A pig intramuscular fat content prediction method based on multi-modal feature fusion and multi-layer perception

Abstract

The application discloses a pig intramuscular fat content prediction method based on multi-modal feature fusion and a multi-layer perception machine and belongs to the technical field of meat quality detection. The method comprises the following steps: collecting a B-ultrasonic image of a longissimus dorsi muscle of a modeling pig sample, individual growth state data and an actually measured value of intramuscular fat content, and forming an original sample data set; determining a target muscle region of the longissimus dorsi muscle based on the B-ultrasonic image of the longissimus dorsi muscle and extracting a tissue structure response feature vector; constructing an individual state correction factor based on the individual growth state data, correcting the tissue structure response feature vector and forming an intramuscular fat tissue state representation; establishing a tissue state constraint prediction model by using the intramuscular fat tissue state representation and the actually measured value of the intramuscular fat content; and predicting the intramuscular fat content of a to-be-detected pig sample by using the tissue state constraint prediction model.

Description

Technical Field

[0001] This invention belongs to the field of meat quality testing technology, specifically relating to a method for predicting intramuscular fat content in pigs based on multimodal feature fusion and multilayer perceptron. Background Technology

[0002] Intramuscular fat content in pigs is a crucial indicator of pork quality, affecting its flavor, tenderness, and overall palatability. Current methods for detecting intramuscular fat content typically employ in vivo testing. This involves acquiring ultrasound images of the longissimus dorsi muscle region of the pig and combining these images with individual phenotypic parameters such as pig weight, age, backfat thickness, and eye muscle area. Regression models, machine learning models, or neural network models are then used to establish a predictive relationship between the input data and intramuscular fat content.

[0003] Existing methods for predicting intramuscular fat content in pigs based on ultrasound images typically extract image features such as grayscale, texture, and gradient in the longissimus dorsi muscle region and fuse these features with individual phenotypic parameters to achieve non-destructive prediction of intramuscular fat content. However, changes in grayscale, texture, and echo distribution in ultrasound images primarily reflect the ultrasound tissue response state and do not have a direct correspondence with the intramuscular fat deposition state. Therefore, simply relying on the mapping relationship between image features and predicted values ​​is insufficient to fully express changes in tissue state.

[0004] Meanwhile, differences in age, weight, backfat thickness, and eye muscle area exist among different pigs. Individual growth status changes affect ultrasound echo intensity, tissue scale, and the display of local structures, resulting in differences in image features between different pig samples. Existing prediction models mainly rely on the statistical relationship between input features and intramuscular fat content, lacking constraints on the patterns of tissue state changes. This makes the prediction results susceptible to individual sample differences and variations in image acquisition conditions.

[0005] Therefore, how to establish a predictive characterization that reflects the state of intramuscular fat deposition based on ultrasound image information and individual growth status information, and improve the stability of the intramuscular fat content prediction process, is an urgent problem to be solved. Summary of the Invention

[0006] This invention overcomes the shortcomings of the prior art and provides a method for predicting intramuscular fat content in pigs based on multimodal feature fusion and multilayer perceptron.

[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for predicting intramuscular fat content in pigs based on multimodal feature fusion and multilayer perceptron, comprising the following steps:

[0008] S1: Collect ultrasound images of the longissimus dorsi muscle, individual growth status data, and measured values ​​of intramuscular fat content of modeling pig samples, and form the original sample dataset based on the sample identifiers;

[0009] S2: Determine the target muscle region of the longissimus dorsi muscle based on the original sample dataset, and extract the tissue structure response feature vector from the target muscle region of the longissimus dorsi muscle.

[0010] S3: Construct individual state correction factors based on individual growth state data, and use individual state correction factors to correct tissue structure response feature vectors to form a characterization of intramuscular adipose tissue status.

[0011] S4: Establish a tissue state-constrained prediction model using intramuscular adipose tissue state characterization and measured values ​​of intramuscular adipose content;

[0012] S5: Process the data of the pig samples to be tested according to the data processing method corresponding to the tissue state constraint prediction model, and output the predicted value of intramuscular fat content of the pig samples to be tested.

[0013] Preferably, in step S1, a data correlation is established between the longissimus dorsi muscle ultrasound image, individual growth status data, and measured intramuscular fat content based on the sample identifier. The individual growth status data includes age, weight, backfat thickness, and eye muscle area.

[0014] Preferably, in step S1, the number of ultrasound images of the longissimus dorsi muscle is 1 to 5 frames, the grayscale bit depth of the ultrasound image of the longissimus dorsi muscle is 8 to 16 bits, and the pixel size of the ultrasound image of the longissimus dorsi muscle is not less than 256×256 pixels.

[0015] Preferably, step S2 includes performing grayscale normalization and size unification processing on the ultrasound image of the longissimus dorsi muscle, determining the target muscle region of the longissimus dorsi muscle based on the boundary of the longissimus dorsi muscle region, and dividing the target muscle region of the longissimus dorsi muscle into multiple local tissue units.

[0016] Preferably, the local tissue units adopt rectangular grids, overlapping sliding windows, or strip regions divided along the direction of muscle fibers, and the number of local tissue units is 4 to 64. Based on the local tissue units, tissue uniformity features, tissue interface response features, and local echo gradient features are extracted respectively, and tissue structure response feature vectors are formed according to the order of local tissue units.

[0017] Preferably, step S3 includes standardizing the individual growth state data to form an individual growth state vector, and calculating an individual state correction factor based on the individual growth state vector that is consistent with the dimension of the tissue structure response feature vector.

[0018] Preferably, the individual state correction factor is obtained through an individual state correction weight matrix, an individual state correction bias vector, and a nonlinear transformation method, and the correction magnitude coefficient of the individual state correction factor is 0.05 to 0.50.

[0019] Preferably, in step S4, the tissue state constraint prediction model includes a tissue state input layer, a multimodal state fusion layer, a multilayer perceptron prediction layer, and a linear output layer. The tissue state constraint prediction model takes the intramuscular adipose tissue state representation as input and the intramuscular fat content prediction value as output.

[0020] Preferably, the tissue state constraint prediction model determines the model parameters during training based on the prediction error between the predicted value and the measured value of intramuscular fat content, the prediction difference between different ultrasound image frames of the longissimus dorsi muscle under the same sample identifier, and the correction magnitude of the individual state correction factor.

[0021] Preferably, after the tissue state constraint prediction model is trained, the intramuscular fat content of the pig samples to be tested is predicted using the image processing method, the method for determining the target muscle region of the longissimus dorsi muscle, the method for dividing local tissue units, the method for calculating features, and the individual state correction parameters determined during the training process.

[0022] This invention addresses the deficiencies in the prior art and has the following beneficial effects:

[0023] This invention extracts tissue structure response feature vectors from ultrasound images of the longissimus dorsi muscle and constructs individual state correction factors by combining individual growth status data such as age, weight, backfat thickness, and eye muscle area. These factors correct the tissue structure response feature vectors to form a representation of intramuscular fat tissue state. Then, a tissue state constraint prediction model containing a multimodal state fusion layer, a multilayer perceptron prediction layer, and a linear output layer is used for prediction processing. This allows the tissue structure information of the longissimus dorsi muscle to be fused with the individual growth status information of the pig under a unified state representation. The state matching relationship during the prediction process is maintained through model training constraints, ensuring that the output predicted value of intramuscular fat content corresponds to the changes in tissue state.

[0024] By dividing the target muscle region of the longissimus dorsi into local tissue units and extracting tissue homogeneity features, tissue interface response features, and local echo gradient features, a tissue structure response feature vector is formed. This allows for a structured expression of the local echo distribution, tissue homogeneity changes, and tissue interface changes in the longissimus dorsi ultrasound image. Furthermore, this invention utilizes age, body weight, backfat thickness, and eye muscle area to construct an individual state correction factor. This factor is then used to perform dimension-wise correction on the tissue structure response feature vector, enabling it to be adjusted in conjunction with the individual growth status of the pig, thus forming a representation of intramuscular adipose tissue status and reducing the impact of individual pig state differences on the consistency of tissue state expression.

[0025] By inputting the intramuscular adipose tissue state representation into a tissue state-constrained prediction model, and utilizing a multimodal state fusion layer, a multilayer perceptron prediction layer, and a linear output layer for prediction processing, the state-corrected tissue information can complete multi-level feature mapping and intramuscular fat content prediction. Simultaneously, this invention introduces prediction error constraints, intra-sample consistency constraints, and correction amplitude constraints during the training process of the tissue state-constrained prediction model. This ensures that the model parameter adjustment process is jointly limited by prediction bias, prediction differences between different ultrasound image frames of the longissimus dorsi muscle of the same sample, and the individual state correction amplitude, thereby reducing the fluctuation of prediction results and maintaining the stability of intramuscular fat content prediction results.

[0026] In this invention, the tissue structure response feature extraction process transforms tissue change information in the longissimus dorsi muscle ultrasound image into a processable tissue structure response feature vector. The individual state correction process uses pig growth status data to conditionally match the tissue structure response feature vector, forming an intramuscular adipose tissue state representation that reflects the tissue state. The tissue state constraint prediction model further fuses and constrains the intramuscular adipose tissue state representation, maintaining a correspondence between image information, individual state information, and prediction results. This reduces the impact of individual differences and data fluctuations on the prediction results during the intramuscular fat content prediction process. Attached Figure Description

[0027] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0028] Figure 1 This is a step-by-step diagram of a method for predicting intramuscular fat content in pigs based on multimodal feature fusion and multilayer perceptron. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Method implementation:

[0031] This embodiment provides a method for predicting intramuscular fat content in pigs based on multimodal feature fusion and multilayer perceptron, used to predict intramuscular fat content based on in vivo longissimus dorsi muscle ultrasound images and individual pig growth status data. Addressing the limitations of existing technologies where surface features such as grayscale and texture of ultrasound images cannot directly characterize intramuscular fat deposition, and the bias in ultrasound tissue characterization caused by differences in age, weight, backfat thickness, and eye muscle area among different pigs, this embodiment constructs a data processing procedure consisting of tissue structure response feature extraction, individual state correction, and tissue state constraint prediction.

[0032] This implementation extracts tissue structure response feature vectors reflecting tissue homogeneity, tissue interface changes, and local echo gradients from ultrasound images of the longissimus dorsi muscle. Individual state correction factors are constructed using individual pig growth data to correct the tissue structure response feature vectors, forming a representation of intramuscular adipose tissue state. Based on this, using the intramuscular adipose tissue state representation as prediction input and measured intramuscular fat content as training labels, a tissue state constraint prediction model is established, comprising a multimodal state fusion layer, a multilayer perceptron prediction layer, and a linear output layer. Prediction parameters are determined through model training, and the trained tissue state constraint prediction model is used to predict intramuscular fat content in the pig samples under test.

[0033] like Figure 1 As shown, a method for predicting intramuscular fat content in pigs based on multimodal feature fusion and multilayer perceptron includes the following steps:

[0034] S1: Collect ultrasound images of the longissimus dorsi muscle, individual growth status data, and measured values ​​of intramuscular fat content of modeling pig samples, and form the original sample dataset based on the sample identifiers;

[0035] S2: Determine the target muscle region of the longissimus dorsi muscle based on the original sample dataset, and extract the tissue structure response feature vector from the target muscle region of the longissimus dorsi muscle.

[0036] S3: Construct individual state correction factors based on individual growth state data, and use individual state correction factors to correct tissue structure response feature vectors to form a characterization of intramuscular adipose tissue status.

[0037] S4: Establish a tissue state-constrained prediction model using intramuscular adipose tissue state characterization and measured values ​​of intramuscular adipose content;

[0038] S5: Process the data of the pig samples to be tested according to the data processing method corresponding to the tissue state constraint prediction model, and output the predicted value of intramuscular fat content of the pig samples to be tested.

[0039] The following description, in conjunction with the specific implementation process of each step, further explains the method so that those skilled in the art can clearly understand the interrelationship of each step in implementation and the resulting technical results.

[0040] S1 is used to collect ultrasound images of the longissimus dorsi muscle, individual growth status data, and measured values ​​of intramuscular fat content of modeling pig samples, and to establish the correspondence between various types of data based on sample identification.

[0041] In this step, modeling pig samples are used as the data collection object. The original sample dataset is formed in the following order: sample identification, acquisition of longissimus dorsi muscle ultrasound images, recording of individual growth status data, and association of measured intramuscular fat content. Sample identification is used to maintain the correspondence between image data, individual status data, and label data of the same modeling pig sample; longissimus dorsi muscle ultrasound images are used to record the ultrasound echo distribution of the longissimus dorsi muscle region of the modeling pig sample; individual growth status data are used to record the growth stage, peripheral fat deposition status, and muscle development status of the modeling pig sample; and measured intramuscular fat content is used to form training labels corresponding to the modeling pig sample.

[0042] A unique sample identifier is established for each modeling pig sample, and ultrasound images of the longissimus dorsi muscle region of the modeling pig sample are acquired according to the sample identifier. The sample identifier can be numbers, letters, or a combination of numbers and letters, and the sample identifier length is 4 to 32 characters. When acquiring multiple frames of longissimus dorsi muscle ultrasound images of the same modeling pig sample, the multiple frames of longissimus dorsi muscle ultrasound images of the same modeling pig sample are bound to the same sample identifier, so that different image frames of the same modeling pig sample maintain consistent attribution at the data recording level. The acquisition position of the longissimus dorsi muscle ultrasound image is a live detection position that can display the longissimus dorsi muscle tissue structure. After the longissimus dorsi muscle ultrasound image acquisition is completed, image frames that reflect the echo distribution, tissue homogeneity, and tissue interface changes of the longissimus dorsi muscle are retained.

[0043] During the acquisition of ultrasound images of the longissimus dorsi muscle, 1 to 5 frames of ultrasound images of the longissimus dorsi muscle were acquired for each modeling pig sample. The grayscale bit depth of the longissimus dorsi muscle ultrasound images ranged from 8 to 16 bits. The pixel size of the longissimus dorsi muscle ultrasound images was no less than 256 × 256 pixels. The longissimus dorsi muscle ultrasound images were saved in PNG, TIFF, DICOM, or other image formats that could retain grayscale information. For the same batch of modeling pig samples, the acquisition location, image saving format, and grayscale information retention method of the longissimus dorsi muscle ultrasound images remained consistent.

[0044] After establishing the sample identifiers and acquiring ultrasound images of the longissimus dorsi muscle for the modeling pigs, individual growth status data were recorded according to the sample identifiers. Individual growth status data included age in days, weight, backfat thickness, and eye muscle area. Age in days and weight were used to express the growth stage of the modeling pigs, backfat thickness was used to express the peripheral fat deposition status, and eye muscle area was used to express the muscle development status. After recording the individual growth status data, a data association was established between the individual growth status data and the longissimus dorsi muscle ultrasound images under the same sample identifier, forming the modeling input data for the same modeling pig sample.

[0045] In recording individual growth data, age in days was measured in days (d), with a recording precision of 1 day; weight was measured in kilograms (kg), with a recording precision of 0.1 kg to 1 kg; backfat thickness was measured in millimeters (mm), with a recording precision of 0.1 mm to 1 mm; and eye muscle area was measured in centimeters. 2 The recording accuracy is 0.1cm. 2 Up to 1cm 2 Each modeling pig sample must include at least three data points: age in days, weight, backfat thickness, and eye muscle area. For the same batch of modeling pig samples, the recorded data items and units must be consistent.

[0046] The measured intramuscular fat content of the modeling pig samples is obtained based on the same sample identifier, and this measured value is used as a training label and bound to the same sample identifier. The measured intramuscular fat content, along with the longissimus dorsi muscle ultrasound image and individual growth status data, belong to the same modeling pig sample. After binding through the sample identifier, each modeling pig sample forms a single original sample data entry, represented as follows: ,in, Indicates the first Sample identifiers for a single modeling pig sample. Indicates the first Ultrasound images of the longissimus dorsi muscle in a modeling pig sample. Indicates the first Individual growth status data of a sample of modeling pigs. Indicates the first Measured intramuscular fat content of a sample of model pigs.

[0047] During the recording of measured intramuscular fat content, the unit of measurement is %, and the recording accuracy is 0.01% to 0.1%. The number of pig samples used for modeling should be no less than 50, preferably no less than 100. Each sample record in the original sample dataset includes a sample identifier, at least one frame of ultrasound image of the longissimus dorsi muscle, a set of individual growth status data, and a measured value of intramuscular fat content. Multiple original sample data are aggregated according to the sample identifier to form a dataset used for modeling the pig intramuscular fat content prediction method.

[0048] S2 is used to determine the target muscle region of the longissimus dorsi muscle using ultrasound images of the longissimus dorsi muscle in the original sample dataset, and to extract the tissue structure response feature vector from the target muscle region of the longissimus dorsi muscle.

[0049] In this step, the longissimus dorsi ultrasound images in the original sample dataset are subjected to grayscale normalization and size unification. The target muscle region of the longissimus dorsi is determined in the unified longissimus dorsi ultrasound images, and the target muscle region of the longissimus dorsi is used as the object of tissue structure response feature extraction.

[0050] After reading the longissimus dorsi muscle ultrasound images from the original sample dataset, grayscale normalization was performed on the images. Grayscale normalization converts pixel grayscale values ​​from different image acquisition conditions to a uniform grayscale range, with the normalized pixel grayscale range being 0 to 1. After grayscale normalization, the longissimus dorsi muscle ultrasound images were then resized to ensure consistent or comparable image scales for ultrasound images of the longissimus dorsi muscle corresponding to different modeled pig samples. The image size of the target muscle region of the longissimus dorsi muscle ranged from 128×128 pixels to 512×512 pixels.

[0051] After grayscale normalization and size unification, the target muscle region of the longissimus dorsi muscle is determined based on its boundary. The target muscle region can be separated from the entire longissimus dorsi ultrasound image by cropping, or by using a mask to retain the longissimus dorsi muscle region within the entire ultrasound image while masking non-longissimus dorsi muscle regions. The target muscle region occupies 10% to 80% of the area of ​​the entire longissimus dorsi ultrasound image. The target muscle region is determined through manual annotation, semi-automatic segmentation, or automatic segmentation; a consistent region determination method is used for pig samples from the same batch of modeling pigs.

[0052] The target muscle region of the longissimus dorsi muscle is divided into multiple local tissue units. These local tissue units are used to represent the echogenic changes of local tissues within the longissimus dorsi muscle. Local tissue units can be rectangular grids, overlapping sliding windows, or strip regions divided along the muscle fiber direction. The number of local tissue units ranges from 4 to 64; the size of the rectangular grid ranges from 16×16 pixels to 64×64 pixels; when using overlapping sliding windows, the overlap rate ranges from 0% to 50%. Within the same modeled pig sample, the local tissue units are arranged according to a preset spatial order, and different modeled pig samples use the same local tissue unit division method and arrangement order.

[0053] Based on the local tissue units obtained from the division, the tissue uniformity characteristics of each local tissue unit are calculated. For the th The first modeling pig sample The formula for calculating the uniformity characteristic of a local tissue unit is as follows: ,in, Indicates the first The modeling pig sample number The average gray value of a local tissue unit. Indicates the first The modeling pig sample number The standard deviation of gray level in a local tissue unit. This represents a constant used to prevent the denominator from being zero. Indicates the first The modeling pig sample number The tissue homogeneity characteristics of a local tissue unit. (Constant) for to .

[0054] After calculating the tissue homogeneity characteristics, the tissue interface response characteristics are calculated for each local tissue unit. The tissue interface response characteristics are used to express the density of gray-level abrupt change boundaries within a local tissue unit. The formula for calculating the tissue interface response characteristics is as follows: ,in, Indicates the first The modeling pig sample number Organizational interface response characteristics of individual organizational units Indicates the first The total number of pixels in a local organizational unit. Indicates the first A collection of pixels within a local organizational unit Indicates the pixel position within a local tissue unit. Indicates pixel position grayscale gradient magnitude at that location Indicates the gradient threshold. This is an indicator function that takes the value 1 when the condition is true and 0 when the condition is false. Gradient threshold. The gradient amplitude is 1.0 to 2.5 times the mean gradient amplitude of the target muscle region of the same longissimus dorsi muscle. The gradient calculation operator is the Sobel operator, Prewitt operator, or first-order difference operator. Tissue interface response characteristics. The value range is from 0 to 1.

[0055] Local echo gradient features are calculated based on the gray-level gradient amplitude within the same local tissue unit. These local echo gradient features are used to express the intensity of gray-level changes within a local tissue unit. The formula for calculating the local echo gradient features is as follows: ,in, Indicates the first The modeling pig sample number Local echo gradient characteristics of individual tissue units Indicates the first The total number of pixels in a local organizational unit. Indicates the first A collection of pixels within a local organizational unit Indicates the pixel position within a local tissue unit. Indicates pixel position The grayscale gradient magnitude at a given location. Local echo gradient characteristics. The value range is determined based on the grayscale normalization range and the size of the local tissue unit.

[0056] The tissue homogeneity features, tissue interface response features, and local echo gradient features of all local tissue units in each modeled pig sample are summarized in order of local tissue unit to form a tissue structure response feature vector. .in, Indicates the first The tissue structure response feature vector of a modeling pig sample. Indicates the first A set of tissue homogeneity features from a sample of modeled pigs. Indicates the first A set of tissue interface response features for a modeling pig sample. Indicates the first A set of local echo gradient features for a modeling pig sample. Tissue structure response feature vector. The dimensions range from 12 to 192, and the organizational structure response feature vector The local tissue units exhibit consistent order in terms of tissue homogeneity characteristics, tissue interface response characteristics, and local echo gradient characteristics.

[0057] When a single modeling pig sample has multiple frames of longissimus dorsi muscle ultrasound images, each frame of the longissimus dorsi muscle ultrasound image from the same modeling pig sample undergoes separate processing including grayscale normalization, size unification, determination of the target muscle region of the longissimus dorsi muscle, local tissue unit segmentation, and feature calculation. The tissue structure response feature vector corresponding to the multiple frames of longissimus dorsi muscle ultrasound images can be obtained by taking the mean, weighted mean, or retaining each frame as an intra-sample feature set for the same modeling pig sample. When using the mean or weighted mean, a single tissue structure response feature vector corresponding to the modeling pig sample is formed; when using the frame-by-frame retention method, multiple tissue structure response feature vectors under the same sample identifier are formed.

[0058] S3 is used to construct individual state correction factors by processing organizational structure response feature vectors and individual growth state data, and to perform dimension-by-dimensional correction of organizational structure response feature vectors using individual state correction factors.

[0059] In this step, the individual growth status data of the modeling pig samples are first standardized to form an individual growth status vector; then, an individual status correction factor with the same dimension as the tissue structure response feature vector is calculated based on the individual growth status vector; then, the individual status correction factor is used to adjust the tissue structure response feature vector dimension by dimension to form a representation of intramuscular adipose tissue status.

[0060] The age, weight, backfat thickness, and eye muscle area of ​​the modeling pig samples were standardized to form individual growth state vectors. For the first... The modeling pig sample number The standardized value calculation formula for the growth status data of each individual is as follows: ,in, Indicates the first The modeling pig sample number The original values ​​of individual growth status data Indicates the first training sample The mean of the growth status data of each individual. Indicates the first training sample Standard deviation of individual growth status data Indicates the first The modeling pig sample number Standardized values ​​of individual growth status data. After standardizing the growth status data of each individual, the first... The individual growth state vector of each modeling pig sample is represented as follows: .

[0061] In the standardization of individual growth status data, the data includes age in days, weight, backfat thickness, and eye muscle area. Individual growth status vector. The dimension is 3 to 4. Mean and standard deviation Calculations are based solely on the training samples. The number of missing individual growth status data points does not exceed 25% of the total number of individual growth status data points. When missing individual growth status data exists, missing values ​​are handled by imputing the mean or median of the same batch of samples, or by deleting severely missing samples. The same missing value handling method and the same standardization calculation method are used for the same batch of modeling pig samples.

[0062] Based on individual growth state vector Calculate individual state correction factor The individual state correction factor is used to scale the tissue structure response feature vector. The formula for calculating the individual state correction factor is: ,in, Indicates the first Individual state correction factors for a sample of modeling pigs. Indicates the correction amplitude coefficient. This represents the individual state correction weight matrix. Indicates the first The individual growth state vector of a modeling pig sample. This represents the individual state correction bias vector. Let represent the hyperbolic tangent function. The hyperbolic tangent function is related to... The calculation results are subjected to a nonlinear transformation to adjust the individual state correction factor. Based on the correction amplitude coefficient Within a defined range, a dimension-by-dimensional correction quantity is formed.

[0063] In the calculation of the individual state correction factor, the correction amplitude coefficient The range is 0.05 to 0.50. Individual state correction factor. The range of values ​​for each dimension is to Individual state correction factor Dimensions and organizational structure response feature vectors The dimensions are consistent. Individual state correction weight matrix The number of rows is equal to the organizational structure response feature vector. The dimension of the individual state correction weight matrix The number of columns is equal to the individual growth state vector. The dimension of the individual state correction weight matrix. and individual state correction bias vector The parameters were obtained during the determination of the organizational state constraint prediction model.

[0064] Using individual state correction factors Organizational structure response feature vector Dimensional correction was performed to form a characterization of intramuscular adipose tissue. Dimensional correction is achieved using element-wise multiplication. The formula for characterizing the state of intramuscular adipose tissue is as follows: ,in, Indicates the first Characterization of intramuscular adipose tissue status in a modeling pig sample. Indicates the first The tissue structure response feature vector of a modeling pig sample. Indicates the first Individual state correction factors for a sample of modeling pigs. This represents element-wise multiplication. Element-wise multiplication divides the organizational structure response feature vector according to the same dimensional position. Each eigenvalue in the equation is related to the individual state correction factor. Multiplying the corresponding correction values ​​in the original text yields the corrected intramuscular adipose tissue state characterization. .

[0065] During the dimension-by-dimensional correction process, the tissue structure response feature vector and individual state correction factor Having the same dimensions. Characterization of intramuscular adipose tissue status. Dimensions and organizational structure response feature vectors Consistent dimensions, characterization of intramuscular adipose tissue state The dimensions range from 12 to 192. The training samples, validation samples, test samples, and pig samples to be tested all use the same individual state correction factor calculation method and the same dimension-by-dimensional correction method.

[0066] S4 is used to establish a tissue state constraint prediction model by processing the intramuscular adipose tissue state characterization and the measured value of intramuscular fat content, and the parameters of the tissue state constraint prediction model are determined by training set samples and validation set samples.

[0067] In this step, the modeling pig samples are first divided into training, validation, and test sets according to sample labels. Then, a tissue state constraint prediction model is established, which includes a tissue state input layer, a multimodal state fusion layer, a multilayer perceptron prediction layer, and a linear output layer. The tissue state constraint prediction model takes the intramuscular adipose tissue state representation as input and the predicted intramuscular fat content as output. During training, the prediction error is calculated based on the measured intramuscular fat content, and the model parameters are determined by combining the intrasample consistency constraint term and the correction amplitude constraint term.

[0068] The modeling pig samples were divided into training, validation, and test sets. The training set was used to update the parameters of the tissue state constraint prediction model, the validation set was used to select the tissue state constraint prediction model configuration, and the test set was used for performance validation after the model configuration was fixed. Sample identifiers were used as the partitioning unit. When the same modeling pig sample had multiple frames of longissimus dorsi muscle ultrasound images, these images under the same sample identifier were assigned to the same subset to avoid different longissimus dorsi muscle ultrasound images from the same modeling pig sample appearing simultaneously in the training, validation, or test sets. The training set comprised 60% to 80% of the data, the validation set 10% to 20%, and the test set 10% to 20%. The distribution of measured intramuscular fat content in the training, validation, and test sets was kept consistent using random or stratified random partitioning.

[0069] The input structure of the tissue state-constrained prediction model is established based on the partitioned training set samples. The tissue state-constrained prediction model receives the intramuscular adipose tissue state representation. Intramuscular adipose tissue state characterization as input data Organizational response feature vector and individual state correction factor Formed through dimensional fusion. Organizational structure response feature vector. Derived from ultrasound images of the longissimus dorsi muscle, individual state correction factor The data originates from individual growth state data modalities. The tissue state-constrained prediction model fuses tissue structure response information from image modalities and correction information from individual growth state data modalities through a multimodal state fusion layer to form a predictive input representation. .

[0070] The calculation formula for the multimodal state fusion layer is as follows: ,in, Indicates the first The predicted input representation for a sample of modeled pigs, This represents the weight matrix of the multimodal state fusion layer. Indicates the first Characterization of intramuscular adipose tissue status in a modeling pig sample. This represents the bias vector of the multimodal state fusion layer. This represents a nonlinear activation function. The input dimension of the tissue state-constrained prediction model is equal to the representation of intramuscular adipose tissue state. The dimension. The number of multimodal state fusion layers is 1 to 2, and the predicted input representation. The dimensions range from 16 to 256, and the activation function is non-linear. For ReLU, LeakyReLU, or tanh, the random inactivation rate of the multimodal state fusion layer is 0 to 0.30.

[0071] A multilayer perceptron prediction layer and a linear output layer are built after the multimodal state fusion layer. The multilayer perceptron prediction layer represents the predicted input. A nonlinear mapping is performed, and the linear output layer generates predicted intramuscular fat content values ​​based on the output of the multilayer perceptron prediction layer. The prediction calculation formula for the tissue state-constrained prediction model is as follows: ,in, Indicates the first Predicted intramuscular fat content of a modeling pig sample The parameter is Organizational state constraint prediction model, This includes parameters for the individual state correction layer, the multimodal state fusion layer, the multilayer perceptron prediction layer, and the linear output layer. Indicates the first The intramuscular adipose tissue state of a modeling pig sample was characterized. The multilayer perceptron prediction layer consisted of 2 to 4 layers, with each layer containing 32 to 256 neurons. The activation function for each prediction layer was ReLU, LeakyReLU, or tanh. The linear output layer had one neuron and a 1-dimensional output dimension. The unit of the tissue state constraint prediction model output value was %.

[0072] The parameters of the tissue state-constrained prediction model are established using training set samples. The training loss function includes a prediction error term, an intra-sample consistency constraint term, and a correction magnitude constraint term. The prediction error term is used to calculate the difference between the predicted and measured intramuscular fat content of the training samples; the intra-sample consistency constraint term is used to calculate the prediction difference between different ultrasound images of the longissimus dorsi muscle under the same sample identifier; the correction magnitude constraint term is used to calculate the deviation of the individual state correction factor from a vector where each dimension is 1. The training loss function is expressed as follows: ,in, This represents the training loss of the organizational state-constrained prediction model. This indicates the number of samples in the training batch. Indicates the first Predicted intramuscular fat content for each training sample. Indicates the first Measured intramuscular fat content of each training sample This represents the set of sample pairs formed by different ultrasound image frames of the longissimus dorsi muscle under the same sample identifier. Indicates the number of sample pairs. and These represent the predicted intramuscular fat content values ​​corresponding to two ultrasound images of the longissimus dorsi muscle under the same sample identifier. Indicates the first Individual state correction factor for each training sample This represents a vector in which all dimensions are 1. Indicates the weight of the in-sample consistency constraint. Indicates the correction magnitude constraint weight. This represents the square operation of the 2-norm.

[0073] During the training of the organizational state constraint prediction model, the batch size ranged from 16 to 64, the learning rate from 0.0001 to 0.01, the number of training epochs from 50 to 300, and the optimizer used was Adam, RMSProp, or a stochastic gradient descent optimizer. Intra-sample consistency constraint weights were used. The value is between 0 and 1, representing the correction amplitude constraint weight. The value ranges from 0.0001 to 0.1. When the modeling pig sample lacks multi-frame ultrasound images of the longissimus dorsi muscle, the intra-sample consistency constraint weight... Set to 0. During parameter training, the optimizer adjusts the parameters according to the training loss function. Organizational state constrained prediction model parameters The update process involves updating parameters of the individual state correction layer, the multimodal state fusion layer, the multilayer perceptron prediction layer, and the linear output layer.

[0074] After each training epoch or at a preset interval, the validation error is calculated using the validation set samples. The formula for calculating the validation error is: ,in, Indicates the verification error. Indicates the number of samples in the validation set. Indicates the first Predicted intramuscular fat content values ​​for each validation set sample. Indicates the first The measured intramuscular fat content of each validation set sample was used. The number of multimodal state fusion layers, the number of multilayer perceptron prediction layers, the number of neurons, the learning rate, the correction amplitude coefficient, and the constraint weights were adjusted based on the validation error. The validation error calculation interval was 1 to 5 training epochs, and the early stopping wait epochs were 10 to 30 epochs. The number of multimodal state fusion layers was adjusted to 1 to 2 layers, the number of multilayer perceptron prediction layers to 2 to 4 layers, the number of neurons in each multilayer perceptron prediction layer to 32 to 256, the learning rate to 0.0001 to 0.01, and the correction amplitude coefficient... The adjustment range is 0.05 to 0.50.

[0075] If the validation error does not decrease within several consecutive training epochs, parameter updates are stopped, and the model parameters with smaller validation errors are selected as the parameters for the tissue state constraint prediction model. Organizational state constraint prediction model parameters This includes parameters for the individual state correction layer, parameters for the multimodal state fusion layer, parameters for the multilayer perceptron prediction layer, and parameters for the linear output layer.

[0076] S5 is used to process the longissimus dorsi muscle ultrasound image and individual growth status data of the pig samples to be tested, and to form the intramuscular fat tissue status characterization of the pig samples to be tested according to the data processing method corresponding to the tissue status constraint prediction model, and output the predicted value of intramuscular fat content of the pig samples to be tested.

[0077] In this step, firstly, sample identification is established for the pig samples to be tested, and ultrasound images of the longissimus dorsi muscle of the pig samples are acquired and individual growth status data of the pig samples are recorded. Then, according to the image processing method, the method for determining the target muscle region of the longissimus dorsi muscle, the method for dividing local tissue units, and the feature calculation method determined in the process of establishing the tissue state constraint prediction model, a response feature vector of the tissue structure to be tested is formed. Then, using the mean, standard deviation, individual state correction weight matrix, and individual state correction bias vector determined in the process of establishing the tissue state constraint prediction model, the individual state correction factor of the individual to be tested is calculated. Finally, the individual state correction factor of the individual to be tested is corrected dimension by dimension with the response feature vector of the tissue structure to be tested to form the intramuscular adipose tissue state characterization of the individual to be tested, and the intramuscular adipose tissue state characterization of the individual to be tested is input into the tissue state constraint prediction model to obtain the predicted value of intramuscular fat content of the pig samples to be tested.

[0078] Sample identifiers were established for the pig samples to be tested. Ultrasound images of the longissimus dorsi muscle of the pig samples were acquired, and the age, weight, backfat thickness, and oculomotor muscle area of ​​the pig samples were recorded. The sample identifiers of the pig samples were used to link the longissimus dorsi muscle ultrasound images, individual growth status data, and prediction results of the same pig sample. Measured intramuscular fat content values ​​were not required for the pig samples to be tested. The number of frames in the longissimus dorsi muscle ultrasound images of the pig samples to be tested ranged from 1 to 5, and the image size of the target muscle region of the longissimus dorsi muscle in the pig samples to be tested was consistent with the image size of the target muscle region of the longissimus dorsi muscle in the modeled pig samples.

[0079] The longissimus dorsi muscle ultrasound images of the pig samples were subjected to grayscale normalization, size standardization, and the determination of the target muscle region. The grayscale range of the normalized pixels was 0 to 1. The method for determining the target muscle region was consistent with the process of establishing the tissue state constraint prediction model; the target muscle region could be obtained through cropping or preserved through masking. The proportion of the target muscle region area of ​​the longissimus dorsi muscle in the total area of ​​the longissimus dorsi ultrasound image of the pig samples was obtained according to the region determination rules used in the process of establishing the tissue state constraint prediction model.

[0080] Based on the target muscle region of the longissimus dorsi muscle in the test pig samples, the region was divided into multiple local tissue units according to the local tissue unit division method determined during the establishment of the tissue state constraint prediction model. The number and arrangement order of local tissue units in the test pig samples were consistent with those in the modeling pig samples. The number of local tissue units ranged from 4 to 64; the local tissue units were rectangular grids, overlapping sliding windows, or strip regions divided along the muscle fiber direction; the size of the rectangular grid ranged from 16×16 pixels to 64×64 pixels; when using overlapping sliding windows, the overlap rate ranged from 0% to 50%.

[0081] For each local tissue unit of the pig sample to be tested, the tissue homogeneity characteristics, tissue interface response characteristics, and local echo gradient characteristics were calculated. The tissue homogeneity characteristics were calculated according to... Calculate, where, Indicates the number of pig samples to be tested. The characteristics of tissue homogeneity in individual tissue units Indicates the number of pig samples to be tested. The average gray value of a local tissue unit. Indicates the number of pig samples to be tested. The standard deviation of gray level in a local tissue unit. A constant is used to prevent the denominator from being zero. for to .

[0082] Organizational interface response characteristics according to Calculate, where, Indicates the number of pig samples to be tested. Organizational interface response characteristics of individual organizational units Indicates the number of pig samples to be tested. The total number of pixels in a local organizational unit. Indicates the number of pig samples to be tested. A collection of pixels within a local organizational unit Indicates the pixel position within a local tissue unit. Indicates pixel position grayscale gradient magnitude at that location Indicates the gradient threshold. This is an indicator function that takes the value 1 when the condition is true and 0 when the condition is false. Gradient threshold. The gradient amplitude is 1.0 to 2.5 times the mean gradient amplitude of the target muscle region of the longissimus dorsi muscle in the same pig sample. The gradient calculation operator is the Sobel operator, Prewitt operator, or first-order difference operator. Tissue interface response characteristics. The value range is from 0 to 1.

[0083] Local echo gradient features according to Calculate, where, Indicates the number of pig samples to be tested. Local echo gradient characteristics of individual tissue units Indicates the number of pig samples to be tested. The total number of pixels in a local organizational unit. Indicates the number of pig samples to be tested. A collection of pixels within a local organizational unit Indicates the pixel position within a local tissue unit. Indicates pixel position The grayscale gradient magnitude at a given location. Local echo gradient characteristics. The value range is determined based on the grayscale normalization range and the size of the local tissue unit.

[0084] The tissue homogeneity characteristics, tissue interface response characteristics, and local echo gradient characteristics of all local tissue units in the pig samples to be tested are summarized in order of local tissue unit to form the response feature vector of the tissue structure to be tested. In the formula, This represents the response feature vector of the tested tissue structure in the pig sample. This represents the set of tissue homogeneity characteristics of the pig samples to be tested. This represents the set of tissue interface response features of the pig sample to be tested. This represents the set of local echo gradient features of the pig sample to be tested. The response feature vector of the tissue structure to be tested. Dimensions and modeling of pig sample tissue structure response feature vectors The dimensions are consistent.

[0085] When a pig sample has multiple frames of ultrasound images of the longissimus dorsi muscle, each frame of the longissimus dorsi ultrasound image undergoes grayscale normalization, size standardization, determination of the target muscle region of the longissimus dorsi, local tissue unit segmentation, and feature calculation. The response feature vector of the tissue structure to be tested corresponding to the multiple frames of the longissimus dorsi ultrasound image can be the mean, weighted mean, or retained frame by frame as an intra-sample feature set of the same pig sample. When using the mean or weighted mean, a single response feature vector of the tissue structure to be tested corresponding to the pig sample is formed; when using the frame-by-frame retention method, multiple response feature vectors of the tissue structure to be tested under the same sample identifier are formed.

[0086] Using the mean and standard deviation determined during the establishment of the tissue state-constrained prediction model, the individual growth status data of the pig samples to be tested were standardized. (The text then abruptly shifts to a different topic: the number of pig samples to be tested...) The standardized values ​​of individual growth status data are based on Calculate, where, Indicates the number of pig samples to be tested. Standardized values ​​of individual growth status data Indicates the number of pig samples to be tested. The original values ​​of individual growth status data Indicates the first training sample The mean of the growth status data of each individual. Indicates the first training sample The standard deviation of the growth status data of each individual pig. The individual growth status vector of the pig sample is represented as follows: Individual growth state vector The dimensions are 3 to 4, and the number of missing individual growth status data does not exceed 25% of the total number of individual growth status data. The missing value handling method is consistent with the missing value handling method used in the establishment of the organizational state constraint prediction model.

[0087] Based on the individual growth state vector of the pig sample to be tested Calculate the state correction factor for the individual to be tested The formula for calculating the individual state correction factor is as follows: In the formula, This represents the correction factor for the individual state of the pigs in the sample being tested. Indicates the correction amplitude coefficient. This represents the individual state correction weight matrix. This represents the individual growth state vector of the pig sample to be tested. This represents the individual state correction bias vector. This represents the hyperbolic tangent function. Individual state correction weight matrix. Individual state correction bias vector and correction amplitude coefficient Fixed values ​​determined by the parameters of the organizational state-constrained prediction model. Individual state correction factor. The dimension and response feature vector of the organization structure under test The dimensions are consistent, and the individual state correction factor is the same. The range of values ​​for each dimension is to Correction amplitude coefficient It ranges from 0.05 to 0.50.

[0088] Using the individual state correction factor Response feature vector of the tissue structure under test Dimensional correction is performed to form a characterization of the intramuscular adipose tissue state to be tested. The formula for characterizing the state of the intramuscular adipose tissue to be tested is as follows: In the formula, This indicates the state of the intramuscular adipose tissue in the pig sample being tested. This represents the response feature vector of the tested tissue structure in the pig sample. This represents the correction factor for the individual state of the pigs in the sample being tested. This represents element-wise multiplication. Element-wise multiplication divides the response feature vector of the tested tissue structure according to the same dimensional position. Each eigenvalue in the model is related to the state correction factor of the individual being tested. Multiply the corresponding correction values ​​in the table. Characterization of the intramuscular adipose tissue state to be tested. Dimensions and Modeling of Intramuscular Adipose Tissue State in Pig Samples The dimensions are consistent.

[0089] Characterizing the state of intramuscular adipose tissue to be tested Input the organizational state constraint prediction model, and use the organizational state constraint prediction model parameters. Perform predictive calculations and output the predicted intramuscular fat content of the pig samples to be tested. The prediction calculation formula is as follows: In the formula, This represents the predicted intramuscular fat content of the pig sample to be tested. The parameter is Organizational state constraint prediction model, This indicates the state of the intramuscular adipose tissue in the pig sample being tested. These represent the parameters of the tissue state-constrained prediction model. Predicted intramuscular fat content value. The unit is %, predicted intramuscular fat content. The output accuracy is 0.01% to 0.1%. Organizational state constraint prediction model parameters. The sample of pigs to be tested remains constant during the prediction process.

[0090] Example 1:

[0091] This embodiment provides a method for predicting intramuscular fat content in pigs based on multimodal feature fusion and multilayer perceptron, including the following steps:

[0092] S1: Collect ultrasound images of the longissimus dorsi muscle, individual growth status data, and measured values ​​of intramuscular fat content of modeling pig samples, and form an original sample dataset based on sample identifiers.

[0093] In this embodiment, 200 commercial pigs at the same growth stage were selected as the modeling pig sample. A unique sample identifier was established for each of the 200 modeling pig samples, using a combination of letters and numbers, and sequentially designated as P0001 to P0200. Three frames of ultrasound images of the longissimus dorsi muscle region were acquired for each modeling pig sample. Animal ultrasound equipment was used for image acquisition, and the probe type, scanning position, image acquisition parameters, and image saving format were kept consistent throughout the experiment. The ultrasound images had a grayscale depth of 8 bits, a pixel size of 512×512 pixels, and were saved in PNG format.

[0094] Record the individual growth status data of each modeling pig sample according to the sample identifier. The individual growth status data includes age in days, weight, backfat thickness, and eye muscle area. The unit of age in days is d, with a recording accuracy of 1 day; the unit of weight is kg, with a recording accuracy of 0.1 kg; the unit of backfat thickness is mm, with a recording accuracy of 0.1 mm; and the unit of eye muscle area is cm. 2 The recording accuracy is 0.1cm. 2 In this embodiment, four individual growth status data were recorded for each modeled pig sample: age, weight, backfat thickness, and eye muscle area.

[0095] The measured intramuscular fat content of each modeling pig sample was obtained, and this measured value was used as a training label and bound to the same sample identifier. The unit of the measured intramuscular fat content was %, and the recording precision was 0.01%. Each modeling pig sample formed one original sample data point, represented as follows: ,in, Indicates the first Sample identifiers for a single modeling pig sample. Indicates the first Ultrasound images of the longissimus dorsi muscle in a modeling pig sample. Indicates the first Individual growth status data of a sample of modeling pigs. Indicates the first The measured intramuscular fat content of a sample of modeling pigs. The original sample dataset consists of 200 original sample data points.

[0096] S2: Determine the target muscle region of the longissimus dorsi muscle based on the original sample dataset, and extract the tissue structure response feature vector from the target muscle region of the longissimus dorsi muscle.

[0097] Ultrasound images of the longissimus dorsi muscle from the original sample dataset were read. Each frame of the longissimus dorsi ultrasound image underwent grayscale normalization, normalizing the pixel grayscale range to 0 to 1. The normalized longissimus dorsi ultrasound images were then resized, and the boundaries of the longissimus dorsi region were determined through manual annotation. The target longissimus dorsi muscle region was then cropped based on these boundaries, resulting in an image size of 256×256 pixels.

[0098] Each frame's target muscle region of the longissimus dorsi muscle is divided into 16 rectangular grid-like local tissue units, each with a size of 64×64 pixels and zero overlap between local tissue units. The 16 local tissue units are arranged in a left-to-right, top-to-bottom order.

[0099] Calculate the tissue homogeneity characteristics for each local tissue unit. For the th The modeling pig sample number Each local tissue unit, the tissue homogeneity characteristics are according to Calculate, where, Indicates the first The modeling pig sample number The average gray value of a local tissue unit. Indicates the first The modeling pig sample number The standard deviation of gray level in a local tissue unit. This represents a constant used to prevent the denominator from being zero. Indicates the first The modeling pig sample number The tissue uniformity characteristics of individual local tissue units. In this embodiment, Pick .

[0100] For each local tissue unit, the tissue interface response characteristics and local echo gradient characteristics are calculated. The Sobel operator is used for gradient calculation, and the gradient threshold is... Take 1.5 times the average gradient amplitude of the target muscle region of the same longissimus dorsi muscle. Tissue interface response characteristics are based on... Calculate, where, Indicates the first The modeling pig sample number Organizational interface response characteristics of individual organizational units Indicates the first The total number of pixels in a local organizational unit. Indicates the first A collection of pixels within a local organizational unit Indicates the pixel position within a local tissue unit. Indicates pixel position grayscale gradient magnitude at that location This is an indicator function that takes the value 1 when the condition is true and 0 when the condition is false.

[0101] Local echo gradient features according to Calculate, where, Indicates the first The modeling pig sample number The local echo gradient features of each local organizational unit are analyzed. The tissue homogeneity features, tissue interface response features, and local echo gradient features of the 16 local organizational units are summarized according to the order of the local organizational units to form a tissue structure response feature vector. In this embodiment, the tissue structure response feature vector corresponding to each frame of the longissimus dorsi muscle ultrasound image is... It is 48-dimensional. The tissue structure response feature vectors corresponding to the three frames of ultrasound images of the longissimus dorsi muscle of the same modeled pig sample are retained frame by frame and bound to the same sample identifier.

[0102] S3: Construct individual state correction factors based on individual growth state data, and use individual state correction factors to correct the tissue structure response feature vector to form a characterization of intramuscular adipose tissue state.

[0103] The age, weight, backfat thickness, and eye muscle area of ​​the modeling pig samples were standardized to form individual growth state vectors. For the first... The modeling pig sample number Individual growth status data, standardized values ​​according to Calculate, where, Indicates the first The modeling pig sample number The original values ​​of individual growth status data Indicates the first training sample The mean of the growth status data of each individual. Indicates the first training sample Standard deviation of individual growth status data Indicates the first The modeling pig sample number Standardized values ​​of individual growth status data. In this embodiment, each modeling pig sample has complete data on age in days, weight, backfat thickness, and eye muscle area. The individual growth status vector is represented as follows: Individual growth state vector The dimension is 4.

[0104] Based on individual growth state vector Calculate individual state correction factor Individual state correction factor according to Calculate, where, Indicates the first Individual state correction factors for a sample of modeling pigs. Indicates the correction amplitude coefficient. This represents the individual state correction weight matrix. Indicates the first The individual growth state vector of a modeling pig sample. This represents the individual state correction bias vector. This represents the hyperbolic tangent function. In this embodiment, the correction amplitude coefficient... Set to 0.20, organizational structure response feature vector The dimension is 48, and the individual state correction weight matrix is... It is a 48×4 matrix, representing the individual state correction bias vector. A 48-dimensional vector, individual state correction factor It is a 48-dimensional vector.

[0105] Using individual state correction factors Organizational structure response feature vector Dimensional correction was performed to form a characterization of intramuscular adipose tissue. The state of intramuscular adipose tissue is characterized according to... Calculate, where, Indicates the first Characterization of intramuscular adipose tissue status in a modeling pig sample. Indicates the first The tissue structure response feature vector of a modeling pig sample. Indicates the first Individual state correction factors for a sample of modeling pigs. This indicates element-wise multiplication. In this embodiment, the intramuscular adipose tissue state is characterized. The dimension is 48.

[0106] S4: Establish a tissue state-constrained prediction model using intramuscular adipose tissue state characterization and measured intramuscular fat content.

[0107] The 200 modeling pigs were divided into training, validation, and test sets according to their sample labels. The training set consisted of 140 modeling pigs, the validation set of 30 modeling pigs, and the test set of 30 modeling pigs. Three frames of ultrasound images of the longissimus dorsi muscle from the same modeling pig sample were assigned to the same subset. A stratified random partitioning method was used to ensure that the distribution of measured intramuscular fat content across the training, validation, and test sets remained largely consistent.

[0108] A tissue state-constrained prediction model was established. This model receives 48 dimensions of intramuscular adipose tissue state representation. As input, the organizational state-constrained prediction model consists of one multimodal state fusion layer, three multilayer perceptron prediction layers, and one linear output layer. The multimodal state fusion layer follows... Compute the predicted input representation ,in, Indicates the first The predicted input representation for a sample of modeled pigs, This represents the weight matrix of the multimodal state fusion layer. This represents the bias vector of the multimodal state fusion layer. This represents a non-linear activation function. In this embodiment, the predicted input represents... The dimension is 64, and the non-linear activation function is... Using ReLU, the random inactivation rate of the multimodal state fusion layer is 0.10.

[0109] Three multilayer perceptron prediction layers are set after the multimodal state fusion layer. The number of neurons in the three multilayer perceptron prediction layers are 128, 64, and 32 respectively, and the activation function of the multilayer perceptron prediction layers is ReLU. The linear output layer has one neuron and a one-dimensional output dimension, with the output value in percentage units. The prediction calculation formula of the organizational state constraint prediction model is as follows: ,in, Indicates the first Predicted intramuscular fat content of a modeling pig sample The parameter is Organizational state constraint prediction model, Indicates the first Characterization of intramuscular adipose tissue status in a modeling pig sample.

[0110] The parameters of the tissue state constraint prediction model are established using training set samples. The training loss function includes a prediction error term, an intra-sample consistency constraint term, and a correction magnitude constraint term. The training loss function is as follows: ,in, This indicates the number of samples in the training batch. Indicates the first Predicted intramuscular fat content for each training sample. Indicates the first Measured intramuscular fat content of each training sample This represents the set of sample pairs formed by different ultrasound image frames of the longissimus dorsi muscle under the same sample identifier. Indicates the number of sample pairs. and These represent the predicted intramuscular fat content values ​​corresponding to two ultrasound images of the longissimus dorsi muscle under the same sample identifier. Indicates the first Individual state correction factor for each training sample This represents a vector in which all dimensions are 1. Indicates the weight of the in-sample consistency constraint. This indicates the weight of the correction magnitude constraint.

[0111] In this embodiment, the training batch size is 32, the learning rate is 0.001, the maximum number of training epochs is 200, the optimizer used is the Adam optimizer, and the intra-sample consistency constraint weights are... Set the value to 0.50, and adjust the amplitude constraint weight accordingly. Set the value to 0.01. After each training epoch, calculate the validation error using the validation set samples. The validation error is calculated according to... Calculate, where, Indicates the number of samples in the validation set. Indicates the first Predicted intramuscular fat content values ​​for each validation set sample. Indicates the first The measured intramuscular fat content of the validation set samples was used. The early stop waiting period was set to 20 rounds. Parameter updates were stopped when the validation error did not decrease within 20 consecutive training rounds, and the model parameters with smaller validation errors were selected as the tissue state constraint prediction model parameters. .

[0112] S5: Process the data of the pig samples to be tested according to the data processing method corresponding to the tissue state constraint prediction model, and output the predicted value of intramuscular fat content of the pig samples to be tested.

[0113] Thirty modeled pigs from the test set were used as the test pig samples. A sample identifier was established for each test pig sample. Three frames of ultrasound images of the longissimus dorsi muscle were read from each test pig sample, and the age, weight, backfat thickness, and oculomotor muscle area of ​​each test pig sample were recorded. Following the grayscale normalization method, the method for determining the target muscle region of the longissimus dorsi muscle, the method for dividing local tissue units, and the feature calculation method determined in S2, the response feature vector of the tested tissue structure for each test pig sample was formed. The response feature vector of the tissue structure under test The dimension is 48.

[0114] The mean calculated using training set samples and standard deviation The age, weight, backfat thickness, and eye muscle area of ​​the pig samples to be tested were standardized to form an individual growth state vector of the pig samples to be tested. The individual state correction weight matrix determined in S4 is used. Individual state correction bias vector and correction amplitude coefficient Calculate the state correction factor for the individual to be tested Among them, the correction amplitude coefficient Set the value to 0.20. Use the individual's state correction factor. Response feature vector of the tissue structure under test Dimensional correction is performed to form a characterization of the intramuscular adipose tissue state to be tested. The calculation formula is Characterizing the state of intramuscular adipose tissue to be tested. Input the organizational state constraint prediction model, according to Output the predicted intramuscular fat content of the pig samples to be tested. Predicted intramuscular fat content The unit is %, and the output precision is 0.01%.

[0115] Comparative Example 1:

[0116] This comparative example provides a method for predicting intramuscular fat content in pigs directly from existing ultrasound images, including the following steps:

[0117] M1: Collect sample data.

[0118] This comparative example uses 200 commercial pigs, the same as in Example 1, as the modeling pig sample. Sample identifiers were established for each of the 200 modeling pigs, sequentially designated as P0001 to P0200. Three frames of ultrasound images of the longissimus dorsi muscle region were acquired for each modeling pig sample. The ultrasound image acquisition used the same animal ultrasound equipment as in Example 1, with the same probe type, scanning position, image acquisition parameters, and image saving format. The ultrasound image grayscale depth was 8 bits, the pixel size was 512×512 pixels, and the image was saved in PNG format.

[0119] The measured intramuscular fat content of each modeling pig sample was obtained, and this measured value was used as a training label and bound to the same sample identifier. The unit of the measured intramuscular fat content was %, and the recording accuracy was 0.01%. This comparative example does not utilize age, weight, backfat thickness, and eye muscle area to construct individual state correction factors.

[0120] M2: Identify the longissimus dorsi muscle region and extract ultrasound image features.

[0121] Each frame of the longissimus dorsi muscle ultrasound image was read, and grayscale normalization was performed on the longissimus dorsi muscle ultrasound image to normalize the pixel grayscale range to 0 to 1. The boundary of the longissimus dorsi muscle region was determined by manual annotation, and the longissimus dorsi muscle region image was cropped to obtain the longissimus dorsi muscle region image with a size of 256×256 pixels.

[0122] Ultrasound image features were extracted from each frame of the longissimus dorsi muscle region image. These features included gray-level mean, gray-level standard deviation, texture contrast, energy, and correlation. Texture contrast, energy, and correlation were calculated based on the gray-level co-occurrence matrix (GLCM). The GLCM had 16 gray-level quantization levels, a pixel distance of 1 pixel, and calculation directions of 0°, 45°, 90°, and 135°. The texture contrast, energy, and correlation were averaged for each of the four directions. Each frame of the longissimus dorsi muscle region image formed a 5D ultrasound image feature vector. The 5D ultrasound image feature vectors corresponding to the three frames of the longissimus dorsi muscle region images from the same modeled pig sample were retained frame by frame and bound to the same sample identifier.

[0123] M3: Establish a multilayer perceptron prediction model.

[0124] The 200 modeling pig samples were divided into training, validation, and test sets according to their sample identifiers. The training set consisted of 140 modeling pig samples, the validation set consisted of 30 modeling pig samples, and the test set consisted of 30 modeling pig samples. Three frames of ultrasound images of the longissimus dorsi muscle from the same modeling pig sample were assigned to the same data subset. The sample division method was the same as in Example 1.

[0125] A multilayer perceptron prediction model was established, which received a 5-dimensional ultrasound image feature vector as input. The multilayer perceptron prediction model consisted of three hidden layers and one linear output layer. The number of neurons in the three hidden layers were 128, 64, and 32 respectively, and the ReLU activation function was used in all hidden layers. The linear output layer had one neuron and a 1-dimensional output dimension, with the output value in units of %.

[0126] A multilayer perceptron prediction model is trained using training set samples. The training loss function is the mean squared error loss, calculated as follows: ,in, This represents the training loss of the multilayer perceptron prediction model. This indicates the number of samples in the training batch. Indicates the first Predicted intramuscular fat content for each training sample. Indicates the first The measured intramuscular fat content of each training sample was used. The batch size for this comparative example was 32, the learning rate was 0.001, the maximum training epochs were 200, and the Adam optimizer was used. After each training epoch, the validation error was calculated using the validation set samples. The early stopping epoch was set to 20 epochs; parameter updates stopped if the validation error did not decrease within 20 consecutive training epochs.

[0127] M4: Outputs predicted intramuscular fat content.

[0128] Thirty modeled pigs from the test set were used as the test pig samples. For each test pig sample, three frames of ultrasound images of the longissimus dorsi muscle were used to extract a 5D ultrasound image feature vector using the M2 method. This 5D ultrasound image feature vector was then input into a trained multilayer perceptron prediction model, which outputs a predicted value of intramuscular fat content for the test pig sample. The predicted intramuscular fat content is expressed as a percentage (%), with an output precision of 0.01%.

[0129] Comparative Example 2:

[0130] In this comparative example, unlike Example 1, after completing S2 and forming the tissue structure response feature vector... Subsequently, individual state correction factors were not constructed based on individual growth state data, and the tissue structure response feature vectors were not utilized based on individual state correction factors. Dimensional correction is performed to avoid forming a characterization of intramuscular adipose tissue state. Instead, it directly uses the 48-dimensional organizational structure response feature vector. As a prediction input.

[0131] In this comparative example, S1 is the same as in Example 1, using a sample of 200 modeling pigs. For each modeling pig sample, three frames of ultrasound images of the longissimus dorsi muscle were acquired, and the measured values ​​of age, weight, backfat thickness, eyelid area, and intramuscular fat content were recorded. S2 is the same as in Example 1, with the target muscle region image size of the longissimus dorsi muscle being 256×256 pixels, the number of local tissue units being 16, and the size of each local tissue unit being 64×64 pixels. The tissue uniformity characteristics... Pick The gradient calculation operator uses the Sobel operator, and the gradient threshold is... Taking 1.5 times the average gradient amplitude of the same target muscle region of the longissimus dorsi, each frame of the longissimus dorsi ultrasound image is used to form a 48-dimensional tissue structure response feature vector. .

[0132] In this comparative example, the individual state correction factor calculation and dimension-by-dimensional correction processing described in S3 of Example 1 are not performed. Individual growth state data are only recorded as acquisition conditions and are not used in the prediction input construction. The prediction model receives a 48-dimensional tissue structure response feature vector. As input, the prediction model employs a multilayer perceptron structure, consisting of three hidden layers and one linear output layer. The number of neurons in the three hidden layers are 128, 64, and 32, respectively, and the activation function for all hidden layers is ReLU. The linear output layer has one neuron and a one-dimensional output dimension, with the output value in units of 0.5.

[0133] In this comparative example, the division of the training set, validation set, and test set is the same as in Example 1. The training set includes 140 modeling pig samples, the validation set includes 30 modeling pig samples, and the test set includes 30 modeling pig samples. Three frames of ultrasound images of the longissimus dorsi muscle from the same modeling pig sample are grouped into the same data subset. The prediction model is trained using the mean squared error loss function, and the training loss function is... ,in, This indicates the number of samples in the training batch. Indicates the first Predicted intramuscular fat content for each training sample. Indicates the first The measured intramuscular fat content of each training sample. The batch size for this comparative example was 32, the learning rate was 0.001, the maximum number of training epochs was 200, the Adam optimizer was used, the validation error was calculated after each training epoch, and the early stopping wait epochs were set to 20.

[0134] In this comparative example, 30 modeling pigs from the test set were used as the test pigs. For each test pig, a response feature vector of the tissue structure was generated in the same manner as in S2. and the response feature vector of the tissue structure to be tested The trained multilayer perceptron prediction model is input, and the predicted intramuscular fat content of the pig samples is output. The unit of the predicted intramuscular fat content is %, and the output accuracy is 0.01%. Except for the differences mentioned above, the sample source, ultrasound image acquisition conditions, method of determining the target muscle region of the longissimus dorsi muscle, method of dividing local tissue units, feature calculation parameters, and dataset division method are all the same as in Example 1.

[0135] Comparative Example 3:

[0136] In this comparative example, unlike Example 1, the same method was used to obtain the characterization of intramuscular adipose tissue. However, the model is trained using a multilayer perceptron loss function, without including intra-sample consistency constraints and correction magnitude constraints.

[0137] In this comparative example, S1 to S3 are the same as in Example 1. Specifically, this comparative example uses a sample of 200 modeling pigs. For each modeling pig sample, three frames of ultrasound images of the longissimus dorsi muscle are acquired. The grayscale bit depth of the ultrasound images is 8 bits, the pixel size of the ultrasound images is 512×512 pixels, and the ultrasound images are saved in PNG format. The image size of the target muscle region of the longissimus dorsi muscle is 256×256 pixels. The number of local tissue units is 16, and the size of each local tissue unit is 64×64 pixels. Tissue structure response feature vector. The dimension is 48; individual growth state vector The dimension is 4-dimensional; correction amplitude coefficient Set to 0.20; Individual state correction factor The dimensions are 48; intramuscular adipose tissue state characterization The dimension is 48.

[0138] In this comparative example, the prediction model received 48 dimensions of intramuscular adipose tissue state characterization. As input, the prediction model employs a multilayer perceptron structure, comprising three hidden layers and one linear output layer. The number of neurons in the three hidden layers are 128, 64, and 32 respectively, with ReLU activation function used in all hidden layers. The linear output layer has one neuron and a one-dimensional output dimension, with output values ​​expressed as a percentage. This comparative model does not include intra-sample consistency constraints or correction magnitude constraints.

[0139] In this comparative example, the division of the training set, validation set, and test set is the same as in Example 1. The training set includes 140 modeling pig samples, the validation set includes 30 modeling pig samples, and the test set includes 30 modeling pig samples. Three frames of ultrasound images of the longissimus dorsi muscle from the same modeling pig sample are grouped into the same data subset. The prediction model is trained using the mean squared error loss function, and the training loss function is... ,in, This indicates the number of samples in the training batch. Indicates the first Predicted intramuscular fat content for each training sample. Indicates the first The measured intramuscular fat content of each training sample. The batch size for this comparative example was 32, the learning rate was 0.001, the maximum number of training epochs was 200, the Adam optimizer was used, the validation error was calculated after each training epoch, and the early stopping wait epochs were set to 20.

[0140] In this comparative example, 30 modeling pigs from the test set were used as the test pigs. For each test pig, the intramuscular adipose tissue state was characterized in the same manner as in Example 1. The state of the intramuscular adipose tissue to be tested was characterized. The trained multilayer perceptron prediction model is input, and the predicted intramuscular fat content of the pig samples is output. The unit of the predicted intramuscular fat content is %, and the output accuracy is 0.01%. Except for the absence of intrasample consistency constraints and correction magnitude constraints, the sample source, ultrasound image acquisition conditions, feature extraction parameters, individual state correction parameters, and dataset partitioning method in this comparative example are the same as in Example 1.

[0141] Experimental Example 1:

[0142] The experimental subjects in this example are the prediction process and prediction output results of intramuscular fat content in pigs corresponding to Example 1 and Comparative Examples 1 to 3.

[0143] The following tests were performed on Example 1 and Comparative Examples 1 to 3 respectively.

[0144] Test 1: Root Mean Square Error Test. The predicted results for the pig samples in the test set are compared with the measured values ​​of intramuscular fat content, and the root mean square error (RMSE) is calculated. RMSE is calculated according to... Calculate, where, Indicates the number of test samples. Indicates the first Predicted intramuscular fat content for each test sample, in percentage. Indicates the first The measured intramuscular fat content of each test sample, expressed as a percentage. RMSE is used to evaluate the deviation between the predicted value and the measured intramuscular fat content.

[0145] Test 2: Mean Absolute Error Test. The predicted results and measured intramuscular fat content of the pig samples in the test set were compared on a sample-by-sample basis to calculate the mean absolute error (MAE). MAE was calculated according to... Calculate, where, Indicates the number of test samples. Indicates the first Predicted intramuscular fat content for each test sample. Indicates the first The measured intramuscular fat content of each test sample. MAE is used to evaluate the overall bias of the prediction results.

[0146] Test 3: Coefficient of Determination Test. The consistency between the predicted results and the measured intramuscular fat content of the pig samples in the test set is calculated to obtain the coefficient of determination. The coefficient of determination is based on... Calculate, where, This represents the average measured value of intramuscular fat content in the test samples. Indicates the first The measured intramuscular fat content of each test sample. Indicates the first The predicted intramuscular fat content of a test sample. The coefficient of determination is used to evaluate the degree of agreement between the predicted and measured values ​​in terms of their trends.

[0147] Test 4: Individual State Adaptation Error Test. The test samples were divided according to age or weight. For each group of individual states, calculate the prediction error for each group and the individual state adaptation error. Individual state adaptation error according to Calculate, where, This indicates the number of individual status groups categorized by age or weight. Indicates the first Group sample prediction error This represents the prediction error for all test samples. Individual state adaptation error is used to evaluate the degree of fluctuation in prediction error under different individual states.

[0148] The results of this experiment are shown in Table 1:

[0149] Table 1

[0150] Comparative Example 1 0.86 0.67 0.71 0.182 Comparative Example 2 0.69 0.52 0.81 0.151 Comparative Example 3 0.61 0.46 0.85 0.124 Example 1 0.52 0.39 0.90 0.086

[0151] Comparative Example 1 employs existing direct prediction methods for ultrasound images, extracting only ultrasound image features such as gray-level mean, gray-level standard deviation, texture contrast, energy, and correlation, and inputting these features into a multilayer perceptron prediction model. Comparative Example 1 does not further convert the longissimus dorsi muscle ultrasound image into a tissue structure response feature vector composed of tissue homogeneity features, tissue interface response features, and local echo gradient features, nor does it introduce individual state correction and tissue state constraints. In the corresponding test results, Comparative Example 1 achieved an RMSE of 0.86% and a MAE of 0.67%. The individual state adaptation error is 0.71. It is 0.182.

[0152] Comparative Example 2 modifies the prediction input construction method based on Comparative Example 1, and adopts the same tissue structure response feature extraction process as Example 1. It extracts tissue homogeneity features, tissue interface response features, and local echo gradient features from the target muscle region of the longissimus dorsi muscle, and forms a tissue structure response feature vector. For Example 2, no individual state correction is performed; the organizational structure response feature vector is directly applied. As a prediction input, compared to Comparative Example 1, Comparative Example 2's RMSE changed from 0.86% to 0.69%, and its MAE changed from 0.67% to 0.52%. The individual state adaptation error changed from 0.71 to 0.81. It changed from 0.182 to 0.151.

[0153] Comparative Example 3 adds an individual state correction process to Comparative Example 2, using age, weight, backfat thickness, and eye muscle area to form an individual growth state vector. And through individual state correction factors Organizational structure response feature vector Dimensional correction was performed to form a characterization of intramuscular adipose tissue. Comparative Example 3 was trained using a multilayer perceptron loss function, without including intra-sample consistency constraints and correction magnitude constraints. Compared to Comparative Example 2, the RMSE of Comparative Example 3 decreased from 0.69% to 0.61%, and the MAE decreased from 0.52% to 0.46%. The individual state adaptation error changed from 0.81 to 0.85. It changed from 0.151 to 0.124.

[0154] Example 1 further employs a tissue state-constrained prediction model based on Comparative Example 3. The training loss function in Example 1 includes a prediction error term, an intra-sample consistency constraint term, and a correction magnitude constraint term. The intra-sample consistency constraint term applies to sample pairs formed from different ultrasound image frames of the longissimus dorsi muscle under the same sample identifier, while the correction magnitude constraint term applies to the individual state correction factor. The degree of deviation relative to a vector where all dimensions are 1. Compared to Comparative Example 3, the RMSE of Example 1 changed from 0.61% to 0.52%, and the MAE changed from 0.46% to 0.39%. The individual state adaptation error changed from 0.124 to 0.086, from 0.85 to 0.90. In Example 1, the tissue state constraint prediction model, while establishing a prediction mapping relationship using the intramuscular adipose tissue state characterization, constrains the prediction fluctuations and correction amplitudes of the individual state correction factor for multiple frames of longissimus dorsi ultrasound images under the same sample identifier.

[0155] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for predicting intramuscular fat content in pigs based on multimodal feature fusion and multilayer perceptron, characterized in that, Includes the following steps: S1: Collect ultrasound images of the longissimus dorsi muscle, individual growth status data, and measured values ​​of intramuscular fat content of modeling pig samples, and form an original sample dataset based on sample identification; S2: Determine the target muscle region of the longissimus dorsi muscle based on the original sample dataset, and extract the tissue structure response feature vector from the target muscle region of the longissimus dorsi muscle. S3: Construct an individual state correction factor based on the individual growth state data, and use the individual state correction factor to correct the tissue structure response feature vector to form an intramuscular adipose tissue state characterization. S4: Establish a tissue state-constrained prediction model using the intramuscular adipose tissue state characterization and the measured value of intramuscular adipose content; S5: Process the data of the pig samples to be tested according to the data processing method corresponding to the tissue state constraint prediction model, and output the predicted value of intramuscular fat content of the pig samples to be tested.

2. The method for predicting intramuscular fat content in pigs according to claim 1, characterized in that, In step S1, a data association relationship is established between the longissimus dorsi muscle ultrasound image, individual growth status data, and measured intramuscular fat content based on the sample identifier. The individual growth status data includes age, weight, backfat thickness, and eye muscle area.

3. The method for predicting intramuscular fat content in pigs according to claim 2, characterized in that, In step S1, the number of ultrasound images of the longissimus dorsi muscle is 1 to 5 frames, the grayscale bit depth of the ultrasound image of the longissimus dorsi muscle is 8 to 16 bits, and the pixel size of the ultrasound image of the longissimus dorsi muscle is not less than 256×256 pixels.

4. The method for predicting intramuscular fat content in pigs according to claim 1, characterized in that, Step S2 includes performing grayscale normalization and size unification processing on the ultrasound image of the longissimus dorsi muscle, determining the target muscle region of the longissimus dorsi muscle based on the boundary of the longissimus dorsi muscle region, and dividing the target muscle region of the longissimus dorsi muscle into multiple local tissue units.

5. The method for predicting intramuscular fat content in pigs according to claim 4, characterized in that, The local tissue units are rectangular grids, overlapping sliding windows, or strip regions divided along the muscle fiber direction. The number of local tissue units is 4 to 64. Based on the local tissue units, tissue uniformity features, tissue interface response features, and local echo gradient features are extracted respectively, and the tissue structure response feature vector is formed according to the order of the local tissue units.

6. The method for predicting intramuscular fat content in pigs according to claim 1, characterized in that, Step S3 includes standardizing the individual growth state data to form an individual growth state vector, and calculating an individual state correction factor based on the individual growth state vector that is consistent with the dimension of the tissue structure response feature vector.

7. The method for predicting intramuscular fat content in pigs according to claim 6, characterized in that, The individual state correction factor is obtained through an individual state correction weight matrix, an individual state correction bias vector, and a nonlinear transformation method. The correction magnitude coefficient of the individual state correction factor is between 0.05 and 0.

50.

8. The method for predicting intramuscular fat content in pigs according to claim 1, characterized in that, In step S4, the tissue state constraint prediction model includes a tissue state input layer, a multimodal state fusion layer, a multilayer perceptron prediction layer, and a linear output layer. The tissue state constraint prediction model takes the intramuscular adipose tissue state characterization as input and the intramuscular adipose content prediction value as output.

9. The method for predicting intramuscular fat content in pigs according to claim 8, characterized in that, The tissue state constraint prediction model determines its parameters during training based on the prediction error between the predicted and measured intramuscular fat content, the prediction difference between different ultrasound images of the longissimus dorsi muscle under the same sample identifier, and the correction magnitude of the individual state correction factor.

10. The method for predicting intramuscular fat content in pigs according to claim 9, characterized in that, After the tissue state constraint prediction model is trained, the intramuscular fat content of the pig samples to be tested is predicted by using the image processing method, the method for determining the target muscle region of the longissimus dorsi muscle, the method for dividing local tissue units, the method for calculating features, and the individual state correction parameters determined during the training process.

Images