT / MRI-style images and perform noise reduction and enhancement processing. The converted CT / MRI-style images have higher resolution and clearer boundaries between the fat layer and the muscle layer.

[0026] Step 3: The animal body fat detection model obtains clearly defined scan lines based on the converted CT / MRI-style images. These scan lines can accurately represent the demarcation points between the fat layer and the muscle layer, thus ensuring the efficiency and accuracy of the subsequent segmentation process.

[0027] Step 4: Identify the upper and lower edges of the fat layer based on the demarcation point information on the scan lines and calculate the number of pixels of the upper and lower edges of the fat layer.

[0028] Step 5: The body fat detection device calculates the actual fat thickness based on the number of pixels of the upper and lower edges of the fat layer and the conversion ratio of the detection device, and constructs a body fat index to reflect the overall body fat situation.

[0029] The steps for generating the animal body fat detection model provided by the present invention include: data preparation, model construction, model training, and model deployment.

[0030] I. In the data preparation step, mainly collect the target modality images and B-ultrasound image data to form a target image dataset, and preprocess the collected images and data to achieve data standardization.

[0031] 1. Target modality images: Collect CT or MRI images of some animals, especially the areas that clearly show the distribution of fat and muscle. Although the cycle adversarial generation network does not require paired data, paired data (B-ultrasound images and CT / MRI images) helps to improve the quality of image conversion.

[0032] 2. Collect B-ultrasound image data: Collect a large number of animal B-ultrasound images. These images should contain obvious structures of the fat and muscle layers, and cover animals of different breeds, ages, and genders as much as possible. By ensuring the diversity of the data, the generalization ability of the model is improved.

[0033] 3. Image data preprocessing: Uniform the size and grayscale of all images to ensure that they meet the input requirements of the model. The standardization operation also helps to accelerate the training convergence speed and improve the model effect.

[0034] II. In the model construction step, design and construct a cycle adversarial generation network model through three aspects: a generator, a discriminator, and a loss function.

[0035] 1. The generator design adopts a convolutional neural network based on residual blocks, which is used to convert B-ultrasound images into CT / MRI-style images, or convert CT / MRI-style images into B-ultrasound images. The generator is usually designed with 9 residual blocks to ensure the retention of the details of the generated images.

[0036] 2. The discriminator design adopts a PatchGAN discriminator. This structure judges the local authenticity of the image through small convolutional kernels, which can ensure the feature consistency of the generated image in the local area, thus improving the image conversion effect.

[0037] 3. The loss function design includes generative adversarial loss, cycle consistency loss, and identity loss. Through the design of these three aspects, it is ensured that the consistency and stability of the image style can be maintained during the image conversion process. Specifically: a. Generative adversarial loss: Through adversarial training, make the generated image approximate the style of the target modality to achieve image style conversion.

[0038] b. Cycle consistency loss: Ensure that after the B-ultrasound image undergoes two generations of conversion (that is, the B-ultrasound image is converted into a CT / MRI-style image, and then the CT / MRI-style image is converted into a B-ultrasound image), it can be consistent with the original B-ultrasound image, thus retaining the main structural features of fat and muscle.

[0039] c. Identity loss: Increase the loss of converting an image of the same modality to itself to ensure stable style during the conversion process.

[0040] III. During the model training process, the model is trained and debugged through the pre-training stage, unsupervised training, model evaluation, and iterative optimization.

[0041] 1. Pre-training stage: First, pre-train on a small-scale B-ultrasound and CT / MRI data to preliminarily evaluate the image conversion effect and adjust the model parameters.

[0042] 2. Unsupervised training: Utilize the unsupervised characteristics of the cyclic adversarial generative network and train using the generative adversarial loss and cyclic consistency loss. First, convert the B-ultrasound image to an image in the CT / MRI style, and then convert it back from the CT / MRI-style image to the B-ultrasound image.

[0043] 3. Model evaluation: By comparing the generated CT / MRI-style image with the real CT / MRI image, evaluate the clarity of the image conversion, the edge effect, and the accuracy of the fat and muscle demarcation, and confirm whether the image quality meets the expectations. The present invention uses quantitative indicators (such as PSNR, SSIM, etc.) to evaluate the image quality.

[0044] 4. Iterative optimization: When the quality of the image conversion does not meet the expectations, gradually adjust the learning rates of the generator and discriminator, and dynamically adjust the loss weights during the training process to balance the image conversion effect and the accuracy of structure retention.

[0045] IV. When the model meets the expectations after training and optimization, deploy an animal body fat detection model based on the cyclic adversarial generative network.

[0046] The animal body fat detection method of the present invention is implemented for body fat detection through a detection device deploying an animal body fat detection model based on the cyclic adversarial generative network. The detection device includes a B-ultrasound probe, a data acquisition module, an FPGA, a central processor, and a display module. The data acquisition module converts the echo signal of the B-ultrasound probe into a digital signal through a high-speed ADC (analog-to-digital converter) and transmits it to the FPGA. The FPGA delays and adds the multi-channel echo signals to generate high-resolution scan line data, and performs preprocessing operations such as denoising and gain compensation. After forming a B-ultrasound image through digital scan conversion, it is sent to the central processing unit. The animal body fat detection model deployed in the central processor converts the B-ultrasound image into an image in the CT / MRI style, generates an image with clear boundaries, then performs boundary detection to find the boundary information of the fat, calculates the fat thickness, and outputs the final calculation result and the clear image to the display module for the user to observe and analyze.

[0047] The present invention adopts the cyclic adversarial generation network technology, which can effectively remove speckle noise and artifacts in B-ultrasound images, significantly improve the boundary clarity of the fat layer and the muscle layer, and thus provide a more accurate basis for subsequent image segmentation and fat thickness measurement. The B-ultrasound image is converted into an image in the style of CT or MRI, which improves the visibility of the fat layer boundary, enhances the contrast and sense of hierarchy of the overall image, makes the demarcation between the fat layer and the muscle layer clearer, and greatly reduces the error of fat thickness measurement. Especially in the case of a thinner fat layer, the effect is particularly prominent. After image conversion, the measurement data has higher consistency and reliability, can accurately reflect the body fat condition of animals, and provides reliable data support for subsequent estimation of the body fat index.

[0048] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and deformations can be made, and these improvements and deformations should also be regarded as the protection scope of the present invention.

Claims

1. An animal body fat detection method based on a cyclic adversarial generation network model, characterized in that: Specific steps: Step 1: Collect B-ultrasound images of animal body fat and perform data preprocessing; Step 2: Use the trained animal body fat detection model based on the cyclic adversarial generation network to convert the collected B-ultrasound images into CT / MRI-style images; Step 3: Obtain scan lines with clear boundaries between the fat layer and the muscle layer according to the converted CT / MRI-style images; Step 4: Identify the upper and lower edges of the fat layer based on the demarcation point information on the scan line and calculate the number of pixels of the upper and lower edges of the fat layer; Step 5: Calculate the actual fat thickness according to the number of pixels of the upper and lower edges of the fat layer and the conversion ratio of the detection device.

2. A method for generating an animal body fat detection model according to claim 1, characterized in that Including: Data preparation: Mainly collect target modality images and B-ultrasound image data to form a target image dataset, and preprocess the collected images and data; Model construction: Build a cyclic adversarial generation network model through the design of a generator, a discriminator, and a loss function; Model training: Train and debug the model through the pre-training stage, unsupervised training, model evaluation, and iterative optimization; When the model meets the expectations after training and optimization, complete the model deployment.

3. A method for generating an animal body fat detection model according to claim 2, characterized in that The target modality images are CT or MRI images; the B-ultrasound image data should contain obvious structures of the fat and muscle layers.

4. A method for generating an animal body fat detection model according to claim 2, wherein In the data preparation step, the preprocessing is to perform size and grayscale processing on all images.

5. The method for generating an animal body fat detection model according to claim 2, wherein The generator design uses a convolutional neural network based on residual blocks.

6. A method for generating an animal body fat detection model according to claim 2, wherein The discriminator is designed as a PatchGAN discriminator.

7. A method for generating an animal body fat detection model according to claim 2, characterized in that, The loss function design includes generative adversarial loss, cycle consistency loss, and identity loss.

8. A method for generating an animal body fat detection model according to claim 7, characterized in that, The loss function design is specifically as follows: a. Generative adversarial loss: Through adversarial training, make the generated images approximate the style of the target modality to achieve image style conversion; b. Cycle consistency loss: Ensure that the B-ultrasound images are consistent with the original B-ultrasound images after two generative conversions; c. Identity loss: Increase the loss of converting images of the same modality into themselves to ensure that the style remains stable during the conversion process.

9. The method for generating an animal body fat detection model according to claim 2, wherein The model training steps are as follows: (1) Pre-training stage: First, perform pre-training on a small-scale B-ultrasound and CT / MRI data, initially evaluate the image conversion effect, and adjust the model parameters; (2) Unsupervised training: Utilize the unsupervised characteristics of the cyclic adversarial generation network and perform training using generative adversarial loss and cycle consistency loss. First, convert the B-ultrasound images into CT / MRI-style images, and then convert them back from the CT / MRI-style images to B-ultrasound-style images; (3) Model evaluation: Compare the generated CT / MRI-style images with the real CT / MRI images, evaluate the clarity, edge effect of the image conversion, and the accuracy of the fat and muscle demarcation, and confirm whether the image quality meets the expectations; (4) Iterative optimization: When the quality of the image conversion does not meet the expectations, gradually adjust the learning rates of the generator and the discriminator, and dynamically adjust the loss weights during the training process to balance the image conversion effect and the accuracy of structure retention.

10. An animal body fat detection device, characterized in that, Execute the animal body fat detection method described in claim 1.

Images