A sow reproductive potential evaluation method, electronic device, storage medium and program product

Abstract

The application provides a sow reproductive potential evaluation method, an electronic device, a storage medium and a program product. The method comprises the following steps: obtaining an original AMH value, an ovary B-ultrasound image, physiological data, feeding environment data, environmental data and health state data of a sow to be evaluated; obtaining an original AMH feature sequence according to the original AMH value and the physiological data; extracting an ovary substance area, a follicle number and a follicle average diameter in the ovary B-ultrasound image; obtaining a corrected AMH feature sequence; obtaining an enhanced image; obtaining an initial predicted litter size and an initial probability of each preset reproductive potential grade of the sow to be evaluated; obtaining a final predicted litter size; obtaining a first intermediate probability of each preset reproductive potential grade; and determining a final probability of the preset reproductive potential grade. The present scheme solves the technical problems of insufficient evaluation robustness and limited precision caused by a single AMH detection index.

Description

Technical Field

[0001] This invention relates to the field of sow reproductive potential assessment technology, and more specifically, to a sow reproductive potential assessment method, electronic device, storage medium, and program product. Background Technology

[0002] In the field of sow reproductive performance assessment, early selection and breeding of replacement gilts directly impacts the efficiency of livestock production. Current technologies primarily rely on the experience and judgment of farm workers to assess sow reproductive potential. This involves a comprehensive evaluation based on observation of the sow's appearance, body condition score, and pedigree records. Some farms also use simple production records as supplementary evidence. While these methods are simple to operate and require no additional testing costs, the results are heavily influenced by the subjective experience of the assessors. It is difficult to standardize judgment criteria among different assessors, resulting in insufficient objectivity and consistency. Furthermore, relying solely on appearance and experience cannot deeply reflect the intrinsic physiological state of the sow's ovarian reserve, limiting the accuracy of the assessment and failing to meet the needs of refined selection and breeding. Summary of the Invention

[0003] In view of this, the purpose of this application is to provide a method, electronic device, storage medium and program product for assessing the reproductive potential of sows, which can improve the problem of limited accuracy in predicting reproductive potential in the prior art.

[0004] To achieve the above technical objectives, the technical solution adopted in this application is as follows:

[0005] In a first aspect, embodiments of this application provide a method for assessing the reproductive potential of sows, the method comprising:

[0006] Obtain the raw AMH value, ovarian ultrasound image, physiological data, feeding environment data, environmental data, and health status data of the sows to be evaluated;

[0007] Based on the original AMH value and physiological data, an original AMH feature sequence is obtained. The original AMH feature sequence includes the original AMH value and several original AMH feature values, which are calculated from the original AMH value and physiological data.

[0008] Based on a preset image processing tool, the ovarian parenchymal area, number of follicles, and average diameter of follicles in the ovarian ultrasound image are extracted.

[0009] Based on the original AMH value, ovarian parenchymal area, number of follicles and average diameter of follicles, the original AMH feature sequence is corrected to obtain the corrected AMH feature sequence, which includes the corrected AMH value and the corrected AMH feature value.

[0010] Based on the corrected AMH value and the specified corrected AMH feature value, the ovarian ultrasound image is enhanced to obtain an enhanced image.

[0011] Based on the corrected AMH feature sequence and the enhanced image, the initial predicted number of piglets to be evaluated and the initial probability of each preset reproductive potential level are obtained according to the pre-trained deep learning model.

[0012] Based on the aforementioned rearing environment data, environmental data, and health status data, the initial predicted litter size is revised to obtain the final predicted litter size.

[0013] The initial probability is corrected based on the final predicted number of offspring to obtain the first intermediate probability for each preset reproductive potential level;

[0014] The final probability of the preset reproductive potential level is determined based on the first intermediate probability, the prior probability of the pre-established preset reproductive potential level, the ovarian parenchymal area, the number of follicles, and the average diameter of follicles.

[0015] Secondly, embodiments of this application also provide an electronic device, the electronic device including a processor and a memory coupled to each other, the memory storing a computer program, and when the computer program is executed by the processor, causing the electronic device to perform the method described in the first aspect.

[0016] Thirdly, embodiments of this application also provide a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when run on a computer, causes the computer to perform the method described in the first aspect.

[0017] Fourthly, embodiments of this application also provide a program product, characterized in that it includes a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0018] The invention employing the above technical solution has the following advantages:

[0019] In the technical solution provided in this application, ovarian morphological information, including ovarian parenchymal area, number of follicles, and average follicle diameter, is extracted from ovarian ultrasound images and introduced into the reproductive potential assessment system. On one hand, the original AMH feature sequence is corrected based on the original AMH value, ovarian parenchymal area, number of follicles, and average follicle diameter. The physiological correlation between morphological parameters and AMH levels is utilized to identify and correct AMH detection bias, resulting in a corrected AMH feature sequence. On the other hand, the ovarian ultrasound image is enhanced based on the corrected AMH value and specified corrected AMH feature values, resulting in an enhanced image that specifically improves the recognition of effective image information. The corrected AMH feature sequence and the enhanced image are input into a pre-trained deep learning model to obtain the initial predicted litter size and initial probabilities for each reproductive potential level, achieving feature-level fusion of numerical and image modalities. Subsequently, the initial predicted litter size is corrected based on rearing environment data, environmental data, and health status data to obtain the final predicted litter size. The initial probability is then corrected to obtain the first intermediate probability, eliminating the interference of external factors on the prediction results. Finally, the final probability is determined based on the first intermediate probability, prior probability, ovarian parenchymal area, number of follicles, and average follicle diameter, completing the decision-level fusion of multi-source information. Through these techniques, mutual verification and complementarity between AMH numerical information and ovarian morphological imaging information are achieved, solving the problem that existing technologies have limited assessment accuracy and cannot meet the needs of refined breeding and selection. Attached Figure Description

[0020] This application can be further illustrated by the non-limiting embodiments given in the accompanying drawings. It should be understood that the following drawings only illustrate some embodiments of this application and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained from these drawings without any inventive effort.

[0021] Figure 1 A flowchart of the sow reproductive potential assessment method provided in the embodiments of this application.

[0022] Figure 2 The S140 sub-flowchart is provided for an embodiment of this application.

[0023] Figure 3 The S170 sub-flowchart is provided for the embodiments of this application.

[0024] Figure 4 The S190 sub-flowchart is provided for the embodiments of this application. Detailed Implementation

[0025] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that similar or identical parts are referred to by the same reference numerals in the drawings or description. Implementations not shown or described in the drawings are forms known to those skilled in the art. In the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] In this embodiment, based on in-depth research, the applicant has demonstrated the rationality of predicting sow reproductive potential based on sow AMH, specifically:

[0027] The applicant used ELISA kits to measure serum AMH levels in sows under 1 month old, 1 month old, 2 months old, 3 months old, 4 months old, 5 months old, 6 months old, 1-3 parities, 4-6 parities, 7-9 parities, and 10-12 parities, and studied the trend of AMH changes in sows of different ages. Blood samples were collected from multiparous sows and AMH concentrations were measured, and their reproductive performance was tracked to study its correlation with reproductive traits. Blood samples were collected from 6-month-old and 8-month-old Duroc gilts and 8-month-old Large White gilts and serum AMH levels were measured. Farrowing data for 2-3 parities were continuously tracked, and the differences in litter size trait among different AMH grade ranges were analyzed to develop a grading index for the reproductive potential of gilts.

[0028] The serum AMH levels of sows of different breeds aged 1-6 months are shown in Table 1. Duroc, Large White, and Landrace sows showed significant variations before 2 months of age, but levels stabilized after 3 months. From 3 to 6 months of age, AMH concentrations remained relatively consistent with small fluctuations, all concentrated between 6-8 ng / ml, and there were no differences between breeds (P>0.05, where P represents the P-value in statistical hypothesis testing, i.e., the probability value).

[0029] Table 1. Distribution of AMH levels in sows of different breeds aged 1-6 months

[0030]

[0031] Blood samples were collected from multiparous sows and divided into four groups according to parity: 1-3, 4-6, 7-9, and 10-12. Statistical analysis revealed highly significant differences in AMH levels among sows of parity 1-3, 4-6, 7-12, and 10-12 (P<0.01), and parity was significantly correlated with AMH concentration (P<0.05). Overall, AMH levels in multiparous sows increased with increasing parity (Table 2).

[0032] Table 2 Serum AMH levels in multiparous sows in different parity groups 1)

[0033] 1) The data in the table are the mean ± standard error. Data in the same column with different capital letters indicates a highly significant difference (P<0.01).

[0034] Correlation analysis of various reproductive traits and serum AMH levels in multiparous sows revealed that, at a significance level of α=0.01, AMH concentration in multiparous sows was positively correlated with current parity (r=0.55), total parity (r=0.52), cumulative total litter size (r=0.54), age at measurement (r=0.53), average litter size (r=0.17), and highest litter size (r=0.34). However, no correlation was detected between AMH concentration and current status (non-pregnant, gestational, lactation) or litter size (P>0.05), as detailed in Table 3. This indicates that AMH is related to ovarian reserve in sows and can be used as an indicator of early reproductive potential in gilts.

[0035] Table 3. Correlation analysis of serum AMH levels with various reproductive traits in multiparous sows. 2)

[0036] 2) **P<0.01, *P<0.05

[0037] Blood samples were collected from 6-month-old and 8-month-old Duroc and Large White gilts, and serum AMH levels were measured. Reproductive records for each sow were continuously tracked for the next two litters. Using total litter size as a representative indicator of reproductive potential, different ranges of AMH levels were established. Significant (P<0.05) or highly significant (P<0.01) differences in litter size were found across different AMH level ranges. For Duroc gilts, 6-month-old gilts with AMH levels <5 ng / ml had significantly fewer total litters in their first two litters than gilts with AMH levels between 5 and 10 ng / ml and >10 ng / ml; similarly, 8-month-old gilts with AMH levels <8 ng / ml had significantly fewer total litters in their first two litters than gilts with AMH levels between 8 and 13 ng / ml and >13 ng / ml. For Large White pigs, 8-month-old gilts with AMH levels <8 ng / ml had significantly fewer piglets in their first two litters than gilts with AMH levels between 8 and 20 ng / ml and >20 ng / ml.

[0038] Table 4. Analysis of differences in litter size among 6-month-old Duroc gilts with different AMH levels. 3)

[0039] 3) For different uppercase letters, P < 0.01; for different lowercase letters, P < 0.05, and so on.

[0040] Table 5. Analysis of differences in litter size among 8-month-old Duroc gilts with different AMH levels.

[0041]

[0042] Table 6. Analysis of differences in litter size among different AMH levels in 8-month-old Large White gilts.

[0043]

[0044] Previous studies have shown that serum AMH levels in sows begin to stabilize after 3 months of age, and are relatively more stable with smaller fluctuations between 6 and 8 months of age. Furthermore, AMH levels are positively correlated with sow reproductive traits and can reflect the sow's reproductive potential. Considering the current timing of boar performance testing and the first mating time of gilts, using serum AMH levels at 6 or 8 months of age is optimal for assessing sow reproductive potential.

[0045] Based on this, the applicant found that the serum AMH level of sows was relatively stable from 3 months of age to before mating, and then increased with the increase of parity. The AMH level was positively correlated with reproductive traits such as total litter size of multiparous sows, and could be used to assess the reproductive performance of sows.

[0046] However, relying solely on AMH as a single serum indicator for assessment, without fully integrating the intrinsic physiological relationship between ovarian morphological imaging information and AMH, results in insufficient robustness of assessment results when errors are introduced during the testing process due to sampling, reagent batches, or operational differences. Furthermore, single-dimensional numerical information is insufficient to comprehensively characterize the morphological features of ovarian reserve, leading to limited accuracy in predicting reproductive potential and failing to meet the needs of refined breeding selection.

[0047] Based on this, please refer to Figure 1 This application provides a method for assessing the reproductive potential of sows, which can be applied to electronic devices and whose steps can be executed or implemented by the electronic devices. The electronic devices can be, but are not limited to, personal computers, smartphones, etc. The method for assessing the reproductive potential of sows may include the following steps:

[0048] S110: Obtain the original AMH value, ovarian ultrasound image, physiological data, feeding environment data, environmental data, and health status data of the sow to be evaluated.

[0049] S120, Based on the original AMH value and physiological data, obtain the original AMH feature sequence, the original AMH feature sequence including the original AMH value and several original AMH feature values, the original AMH feature values ​​being calculated from the original AMH value and physiological data;

[0050] S130, based on a preset image processing tool, extract the ovarian parenchymal area, number of follicles and average diameter of follicles from the ovarian ultrasound image;

[0051] S140, based on the original AMH value, ovarian parenchymal area, number of follicles and average diameter of follicles, the original AMH feature sequence is corrected to obtain the corrected AMH feature sequence, which includes the corrected AMH value and the corrected AMH feature value.

[0052] S150, based on the corrected AMH value and the specified corrected AMH feature value, the ovarian ultrasound image is enhanced to obtain an enhanced image;

[0053] S160, Based on the corrected AMH feature sequence and the enhanced image, the initial predicted number of piglets to be evaluated and the initial probability of each preset reproductive potential level are obtained according to the pre-trained deep learning model.

[0054] S170, Based on the feeding environment data, environmental data, and health status data, the initial predicted litter size is corrected to obtain the final predicted litter size;

[0055] S180, Correct the initial probability according to the final predicted number of offspring to obtain the first intermediate probability for each preset reproductive potential level;

[0056] S190, based on the first intermediate probability, the prior probability of the preset reproductive potential level, the ovarian parenchymal area, the number of follicles, and the average diameter of follicles, determine the final probability of the preset reproductive potential level.

[0057] In this embodiment, morphological parameters such as ovarian parenchymal area, number of follicles, and average follicle diameter are first extracted from ovarian ultrasound images. On one hand, these morphological parameters are used to correct the original AMH feature sequence. Given the natural physiological correlation between follicle number and AMH level, when the ovarian reserve level reflected by the morphological information deviates significantly from the original AMH detection results, potential biases in the original AMH detection data can be identified, and the correction process can reduce the interference of detection errors on subsequent evaluation results. On the other hand, adaptive enhancement is performed on the ovarian ultrasound image based on the corrected AMH value and a specified corrected AMH feature value. Different AMH levels correspond to differentiated follicle size distribution characteristics. Based on the corrected AMH information, the image target features are specifically enhanced: fine follicle texture information is enhanced at high AMH levels, and follicle contour features are enhanced at low AMH levels, improving the image's effective information recognition. Subsequently, the corrected AMH feature sequence and the enhanced image are input into a pre-trained deep learning model to achieve the fusion of numerical and image modal information, outputting the initial predicted litter size and the initial probabilities of each reproductive potential level. Subsequently, based on feeding, environmental, and health-related data, the initial predicted litter size was revised to obtain the first intermediate probability after initial probability correction. Then, combining the first intermediate probability, the prior probability of reproductive potential level, and ovarian morphological parameters, multi-source information fusion was performed to determine the final probability of each preset reproductive potential level. This scheme simultaneously integrates AMH numerical information representing endocrine levels with imaging morphological information representing ovarian tissue status, enabling mutual verification and information complementarity between the two modalities. This avoids the assessment limitations of a single AMH detection indicator, comprehensively characterizes the true level of sow ovarian reserve from multiple dimensions, and effectively improves the accuracy and robustness of sow reproductive potential assessment results.

[0058] The following is a detailed explanation of each step in the method for assessing the reproductive potential of sows:

[0059] In S110, in this embodiment, 6-month or 8-month-old gilts to be evaluated are selected as sampling subjects. Blood samples are collected from the sows using the anterior vena cava and a 5mL coagulation-promoting vacuum tube. After collection, the samples are centrifuged at 3000 rpm for 15 to 20 minutes to separate the supernatant serum, which is then frozen and stored at -20℃ for later testing. Subsequently, the AMH concentration of the serum samples is detected by enzyme-linked immunosorbent assay (ELISA) or chemiluminescence immunoassay. The detection value in nanograms per milliliter is used as the original AMH value. At the same time, the detection time, reagent batch number, and intra-batch coefficient of variation are recorded to provide a basis for subsequent data quality verification and evaluation result correction.

[0060] In this embodiment, ovarian ultrasound image acquisition was performed simultaneously with AMH blood collection to ensure temporal consistency between serological and ovarian morphological data. A B-mode ultrasound diagnostic instrument equipped with a 3.5 to 5 MHz convex array probe was used for acquisition. The grayscale acquisition mode was switched, and the image depth was adaptively adjusted according to the body size of the sow to ensure complete presentation of the ovarian parenchyma structure. The operator placed the probe on the abdominal wall lateral to the last pair of teats of the sow, keeping it perpendicular to the ovarian surface, and explored and acquired clear images of the largest longitudinal sections of the left and right ovaries, completely preserving the ovarian parenchyma and follicular texture structure. Finally, the bilateral ovarian images were stored according to a unified naming rule of ear number-test date-left / right side, forming a standardized grayscale image dataset, providing high-quality image support for subsequent ovarian morphological parameter extraction.

[0061] In this embodiment, the sow to be evaluated can be a Large White, Landrace, or Duroc pig.

[0062] Physiological data include the age, breed, body weight, and backfat thickness of the sows to be evaluated. The methods for obtaining each indicator are as follows:

[0063] (1) Age in days: Calculated based on the sow’s birth date and test date, in days.

[0064] (2) Breed: Determined based on sow pedigree records, including breeds such as Large White or Duroc or their hybrid combinations.

[0065] (3) Weight: The sows were weighed in a fasting state using an electronic scale, and the weight was measured in kilograms (kg).

[0066] (4) Back fat thickness: Back fat thickness is measured by ultrasound and measured in millimeters (mm).

[0067] Feeding environment data refers to parameters of feeding conditions that are directly related to the pig farm's feeding and management strategies and are artificially controlled. These parameters reflect the level of feeding and management of sows and are considered externally controllable factors. Specifically, they include feeding programs, daily feed intake, stocking density, and pen type. Among these, the feeding program represents the type of feeding strategy adopted by the sows and the nutritional level of the feed, which is obtained from the pig farm's feeding and management records. Daily feed intake represents the average amount of feed actually ingested by the sows each day, obtained through statistical analysis of feed intake records from automatic feeding systems or by manually weighing and averaging over multiple days. Stocking density represents the sufficiency of space in the pen where the sows are located, calculated from the pen area and the actual number of sows. Pen type represents the form of the feeding space where the sows are located, determined by the pig farm's infrastructure records.

[0068] Environmental data refers to parameters of natural environmental conditions inside and outside the pigsty, reflecting the stress level of the physical environment in which the sows live. These are external influencing factors affected by region, season, and climate, and specifically include indoor temperature, indoor relative humidity, temperature and humidity index, number of heat stress days, and light duration. Among them, indoor temperature represents the air temperature inside the pigsty, and indoor relative humidity represents the humidity level of the air inside the pigsty. Both are obtained by automatically and continuously collecting data from temperature and humidity sensors placed in the pigsty and taking the daily average value. The temperature and humidity index comprehensively reflects the degree of synergistic influence of temperature and humidity on heat stress and is calculated from the synchronously collected temperature and humidity data according to a standard formula. The number of heat stress days represents the cumulative number of days in the period prior to sampling where the temperature and humidity index exceeded a preset threshold. This is obtained from historical temperature and humidity data. Light duration represents the daily light exposure time in the pigsty and is determined by the pig farm's light management system.

[0069] Health status data refers to parameters reflecting the individual health and disease status of sows. It characterizes the degree to which the sow's own health level affects reproductive performance and belongs to the internal influencing factors at the individual level. Specifically, it includes health grade, disease history, immune status, and body condition score. Among them, the health grade represents the overall health status of the sow, which is recorded in the health management file after comprehensive assessment by the pig farm veterinarian or feeder. The disease history represents the sow's past illnesses and is obtained by querying the pig farm veterinary diagnosis and treatment record system or disease prevention and control file. The immune status represents the sow's vaccination status and is determined by the pig farm's immunization record. The body condition score represents the sow's body fat reserves and nutritional status and is obtained by trained technicians through visual observation and touch assessment.

[0070] In S120, several raw AMH eigenvalues ​​include AMH population quantiles, long-term AMH slope, AMH per unit body weight, and the interaction value between AMH and backfat thickness, among which:

[0071] The AMH population quantile is defined as the percentile ranking of the sow's original AMH value within the same age- and breed-specific population, ranging from 0 to 100. This feature reflects the sow's relative position within the population and exhibits strong robustness to extreme values, serving as a supplementary statistical perspective to intrinsic AMH. The specific calculation method is as follows: AMH reference dataset of sows of the same age and breed is pre-established. The original AMH value of the sow to be evaluated is compared with this reference dataset. The proportion of samples in the reference dataset with values ​​lower than the original AMH value is counted out of the total sample size. This proportion is multiplied by 100 to obtain the sow's AMH population quantile.

[0072] If multiple AMH test records exist for the same sow, the long-term slope of AMH is calculated by performing a linear regression with the test time as the x-axis and the AMH value as the y-axis. The regression slope is the long-term slope of AMH, reflecting the trend of AMH changes over time. For gilts with only a single test record, this feature is filled with zero or the population mean. The specific calculation method is as follows: obtain the time points of each AMH test for the same sow and the corresponding AMH test values, and perform a univariate linear regression fitting with the test time (in days) as the independent variable and the AMH test value as the dependent variable. The slope of the obtained regression equation is the long-term slope of AMH, and its unit is ng / mL / day. For gilts with only a single test record, since the slope cannot be calculated, this feature value is filled with zero or the average slope value of the same age and breed population.

[0073] The AMH per unit body weight is calculated by dividing the original AMH value by the sow's body weight, expressed in ng / mL / kg. This feature corrects for body size differences, eliminating the blood dilution effect caused by different body weights and ensuring comparability of AMH levels in sows of different body sizes. The specific calculation method is as follows: obtain the original AMH value and corresponding body weight of the sow to be evaluated, divide the original AMH value by the sow's body weight, and the quotient is the AMH per unit body weight.

[0074] The AMH backfat thickness interaction value is obtained by multiplying the original AMH value by the backfat thickness. This feature characterizes the interaction between fat deposition and endocrine function, reflecting the regulatory effect of body fat status on ovarian reserve, and providing a multidimensional perspective of the fat-endocrine axis for reproductive potential assessment. The specific calculation method is as follows: obtain the original AMH value and corresponding backfat thickness data of the sow to be evaluated, multiply the original AMH value by the backfat thickness, and the resulting product is the AMH backfat thickness interaction value.

[0075] In this embodiment, the AMH population quantile, long-term AMH slope, AMH per unit body weight, and AMH backfat thickness interaction value are selected as the original AMH feature values ​​because a single original AMH value can only reflect the absolute concentration level at the time of detection and is difficult to comprehensively characterize the multidimensional features of ovarian reserve. Specifically, the AMH population quantile, from a relative statistical perspective, uses the percentile ranking of the sow under evaluation within the same age and breed population to characterize the relative position of its AMH level. This effectively avoids interference from extreme values ​​and demonstrates strong robustness to fluctuations in the test results. The long-term slope of AMH, from a time perspective, reflects the dynamic changes in AMH levels through linear regression trends from multiple tests, providing a basis for judging the changing trends of ovarian reserve. For gilts with only a single test record, the slope is filled with zero or the population mean to ensure consistency of characteristic dimensions. The AMH per unit weight, from an individual correction perspective, eliminates the blood dilution effect caused by body size differences through body weight normalization, making the AMH levels of sows of different body sizes comparable. The AMH backfat thickness interaction value, from a metabolic interaction perspective, characterizes the interaction between fat deposition and endocrine function through the product of the original AMH value and backfat thickness, revealing the regulatory effect of body fat status on ovarian reserve. The above four features supplement and expand the original AMH value from four different dimensions: relative position, time trend, body shape correction, and metabolic interaction. They can depict the true state of ovarian reserve in sows from multiple dimensions and levels, providing a more comprehensive and robust numerical feature foundation for subsequent model inference and reproductive potential assessment.

[0076] In this embodiment, several original AMH feature values ​​include intrinsic AMH values. The intrinsic AMH values ​​represent the AMH values ​​after removing the influence of age and variety from the original AMH values. Based on this, the intrinsic AMH values ​​are obtained as follows:

[0077] The step of obtaining the original AMH feature sequence based on the original AMH value and physiological data includes: inputting the age and variety into a pre-established first regression model, outputting an AMH baseline value through the first regression model, wherein the first regression model is used to characterize the correspondence between the age and variety and the AMH baseline value; and obtaining the intrinsic AMH value through the difference between the original AMH value and the AMH baseline value.

[0078] The first regression model in this embodiment is:

[0079]

[0080] in, This is the regression intercept term; The regression coefficient corresponding to age represents the rate of change of AMH level with increasing age. The regression coefficient for the m-th breed represents the difference in AMH baseline levels among different breeds. The breeds are coded using dummy variables, with mainstream breeds (such as Large White pigs) as the reference group and other breeds (such as Landrace and Duroc) having their own corresponding dummy variables.

[0081] In this embodiment, the model coefficients obtained by fitting large sample historical data using the least squares method are shown in the following example: Intercept term 1.20 ng / mL; age coefficient It is 0.008 ng / mL / day; the coefficient corresponding to the Changbai variety is... The value was 0.35 ng / mL; the coefficient corresponding to the Duroc variety was... The value is -0.20 ng / mL. Taking a 170-day-old Large White gilt as an example, its AMH baseline value is calculated as: 1.20 + 0.008 × 170 + 0.35 = 2.91 ng / mL; if the sow's original AMH value is 3.12 ng / mL, then the intrinsic AMH value is 3.12 - 2.91 = 0.21 ng / mL, indicating that its AMH level is higher than the baseline value of 0.21 ng / mL for the same age and breed population.

[0082] In this embodiment, the regression coefficients of the first regression model are obtained as follows: Historical data of a large sample of replacement gilts are collected, covering gilts of different ages and breeds. Each sample includes age, breed information, and corresponding serum AMH values. Breed information is converted into dummy variable codes, with mainstream breeds used as the reference group and other breeds having corresponding dummy variables. The age and breed dummy variables are used as independent variables, and the AMH values ​​as the dependent variable. The least squares method is used to fit and solve the multiple linear regression model, obtaining estimated values ​​for each regression coefficient. After fitting, ten-fold cross-validation is used to verify the model's goodness of fit and generalization ability, ensuring that the model's R² reaches a preset threshold. After successful verification, these coefficients are solidified as the parameters of the first regression model and can be updated periodically based on new data.

[0083] This embodiment uses a first regression model to establish the correspondence between age, breed, and AMH baseline values. This is because sow serum AMH levels are affected by age and breed: on the one hand, ovarian reserve changes regularly with age, and AMH levels show a corresponding trend; on the other hand, different breeds of sows exhibit systematic differences in their baseline AMH levels due to differences in genetic background. If raw AMH values ​​are used directly for assessment, AMH levels between sows of different ages and breeds are not comparable and cannot accurately reflect the true differences in individual ovarian reserve. The regression model can quantitatively characterize the systematic influence of age and breed on AMH levels, establishing a unified benchmark reference system and providing a basis for subsequent cross-individual comparisons.

[0084] In this embodiment, the original AMH value contains systematic biases due to age and breed differences, making it unable to directly represent the individual's true ovarian reserve level relative to the population. By subtracting the baseline AMH value determined by age and breed from the original AMH value, the intrinsic AMH value is obtained, removing the influence of age and breed. It reflects the degree of deviation of the sow from the average level of the same age and breed population, i.e., the individual-specific ovarian reserve difference. An intrinsic AMH value greater than zero indicates that the sow's ovarian reserve is better than the population average, while a value less than zero indicates that it is worse than the population average. This allows for cross-sectional comparisons of sows of different ages and breeds under a unified benchmark, eliminating the interference of systematic confounding factors and providing more accurate and comparable feature inputs for subsequent reproductive potential assessment.

[0085] In S130, the ovarian parenchymal area is obtained as follows: The ovarian ultrasound image is input into a preset image processing tool, and semantic segmentation is performed on the image to obtain a binary mask of the ovarian parenchymal region. In the mask, the area with a pixel value of 1 corresponds to the ovarian parenchyma, and the area with a pixel value of 0 corresponds to the background. The total number of pixels in the ovarian parenchymal region in the binary mask is counted, and combined with the preset scale parameter of the ultrasound probe (i.e., the actual physical size corresponding to each pixel), the pixel area is converted into the actual physical area in square millimeters, which is the ovarian parenchymal area.

[0086] The number of follicles is obtained as follows: semantic segmentation of the ovarian ultrasound image is performed using a preset image processing tool to obtain a binary mask of the follicular cavity region; connected component analysis is performed on the binary mask of the follicular cavity to identify each independent connected region of the follicular cavity; a minimum area threshold is set to filter out noise regions or artifacts with an area smaller than the threshold, and the number of remaining effective connected regions is counted, which is the number of follicles.

[0087] The average diameter of follicles is obtained as follows: Based on the analysis of connected follicle regions, the equivalent diameter of each effective follicle region is calculated, which is the diameter of a circle with the same area as the region. The equivalent diameters of all effective follicles are then arithmetically averaged, and the resulting average value is the average diameter of follicles, in millimeters.

[0088] The preset image processing tool can be a lightweight semantic segmentation neural network, such as a U-shaped segmentation network built on a lightweight backbone network like MobileNet. It has a small number of parameters, fast inference speed, and is suitable for real-time front-end processing. In addition, the image processing tool can also be an image analysis process based on traditional image processing methods such as threshold segmentation, edge detection, and morphological opening and closing operations.

[0089] In S140, serum AMH testing can be affected by various factors such as sampling procedures, reagent batches, and experimental conditions, leading to random errors or systematic biases in the measured values. Directly using the biased raw AMH values ​​and their derived features for subsequent evaluation will affect the accuracy of the final results. However, morphological parameters such as ovarian parenchymal area, follicle number, and average follicle diameter have a natural physiological correlation with AMH levels—AMH is secreted by follicular granulosa cells; the more follicles and the richer the ovarian parenchyma, the higher the AMH secretion level is usually. Therefore, morphological parameters can independently reflect ovarian reserve status from another dimension. Cross-validation and correction of raw AMH values ​​using morphological parameters can identify and reduce the impact of detection errors on the AMH feature sequence, improving the reliability and robustness of numerical features. Simultaneously, the corrected AMH values ​​need to be updated synchronously with each derived AMH feature value to ensure the consistency of the entire feature sequence, providing a more accurate numerical basis for subsequent image enhancement guidance and model inference.

[0090] Based on this, such as Figure 2 As shown, S140 specifically includes the following steps:

[0091] S141: Input the ovarian parenchymal area, number of follicles and average follicle diameter into the second regression model, and obtain the AMH predicted value through the second regression model. The second regression model represents the correspondence between the ovarian parenchymal area, number of follicles and average follicle diameter and the AMH predicted value.

[0092] S142: Calculate the deviation rate between the original AMH value and the predicted AMH value;

[0093] S143: When the deviation rate is less than or equal to a preset value, the original AMH value and the AMH prediction value are weighted to obtain the corrected AMH value;

[0094] S144: Based on the corrected AMH value, obtain the corrected AMH feature sequence.

[0095] The second regression model in this embodiment is a multiple linear regression model, used to characterize the correspondence between ovarian parenchymal area, number of follicles, and average follicle diameter and AMH predicted values. Its expression is as follows:

[0096]

[0097] in, This is the intercept term of the second regression model; The regression coefficient corresponding to the number of follicles represents the average contribution of a single follicle to serum AMH levels. The regression coefficient corresponding to the ovarian parenchymal area represents the effect of ovarian tissue volume on AMH secretion. The regression coefficients corresponding to the average follicle diameter represent the regulatory effect of follicle size differences on AMH secretion levels. All coefficients were obtained by least-squares fitting of large-sample paired data (simultaneous acquisition of ovarian ultrasound images and serum AMH levels), and cross-validation confirmed that the model's goodness of fit met the requirements.

[0098] In S142 and S143, the specific calculation process of the deviation rate is as follows: the absolute difference between the original AMH value and the predicted AMH value is divided by the predicted AMH value, and then multiplied by 100% to obtain the deviation rate, which is used to quantify the degree of deviation between the serum test result and the morphological estimation result. When the deviation rate is less than or equal to the preset value, it indicates that the serum AMH test value and the reserve level reflected by the ovarian morphological parameters are in good consistency, and both are within the confidence range. At this time, the original AMH value and the predicted AMH value are weighted and fused. The original AMH value, as a direct test result, is given a higher weight, and the predicted AMH value, as an independent morphological verification, is given a lower weight. The weighted fusion further reduces the random noise in the detection process, making the corrected AMH value closer to the true level. The reason for adopting this processing method is that there is a natural physiological correlation between morphological parameters and AMH levels. When the deviation between the two is small, morphological estimation can provide cross-validation information independent of serum testing. Weighted fusion can achieve complementarity between the two types of information and improve the robustness of numerical features. When the deviation rate is greater than the preset value, it indicates that there is a significant contradiction between the detected value and the morphological estimate, suggesting that there may be abnormal detection or special physiological state. At this time, the original AMH value is not forcibly corrected, but is retained as the corrected AMH value. At the same time, an abnormal label of detection-morphological contradiction is attached, marking the sample as a suspicious sample that needs to be focused on, for subsequent manual review or further detection and confirmation, so as to avoid introducing a larger systematic deviation due to blind correction.

[0099] In this embodiment, the coefficients of the fitted second regression model are shown in the following example: intercept term It was 0.05 ng / mL; follicle count coefficient The concentration was 0.10 ng / mL per ovary; ovarian parenchymal area coefficient It is 0.003 ng / mL / mm²; the average follicle diameter coefficient is 0.003 ng / mL / mm². The value is 0.20 ng / mL / mm. Taking a gilt with 18 follicles, an ovarian parenchymal area of ​​320 mm², and an average follicle diameter of 2.8 mm as an example, its AMH predicted value is calculated as: 0.05 + 0.10 × 18 + 0.003 × 320 + 0.20 × 2.8 = 3.37 ng / mL. If the original AMH value of this gilt is 3.12 ng / mL, then the deviation rate is |3.12 - 3.37| / 3.37 × 100% ≈ 7.4%, which is less than the preset threshold (e.g., 15%), indicating that the detected value is in good agreement with the morphological estimate, and the corrected AMH value can be obtained through weighted fusion.

[0100] The weighted processing method is linear weighted fusion, and the specific expression is:

[0101]

[0102] in, These are the weighting coefficients corresponding to the original AMH values. Let be the weighting coefficients corresponding to the AMH predicted values, and satisfy . + =1. and These are the original AMH value and the predicted AMH value, respectively.

[0103] In this embodiment, the raw AMH value is used as the direct serum test result, which has relatively higher reliability, therefore its weight is... The weight is set to 0.7; the AMH predicted value is used as the morphologically independent estimation result to assist in the correction of the detected value, therefore the weight is... The weight is set to 0.3. The reason for this weighting is that the raw AMH value is direct experimental data and the primary basis for evaluation, thus it should be given a higher weight to ensure the objectivity of the results. The morphological prediction value, on the other hand, provides cross-validation information independent of serum testing; assigning it a lower weight allows it to fulfill its corrective role, reduce random detection errors, and avoid excessive interference from the uncertainty of morphological estimation itself, ensuring the robustness of the correction process. For example, when the raw AMH value is 3.12 ng / mL, the predicted AMH value is 3.37 ng / mL, and the deviation rate meets the preset conditions, the corrected AMH value is calculated as: 0.7 × 3.12 + 0.3 × 3.37 = 3.195 ng / mL.

[0104] Based on this, in S144, if the corrected AMH value is equal to the original AMH value, then the corrected AMH feature sequence is equal to the original AMH feature sequence. When the deviation rate is less than or equal to the preset value, the corrected AMH value is a new value obtained by weighted fusion of the original AMH value and the predicted AMH value. At this time, based on the corrected AMH value, each AMH derived feature value is recalculated according to the same calculation method as the original AMH feature sequence to obtain the corrected AMH feature sequence. Specifically, the original AMH value used in the calculation of the original AMH feature sequence is uniformly replaced with the corrected AMH value, and the intrinsic AMH value, AMH population quantile, AMH per unit weight, and AMH backfat thickness interaction value are recalculated to ensure that the corrected feature sequence is consistent with the corrected AMH value and to avoid numerical mismatch between features, thereby providing an internally consistent numerical feature basis for subsequent image enhancement guidance and model inference.

[0105] In S150, the texture features of ovarian ultrasound images are closely related to the follicle size distribution, and there are significant differences in follicle size distribution corresponding to different AMH levels. Sows with higher AMH levels typically have a large number of small antral follicles, making the texture information of these small follicles in the image more crucial; sows with lower AMH levels have fewer follicles but larger sizes, making the contour features of the follicles in the image more important. If a general image enhancement method is used, it is impossible to differentiate the enhancement based on the follicle distribution characteristics under different AMH levels, which may lead to insufficient enhancement of key scale features and excessive amplification of noise at non-key scales. By using corrected AMH values ​​and related feature values ​​to guide the image enhancement process, the enhancement strategy can be adaptively adjusted according to the individual's AMH level, specifically enhancing the follicle texture or contour features at the corresponding scale, improving the recognition of effective information in the image, providing higher-quality image input for subsequent feature extraction and inference by deep learning models, and thus improving the accuracy of reproductive potential assessment.

[0106] Based on this, S150 is specifically as follows:

[0107] Based on the corrected AMH value and the specified corrected AMH feature value, enhancement parameters are obtained according to a specified enhancement rule. The specified enhancement rule includes a mapping between the first range of the corrected AMH value and the second range of the specified corrected AMH feature value and the enhancement parameters. The enhancement parameters include enhancement weight, a first Gaussian standard deviation, and a second Gaussian standard deviation, wherein the first Gaussian standard deviation is less than the second Gaussian standard deviation. The ovarian ultrasound image is subjected to Gaussian blurring using the first Gaussian standard deviation as a parameter to obtain a first blurred image. The ovarian ultrasound image is then subjected to Gaussian blurring using the second Gaussian standard deviation as a parameter to obtain a second blurred image. The first blurred image and the second blurred image are subtracted pixel by pixel to obtain a Gaussian difference image. The gray value of each pixel in the ovarian ultrasound image is multiplied by the enhancement weight and the corresponding Gaussian difference value to obtain the enhanced gray value of that pixel. The Gaussian difference value is obtained from the Gaussian difference image. The enhanced gray value is truncated to ensure that the gray value is within a pre-set valid value range, and finally, the enhanced image is obtained.

[0108] In this embodiment, the specified corrected AMH eigenvalue is the corrected AMH population quantile. For example, the specified enhancement rule can be represented in the following table.

[0109] Table 7 Specified Enhancement Rules Table

[0110]

[0111] The high AMH setting employs a small-scale texture enhancement strategy, setting relatively small first and second Gaussian standard deviations with an enhancement weight of 0.3, focusing on enhancing the texture information of small follicles in the image; the medium AMH setting employs a balanced enhancement strategy, setting medium-sized first and second Gaussian standard deviations with an enhancement weight of 0.2, balancing the enhancement of texture details and contour features; and the low AMH setting employs a large-scale contour enhancement strategy, setting relatively large first and second Gaussian standard deviations with an enhancement weight of 0.3, focusing on enhancing the contour features of follicles.

[0112] The reason for adopting the above-mentioned differentiated enhancement strategy is that there are significant physiological differences in the distribution of ovarian follicles corresponding to different AMH levels: when the AMH level is high, there are many small antral follicles in the ovary, and the key information for evaluation lies in the texture details of small follicles. Therefore, small-scale enhancement is used to strengthen fine texture features; when the AMH level is moderate, the follicle size distribution is relatively uniform, and information at all scales has evaluation value. Therefore, balanced enhancement is used to take into account feature extraction at different scales; when the AMH level is low, there are fewer follicles, but the individual size is larger. The key information for evaluation lies in the contour morphology and number statistics of follicles. Therefore, large-scale enhancement is used to strengthen contour features. Through an enhancement strategy that adaptively matches the AMH level, the recognition of key information at each level can be improved in a targeted manner, avoiding the problems of insufficient enhancement of key features or excessive amplification of non-key noise caused by general enhancement methods, thus providing higher-quality image input for subsequent model inference.

[0113] In this embodiment, Gaussian blurring can filter out high-frequency details smaller than the Gaussian kernel scale in the image. The larger the Gaussian standard deviation, the more details are filtered out, and the more blurred the image becomes. Subtracting the Gaussian blur results with two different standard deviations is equivalent to constructing a bandpass filter, retaining only the texture information between the two scales and filtering out noise at smaller scales and background at larger scales. The first Gaussian standard deviation defines the minimum scale boundary for enhancement, and the second Gaussian standard deviation defines the maximum scale boundary for enhancement. The difference between the two determines the range of texture scales extracted and enhanced. By multiplying the Gaussian difference result by the enhancement weight and then superimposing it back onto the original image, targeted enhancement of texture features within a specific scale range is achieved, thereby improving the recognizability of follicle structures at the corresponding scale.

[0114] For example, S150 can be:

[0115] Taking gilts as an example, their corrected AMH value was 3.195 ng / mL, and their AMH population quantile was 58th percentile. According to the enhancement rules, they were determined to be in the medium AMH range, with the corresponding enhancement parameters being: enhancement weight 0.2, first Gaussian standard deviation 1.5, and second Gaussian standard deviation 3.0. A 3×3 pixel region near the edge of the follicle in the original B-ultrasound image was selected, with the original gray values ​​of 100 for the left ovarian parenchyma region and 150 for the right follicular cavity region. The original images were Gaussian blurred using the first Gaussian standard deviation of 1.5 and the second Gaussian standard deviation of 3.0, respectively, to obtain a first blurred image and a second blurred image. The gray value of the edge transition region in the first blurred image was approximately 112, and the gray value of the edge transition region in the second blurred image was approximately 118. The first blurred image and the second blurred image were subtracted pixel by pixel to obtain a Gaussian difference image, with a difference value of approximately -6 on the left side of the edge and approximately +6 on the right side of the edge.

[0116] The enhanced grayscale value is obtained by multiplying the grayscale value of each pixel in the original image by the enhancement weight and the corresponding Gaussian difference value. Taking the center pixel on the left side of the edge as an example, the original grayscale value is 100, the Gaussian difference is -6, and the enhancement weight is 0.2, so the enhanced grayscale value is calculated as 100 + 0.2 × (-6) = 98.8. Taking the center pixel on the right side of the edge as an example, the original grayscale value is 150, the Gaussian difference is +6, so the enhanced grayscale value is calculated as 150 + 0.2 × 6 = 151.2. The enhanced grayscale values ​​are all within the effective range of 0 to 255, and no truncation is required. After enhancement, the grayscale difference between the two sides of the edge is increased from the original 50 to 52.4, the edge contrast is improved by about 4.8%, and the follicle contour texture features are effectively enhanced.

[0117] In S160, the corrected AMH feature sequence and the enhanced image are input into a pre-trained deep learning model, which outputs the initial predicted number of offspring and the initial probability of each preset reproductive potential level.

[0118] The deep learning model in this embodiment is a dual-branch multimodal fusion deep learning model, comprising two parallel branches: a numerical processing branch and an image processing branch. These branches extract features from the corrected AMH feature sequence and the enhanced ovarian ultrasound image, respectively. A fusion module then integrates the features from the numerical and image modalities. Finally, the output layer simultaneously outputs the initial predicted litter size and the initial probabilities of each reproductive potential level. This type of model can simultaneously utilize structured numerical information and unstructured image information, fully exploring the complementary relationship between the two modalities. Compared to single-modal models, it has stronger feature representation capabilities and prediction accuracy, making it suitable for comprehensive evaluation scenarios involving multimodal data.

[0119] The basic architecture of the deep learning model consists of four parts: a numerical branch, an image branch, a feature fusion layer, and dual output heads. The numerical branch adopts a multilayer perceptron structure, which is composed of several fully connected layers stacked together. Each layer is configured with batch normalization and activation functions. The input is the corrected AMH feature sequence, and the output is a fixed-dimensional numerical feature vector. The image branch adopts a lightweight convolutional neural network structure, which is composed of several convolutional layers, pooling layers, and residual modules stacked together. The input is the enhanced B-ultrasound image, and the output is a fixed-dimensional image feature vector. The feature fusion layer uses concatenation fusion or attention fusion to fuse the numerical feature vector with the image feature vector to obtain a fused feature vector. The dual output heads include a regression output head and a classification output head. The regression output head outputs the initial predicted number of offspring, and the classification output head outputs the initial probability of each reproductive potential level after softmax activation.

[0120] The deep learning model can be trained using supervised training, with historical data from multiparous sows as training samples. Each sample includes an AMH feature sequence, an ovarian ultrasound image, and a label for the actual number of piglets born. During training, the AMH feature sequence and ultrasound image are input into the model, and forward propagation is used to obtain the predicted number of piglets born and the grading probability. The mean squared error between the predicted number of piglets born and the actual number of piglets born is used as the regression loss, and the cross-entropy between the grading probability and the actual grading is used as the classification loss. The two losses are weighted and summed as the total loss function. A gradient descent optimization algorithm (such as the Adam optimizer) is used for backpropagation to iteratively update the network parameters. Ten-fold cross-validation can be used to evaluate the model's generalization ability during training, and an early stopping strategy is used to prevent overfitting. After training, the model parameters are saved for direct use in the inference stage.

[0121] In this embodiment, the preset reproductive potential levels are divided into five levels: S, A, B, C, and D. S level indicates excellent reproductive potential, superb ovarian reserve, and a significantly higher expected litter size than the population average, belonging to the core population retention level; A level indicates good reproductive potential, sufficient ovarian reserve, and a litter size expected higher than the population average, suitable for normal reproductive production; B level indicates acceptable reproductive potential, adequate ovarian reserve, and a litter size expected at the population average, suitable for routine feeding and management; C level indicates poor reproductive potential, low ovarian reserve, and a litter size expected lower than the population average, requiring optimized feeding conditions and consideration of shortening the usage period; D level indicates a reproductive potential warning, severely insufficient ovarian reserve, and a litter size expected significantly lower than the population average, recommending culling. The levels are arranged in descending order of reproductive performance, and the initial probability is the probability distribution of the model for each of the five levels, with the sum of the probabilities of the five levels being 1.

[0122] In S170, such as Figure 3 As shown, the specific steps include the following:

[0123] S171: Based on the feeding environment data, environmental data, and health status data, the current feeding level index, the current environmental stress index, and the current health index are obtained, wherein the current feeding level index, the current environmental stress index, and the current health index are all greater than 0 and less than 1.

[0124] S172: Input the current feeding level index, current environmental stress index and current health index into the third regression model, and calculate the current confounding effect size through the third regression model. The current confounding effect size represents the contribution of the current feeding conditions corresponding to the current feeding level index, current environmental stress index and current health index to the initial predicted number of piglets.

[0125] S173: Input the pre-set baseline feeding level index, baseline environmental stress index, and baseline health index into the third regression model, and calculate the baseline confounding effect size through the third regression model;

[0126] S174: Calculate the difference between the initial predicted number of offspring and the current confounding effect;

[0127] S175: Calculate the sum of the difference and the baseline confounding effect to obtain the final predicted litter size, thereby correcting the initial predicted litter size.

[0128] In S171, the calculation process for the feeding level index is as follows: First, the four feeding environment parameters—feeding program, daily feed intake, stocking density, and pen type—are quantified and normalized. The feeding program is a categorical parameter, mapped to different scores according to feeding standards from low to high. For example, restricted feeding is mapped to 0.3, regular free-feeding to 0.7, and phased optimal feeding to 1.0. Daily feed intake is a numerical parameter, using a min-max normalization method, with a linear mapping of the minimum daily feed intake as 0 and the maximum daily feed intake as 1. Stocking density is a numerical parameter, using inverse normalization; higher density results in a lower score, with a linear mapping of the optimal density as 1 and the maximum tolerated density as 0. Pen type is a categorical parameter, with gestation pens mapped to 0.5, semi-gestation pens to 0.7, and large-group pens to 1.0. Subsequently, the normalized scores of the four items are weighted and summed according to the preset weights. The weights can be set according to the degree of influence of each parameter on reproductive performance. For example, the weight of feeding program is 0.3, the weight of daily feed intake is 0.4, the weight of stocking density is 0.15, and the weight of pen type is 0.15. Finally, a feeding level index with a value range of 0 to 1 is obtained. The higher the value, the better the feeding management conditions and the more sufficient the nutrition supply.

[0129] The calculation process for the environmental stress index is as follows: First, the three environmental parameters—temperature and humidity index, number of heat stress days, and light duration—are quantified and normalized. The temperature and humidity index is a numerical parameter, and min-max normalization is used, with a linear mapping based on the lower limit of the suitable temperature and humidity range being 0 and the upper limit of high temperature and humidity stress being 1. The higher the temperature and humidity index, the higher the score. The number of heat stress days is a numerical parameter, and the normalized value is the proportion of heat stress days in the 30 days prior to sampling to the total number of days. The more days, the higher the score. The light duration is a numerical parameter, and the normalization is based on the degree of deviation between the actual light duration and the suitable light duration. The greater the deviation, the higher the score. Subsequently, the normalized scores of the three items are weighted and summed according to the preset weights. The weights can be set according to the influence of each parameter on the stress level. For example, the weight of temperature and humidity index is 0.5, the weight of heat stress days is 0.35, and the weight of light duration is 0.15. Finally, the environmental stress index with a value range of 0 to 1 is obtained. The higher the value, the more severe the stress caused by environmental factors to sows.

[0130] The calculation process for the health index is as follows: First, the four health parameters—health level, medical history, immune status, and physical condition score—are quantified and normalized. Health level is a graded parameter, mapped to different scores of 0.2, 0.5, 0.8, and 1.0, respectively, from poor to excellent. Medical history is a categorical parameter, with 1.0 for no reproductive-related disease history, 0.7 for a history of mild disease, and 0.3 for a history of severe reproductive disease. Immune status is a categorical parameter, with 1.0 for complete reproductive-related vaccinations, 0.6 for partial vaccinations, and 0.3 for severe vaccinations. Physical condition score is a numerical parameter, based on a suitable physical condition (3 points), with lower scores for deviations from the suitable range, using a distance normalization method. Subsequently, the normalized scores of the four items are weighted and summed according to the preset weights. The weights can be set according to the degree of influence of each parameter on reproductive performance. For example, the weight of health level is 0.3, the weight of disease history is 0.35, the weight of immune status is 0.15, and the weight of body condition score is 0.2. Finally, a health index with a value range of 0 to 1 is obtained. The higher the value, the better the overall health of the sow.

[0131] The specific calculation process of the feeding level index by the electronic device is as follows: First, it reads the feeding environment parameters such as the feeding plan, daily feed intake, stocking density, and pen type corresponding to the sow to be evaluated from the pig farm feeding management system; then, it calls the preset normalization mapping rules to quantify each parameter respectively—the categorical parameters such as feeding plan and pen type are converted into scores in the range of 0 to 1 according to the preset mapping table, and the numerical parameters such as daily feed intake and stocking density are linearly mapped to the range of 0 to 1 according to the min-max normalization formula; next, it reads the pre-configured weight coefficients from the storage unit, and multiplies the normalized score of each parameter by the corresponding weight; finally, it sums up the weighted scores to obtain the current feeding level index with a value range of 0 to 1, and the higher the value, the better the feeding management conditions.

[0132] The specific calculation process of the electronic device for the environmental stress index is as follows: First, it reads environmental data such as the temperature and humidity index, the number of heat stress days in the 30 days prior to sampling, and the average daily light duration of the pen where the sow to be evaluated is located from the pig house environmental monitoring system; then, it calls the preset normalization calculation rules to quantify each parameter respectively—for numerical parameters such as temperature and humidity index and number of heat stress days, it linearly maps them to the 0 to 1 range according to the stress level from low to high; for the light duration parameter, it calculates the deviation based on the appropriate light duration, and the larger the deviation, the higher the normalization score; next, it reads the pre-configured weight coefficients from the storage unit, and multiplies the normalized score of each parameter by the corresponding weight; finally, it sums up the weighted scores to obtain the current environmental stress index with a value range of 0 to 1, and the higher the value, the more severe the environmental stress.

[0133] The specific calculation process for the health index by the electronic device is as follows: First, it reads health status data such as the health level, disease history, immune status, and body condition score of the sow to be evaluated from the pig farm health management system; then, it calls the preset normalization mapping rules to quantify each parameter—the level or category parameters such as health level, disease history, and immune status are converted into scores in the range of 0 to 1 according to the preset mapping table, and the numerical parameters such as body condition score are normalized based on suitable body condition, with the score decreasing as the deviation from the suitable range increases; next, it reads the pre-configured weight coefficients from the storage unit and multiplies the normalized score of each parameter by the corresponding weight; finally, it sums up the weighted scores to obtain the current health index, which ranges from 0 to 1, with a higher value indicating a better health condition.

[0134] Taking gilts as an example, the electronic device reads the corresponding feeding environment data of the sow from the pig farm feeding management system: the feeding plan is free access, the daily feed intake is 2.5 kg, the stocking density is 1.2 heads per square meter, and the pen type is large pen feeding; after calling the normalized mapping rule, the feeding plan score is 0.7, the daily feed intake score is 0.9, the stocking density score is 0.85, and the pen type score is 1.0; after reading the preset weight coefficient, the scores of each item are multiplied by the corresponding weight and accumulated to obtain the current feeding level index of 0.85, indicating that the feeding conditions of the sow are relatively good.

[0135] Taking gilts as an example, the electronic device reads the environmental data of the pen where the gilt is located from the pig house environmental monitoring system: the average temperature and humidity index for the 30 days before sampling is 68, the number of heat stress days is 2, and the average daily light duration is 14 hours; after calling the normalization calculation rules, the temperature and humidity index score is 0.2, the heat stress day score is 0.07, and the light duration score is 0.1; after reading the preset weight coefficients, the scores of each item are multiplied by the corresponding weights and accumulated to obtain the current environmental stress index of 0.15, indicating that the stress level of the environment in which the gilt is located is relatively mild.

[0136] Taking gilts as an example, the electronic device reads the gilt's health status data from the pig farm's health management system: the health level is excellent, there is no history of reproductive-related diseases, all vaccines are complete, and the body condition score is 3.5. After calling the normalized mapping rule, the health level score is 1.0, the disease history score is 1.0, the immunization status score is 1.0, and the body condition score is 0.85. After reading the preset weight coefficients, each score is multiplied by its corresponding weight and summed to obtain the current health index of 0.95, indicating that the gilt is in good health.

[0137] The third regression model in this embodiment is a multiple linear regression model, used to characterize the correspondence between the feeding level index, environmental stress index, health index, and confounding effect size. Its expression is:

[0138]

[0139] in, This is the intercept term of the third regression model; The regression coefficient corresponding to the feeding level index represents the degree of influence of feeding management conditions on the initial predicted number of piglets. It is usually a positive value, that is, the better the feeding conditions, the higher the model prediction value. The regression coefficient corresponding to the environmental stress index represents the degree of influence of environmental stress on the initial predicted number of offspring. It is usually a negative value, that is, the more severe the stress, the lower the model prediction value. The regression coefficients corresponding to the health index represent the degree of influence of health status on the initial predicted number of offspring. They are typically positive, meaning the better the health status, the higher the model's predicted value. All the above coefficients were obtained by fitting a large sample of historical data using the least squares method, and cross-validation confirmed that the model's goodness of fit met the requirements.

[0140] In this embodiment, the coefficients of the fitted third regression model are shown below: Intercept term 2.0 heads; Feeding level index coefficient +3.0 heads; Environmental stress index coefficient -2.5 heads; Health Index Coefficient The result is +2.0 heads. Taking gilts as an example, their feeding level index is 0.85, environmental stress index is 0.15, and health index is 0.95. The current confounding effect is calculated as: 2.0 + 3.0 × 0.85 + (-2.5) × 0.15 + 2.0 × 0.95 = 2.0 + 2.55 - 0.375 + 1.9 = 6.075 heads. This indicates that under the current feeding, environmental, and health conditions, the contribution of external factors to the initial predicted litter size is approximately 6.075 heads.

[0141] The current confounding effect size characterizes the contribution of three external factors—feeding level, environmental stress, and health status—to the initial predicted litter size. In other words, it represents the increase or decrease in the initial predicted value due to external non-genetic factors. This value does not reflect the sow's intrinsic reproductive potential, but rather the combined impact of external conditions on the model's prediction results—superior feeding and management conditions will increase the predicted value, strong environmental stress will decrease it, and good health will raise it. The confounding effect size is the net impact of these three external factors on the prediction result.

[0142] In this embodiment, the initial predicted litter size output by the deep learning model is an apparent composite value, which includes both the sow's own genetic reproductive potential and the influence of external factors such as feeding, environment, and health. Directly using this initial prediction to assess reproductive potential could easily misjudge the quality of external conditions as differences in the sow's inherent abilities, leading to unfair and inaccurate assessment results. By quantifying the confounding effect size, the contribution of external factors to the initial prediction can be clearly separated. This provides a basis for subsequently removing external influences from the initial prediction to obtain a corrected value reflecting the sow's true intrinsic reproductive potential, thereby improving the objectivity and accuracy of reproductive potential assessment.

[0143] In this embodiment, the pre-set baseline feeding level index, baseline environmental stress index, and baseline health index are input into the third regression model to calculate the baseline confounding effect size. This baseline value represents the degree of contribution of external factors to the predicted value under the unified baseline conditions. Next, the difference between the initial predicted litter size and the current confounding effect size is calculated. This difference is the predicted value after removing the influence of the current external conditions, reflecting the prediction level after removing the interference of the current external factors. Finally, the above difference is added to the baseline confounding effect size to obtain the final predicted litter size, thereby correcting the predicted value from the current actual conditions to the unified baseline conditions, completing the correction of the initial predicted litter size.

[0144] In this embodiment, the initial predicted litter size is the apparent prediction output by the deep learning model under the current actual feeding, environmental, and health conditions of the sows. This prediction is influenced by external factors, and since different sows face different external conditions, the initial predicted values ​​lack fair horizontal comparability. By subtracting the current confounding effect size, the influence of the sows' current external conditions is removed, resulting in a basic predicted value that only reflects their own potential. Adding the baseline confounding effect size maps all sows to the same unified baseline condition for evaluation, thus eliminating the unfairness caused by differences in external conditions. This counterfactual correction method unifies the predicted values ​​under different external conditions to the same baseline space, ensuring that the final prediction truly reflects the sow's intrinsic reproductive potential level, rather than the quality of external feeding conditions.

[0145] Taking gilts as an example, the electronic device first calculates the current feeding level index (0.85), current environmental stress index (0.15), and current health index (0.95) based on the gilt's feeding environment data, environmental data, and health status data. These three indices are then input into the third regression model, resulting in a current confounding effect size of 6.075 heads, representing the contribution of the current feeding conditions to the initial predicted litter size. Subsequently, the pre-set baseline feeding level index (0.6), baseline environmental stress index (0.3), and baseline health index (0.8) are input into the third regression model, resulting in a baseline confounding effect size of 4.65 heads, representing the contribution of external factors to the predicted value under a unified baseline condition. If the sow's initial predicted litter size is 13.2 piglets, first calculate the difference between the initial predicted litter size and the current confounding effect: 13.2 - 6.075 = 7.125 piglets, obtaining the baseline predicted value after removing the influence of current external conditions. Then calculate the sum of this difference and the baseline confounding effect: 7.125 + 4.65 = 11.775 piglets, thus obtaining the final predicted litter size of 11.775 piglets. Through the above correction process, the initial predicted value is corrected from 13.2 piglets under the current favorable conditions to 11.775 piglets under unified baseline conditions, removing the predictive bonus brought by good feeding conditions, thereby obtaining a final predicted result that more accurately reflects the sow's intrinsic reproductive potential.

[0146] Specifically, S180 can be:

[0147] The final predicted litter size, the preset reproductive potential level threshold, and the initial probability are input into the Newton iterative model, which is based on Newton's method and outputs the first prediction standard deviation. Based on the final predicted litter size, the initial predicted litter size, and the first prediction standard deviation, the first intermediate probability is determined based on the cumulative distribution function of the standard normal distribution.

[0148] In this embodiment, the preset level threshold for reproductive potential corresponds to the grading boundary parameter used to divide each reproductive potential level during the iteration process, and is the interval basis for calculating the theoretical probability of each level; the initial probability corresponds to the initial five-level probability distribution used as the target reference during the iteration process, where the A-level probability serves as the target value of the error function, driving the iteration to converge in the correct direction; the final predicted litter size is one of the input parameters of the Newton iteration model, and in the iterative calculation, it serves as the mean parameter μ0 of the normal distribution, together with the level threshold, to determine the theoretical probability of each level interval; the Newton iteration model corresponds to the complete iterative solution process based on Newton's method, specifically: firstly, the initial guess value of the first prediction standard deviation (hereinafter referred to as σ) is set as the iteration starting point. The process then substitutes the current σ value and the initial mean μ0 into the normal distribution formula to calculate the z-score corresponding to each level threshold. The theoretical probability of each level is obtained using the cumulative distribution function of the standard normal distribution. Next, the difference between the theoretical probability of level A and the initial probability of level A is taken as the error function value. Then, the derivative of the error function with respect to σ, i.e., the slope of the tangent line, is calculated, reflecting the degree of influence of small changes in σ on the level A probability. Subsequently, the σ value is updated according to Newton's iteration formula, i.e., approximating the point where the error is zero along the tangent direction. Finally, it is determined whether the difference between the old and new σ values ​​is less than the preset convergence threshold. If it is, the iteration stops; otherwise, it returns to recalculation until convergence. Sufficient computational accuracy is usually achieved after 3 to 5 iterations. The first predicted standard deviation corresponds to the σ value obtained after the iteration process converges, i.e., the standard deviation of the predicted normal distribution of litter size.

[0149] The final predicted litter size is used as the new mean μ of the normal distribution in the calculation, which is the central location parameter of the probability distribution. The initial predicted litter size is used as the initial mean in the previous stage when Newton iteratively solves for the first predicted standard deviation, providing the basis for the standard deviation parameter in the solution of the first intermediate probability. The first predicted standard deviation is used as the standard deviation σ of the normal distribution, characterizing the dispersion of the probability distribution. The cumulative distribution function based on the standard normal distribution corresponds to the specific probability calculation process: that is, using the final predicted litter size as the new mean μ and the first predicted standard deviation σ to construct a corrected normal distribution, for each A threshold for the reproductive potential level is preset, and the corresponding z-score is calculated, which is the difference between the threshold and the mean divided by the standard deviation. The cumulative probability corresponding to each threshold is calculated using the cumulative distribution function Φ(z) of the standard normal distribution. The cumulative probability of each level is obtained by subtracting the cumulative probability of the lower bound from the cumulative probability of the upper bound. The probability of level S is 1 minus the cumulative probability corresponding to the lower bound threshold of level S, and the probability of level D is the cumulative probability corresponding to the lower bound threshold of level D. The first intermediate probability is determined to correspond to the five-level probability distribution obtained after the above calculation, which is the probability of each reproductive potential level after mean correction.

[0150] In this embodiment, the initial five-level probability distribution output by the deep learning model is essentially the area of ​​probability across the five reproductive potential level intervals when the predicted litter size follows a normal distribution. The sharpness of the distribution implies the uncertainty of the prediction—a more concentrated probability distribution indicates a more certain model and a smaller standard deviation, while a flatter probability distribution indicates a more uncertain model and a larger standard deviation. The first predicted standard deviation is derived from the initial probabilities using Newton's iteration method, which is equivalent to restoring the discrete classification probability distribution to continuous normal distribution parameters, extracting the inherent uncertainty of the model's prediction for that sample. Counterfactual confounding correction modifies the mean of the predicted litter size, that is, correcting from the initial predicted litter size to the final predicted litter size, while the prediction uncertainty, i.e., the standard deviation σ, remains unchanged because external feeding, environmental, and health factors mainly affect the average level of the litter size rather than the predicted fluctuation range. Therefore, by using the corrected final predicted number of offspring as the new mean and the inversely derived first prediction standard deviation as the standard deviation, a normal distribution is reconstructed. Then, by calculating the probability area of ​​this distribution in each level interval through the cumulative distribution function of the standard normal distribution, the five-level probability distribution after mean correction, i.e. the first intermediate probability, is obtained. This achieves the overall translation of the probability distribution with mean correction, ensuring that the probability distribution is consistent with the corrected predicted value.

[0151] In this embodiment, as Figure 4 As shown, S190 may specifically include the following steps:

[0152] S191: Input the ovarian parenchymal area, number of follicles and average follicle diameter into the fourth regression model, and output the predicted number of litters through the fourth regression model. The fourth regression model represents the correspondence between the ovarian parenchymal area, number of follicles and average follicle diameter and the predicted number of litters.

[0153] S192: Calculate the second intermediate probability of the preset reproductive potential level by combining the predicted number of offspring, the first intermediate probability, and the pre-calibrated second prediction standard deviation;

[0154] S193: Based on the first intermediate probability and the pre-calculated numerical reliability coefficient, a first mass value is obtained; based on the second intermediate probability and the image reliability coefficient, a second mass value is obtained; and based on the prior probability and the preset prior reliability coefficient, a third mass value is obtained, wherein the image reliability coefficient is equal to the ratio of the number of pixels corresponding to the ovarian parenchymal area to the total number of pixels in the ovarian ultrasound image.

[0155] S194: Combine the first mass value, the second mass value, and the third mass value to obtain the final probability.

[0156] In S191, the fourth regression model in this embodiment is a multiple linear regression model, used to characterize the correspondence between ovarian parenchymal area, number of follicles, and average follicle diameter and the predicted number of litters. Its expression is:

[0157]

[0158] in, This is the intercept term of the fourth regression model; The regression coefficient corresponding to the number of follicles represents the average contribution of a single follicle to the number of offspring. The regression coefficient corresponding to the ovarian parenchymal area represents the influence of ovarian tissue volume on litter size. The regression coefficients corresponding to the average follicle diameter represent the regulatory effect of follicle development stage on litter size. All coefficients were obtained by least squares fitting of large-sample paired data (simultaneously acquired ovarian ultrasound images and actual litter size records), and cross-validation confirmed that the model fit met the requirements.

[0159] The principle behind the fourth regression model lies in the natural physiological correlation between ovarian morphological parameters and litter size. The number of follicles directly reflects the size of the ovarian reserve pool, the ovarian parenchymal area reflects the overall developmental level of the ovary, and the average follicle diameter reflects the developmental stage of the follicle. These three factors collectively determine the sow's reproductive potential from different dimensions. The fourth regression model quantitatively characterizes the correspondence between morphological parameters and litter size through statistical fitting, enabling independent prediction of litter size from an image morphology perspective. This provides image-based prediction results completely independent of numerical modalities for subsequent multi-source evidence fusion, achieving cross-validation and complementarity between the two modalities.

[0160] In S192, the predicted number of litters output by the fourth regression model is used as the mean of the normal distribution on the image side, and the pre-calibrated second prediction standard deviation is used as the standard deviation of the normal distribution on the image side. The second prediction standard deviation is a fixed value estimated from the prediction residuals of the fourth regression model during the training phase. It is shared by all evaluation samples and reflects the inherent uncertainty of predicting the number of litters based on ovarian morphological parameters.

[0161] Then, a preset reproductive potential level threshold, which is exactly the same as the first intermediate probability, is adopted, including the lower limit of level S, level A, level B, and level C. By adopting a unified level threshold system, it is ensured that the second intermediate probability on the image side and the first intermediate probability on the numerical side are under the same level classification standard, providing a comparable probability distribution basis for subsequent multi-source probability fusion.

[0162] Next, the z-score corresponding to each level threshold is calculated. That is, for each level threshold, the difference between the threshold and the predicted number of offspring is divided by the second prediction standard deviation to obtain the standardized z-score value. The z-score represents the position of the level threshold in the standard normal distribution and is a bridge connecting the original value and the probability distribution.

[0163] Next, the cumulative probability is calculated, that is, the cumulative probability corresponding to the z-score for each level threshold is calculated using the cumulative distribution function Φ(z) of the standard normal distribution. The cumulative probability represents the sum of probabilities of the value being less than or equal to the threshold in the normal distribution, which provides the basis for the subsequent calculation of interval probabilities.

[0164] Finally, to calculate the probability of each level and output the second intermediate probability, the cumulative probability of the upper bound of each level is subtracted from the cumulative probability of the lower bound to obtain the interval probability corresponding to that level; where the S-level probability is 1 minus the cumulative probability corresponding to the lower limit threshold of S-level, and the D-level probability is the cumulative probability corresponding to the lower limit threshold of D-level; the five-level probability distribution obtained after the calculation is the second intermediate probability, which represents the probability distribution of reproductive potential levels based on image morphological parameters.

[0165] For example, S192 can be:

[0166] In this embodiment, taking gilts as an example, the first step is to determine the parameters of the normal distribution on the image side, where the mean is taken as the predicted number of piglets output by the fourth regression model, which is 13.0 heads, and the standard deviation is taken as the pre-calibrated second prediction standard deviation, which is 1.5 heads. This standard deviation is a fixed value estimated from the prediction residuals of the fourth regression model during the training phase, which is shared by all evaluation samples, and reflects the inherent uncertainty of predicting the number of piglets based on ovarian morphological parameters.

[0167] The second step is to determine the level thresholds. The preset reproductive potential level thresholds are exactly the same as the first intermediate probability. The lower limit for level S is 14.0 heads, level A is 12.0 heads, level B is 10.0 heads, and level C is 8.0 heads. By adopting a unified level threshold system, it is ensured that the second intermediate probability on the image side and the first intermediate probability on the numerical side are under the same level classification standard, providing a comparable probability distribution basis for subsequent multi-source probability fusion.

[0168] The third step is to calculate the z-score corresponding to each level threshold. Divide the difference between each level threshold and the predicted number of offspring by the second prediction standard deviation to obtain the standardized z-score value. The specific calculation results are as follows: the z-score corresponding to the lower limit of level S is (14.0-13.0) / 1.5=0.67, the z-score corresponding to the lower limit of level A is (12.0-13.0) / 1.5=-0.67, the z-score corresponding to the lower limit of level B is (10.0-13.0) / 1.5=-2.00, and the z-score corresponding to the lower limit of level C is (8.0-13.0) / 1.5=-3.33. The z-score represents the position of the level threshold in the standard normal distribution and is a bridge connecting the original value and the probability distribution.

[0169] The fourth step is to calculate the cumulative probability. Using the cumulative distribution function Φ(z) of the standard normal distribution, the cumulative probability corresponding to the z-score for each threshold level is calculated. The specific results are as follows: the cumulative probability corresponding to z=0.67 is 0.7486, the cumulative probability corresponding to z=-0.67 is 0.2514, the cumulative probability corresponding to z=-2.00 is 0.0228, and the cumulative probability corresponding to z=-3.33 is 0.0004. The cumulative probability represents the sum of probabilities of values ​​less than or equal to the threshold in the normal distribution, providing a basis for the subsequent calculation of interval probabilities.

[0170] The fifth step calculates the probability of each level and outputs the second intermediate probability. The cumulative probability of the upper bound of each level is subtracted from the cumulative probability of the lower bound to obtain the interval probability corresponding to that level. Specifically, the probability of level S is 1 minus the cumulative probability corresponding to the lower limit of level S, i.e., 1 - 0.7486 = 25.14%; the probability of level A is the cumulative probability of the lower limit of level S minus the cumulative probability of the lower limit of level A, i.e., 0.7486 - 0.2514 = 49.72%; the probability of level B is 0.2514 - 0.0228 = 22.86%; the probability of level C is 0.0228 - 0.0004 = 2.24%; and the probability of level D is the cumulative probability corresponding to the lower limit of level D, i.e., 0.04%. The sum of the five probabilities is 100%, verifying the calculation is correct. The final five-level probability distribution is the second intermediate probability, representing the probability distribution of reproductive potential levels based on image morphological parameters.

[0171] In S193, the first mass value is the numerical mass, which is calculated by multiplying the probability value of each level in the first intermediate probability by the numerical reliability coefficient to obtain the first mass value of the corresponding level. The numerical reliability coefficient is a pre-calculated value, which is calculated based on the intra-batch coefficient of variation (CV) value of the AMH detection. The specific calculation formula is that the numerical reliability coefficient is equal to max(0, 1-CV / CV_threshold). Its meaning is that the smaller the detection CV, the higher the reliability of the numerical evidence and the closer the mass value is to the original probability. The larger the CV, the lower the reliability and the more the mass value is discounted. For example, if the intra-batch CV is 3% and the CV threshold is 10%, the numerical reliability coefficient is 0.7. Multiplying each level of the first intermediate probability by 0.7 respectively gives the first mass value.

[0172] The second mass value, or image mass, is calculated as follows: the probability value of each level in the second intermediate probability is multiplied by the image reliability coefficient to obtain the corresponding second mass value. The image reliability coefficient is directly calculated from the ovarian ultrasound image. The specific formula is that the image reliability coefficient equals the number of pixels corresponding to the ovarian parenchymal area divided by the total number of pixels in the ovarian ultrasound image. This means that the larger the proportion of the ovary in the image, the more effective information the image contains, and the more reliable the image evidence. For example, if the total number of pixels in the image is 65,536 and the number of pixels in the ovarian parenchyma is 55,706, then the image reliability coefficient is approximately 0.85. Multiplying each level of the second intermediate probability by 0.85 yields the second mass value.

[0173] The third mass value, also known as the prior mass, is calculated as follows: multiply the probability value of each level in the prior probability by the prior reliability coefficient to obtain the third mass value for the corresponding level. The prior probability comes from large-sample population statistics, representing the distribution proportion of each reproductive potential level in the entire population. The prior reliability coefficient is a preset fixed value, set according to the reliability of prior knowledge, usually set to a small value such as 0.2, because the prior only reflects the average level of the population and has limited reference value for individual assessment. Setting a small coefficient can prevent the prior from excessively influencing individual judgment. Multiplying each level of the prior probability by the prior reliability coefficient yields the third mass value.

[0174] Each mass value is the probability distribution of the corresponding evidence source multiplied by a reliability coefficient. Higher reliability results in a larger coefficient, and the mass value is closer to the original probability; lower reliability results in a smaller coefficient, and the mass value is discounted more. The mass value characterizes the degree of support each evidence source provides for each reproductive potential level; it is the evidence representation after reliability discounting of the original probability distribution. Each evidence source, including numerical evidence, video evidence, and prior evidence, has its corresponding probability distribution. Multiplying the probability value of each level in the probability distribution by the reliability coefficient of the corresponding evidence source yields the mass distribution of that evidence source. A higher reliability coefficient means a mass value closer to the original probability value, indicating higher credibility of the evidence source and a greater weight in influencing the final decision; a lower reliability coefficient means a smaller mass value, indicating lower credibility of the evidence source and a smaller weight in influencing the final decision. Unlike the standard probability distribution, the sum of the five levels in the mass value is not equal to 1, because multiplying by the reliability coefficient is equivalent to "discounting" the evidence; the discounted portion represents the uncertainty of the evidence source. The mass value, as an intermediate representation of multi-source evidence fusion, provides a unified input format for subsequent fusion calculations between different evidence sources.

[0175] In this embodiment, the effective information in the ovarian ultrasound image is mainly concentrated in the ovarian parenchyma region. The proportion of the ovarian parenchyma in the image directly determines the information content and reliability of the imaging evidence. Specifically, when the ovary occupies a large proportion in the image, the details of the ovarian parenchyma and follicular structure are richer, the extraction accuracy of morphological parameters is higher, and the prediction results based on morphological parameters are more reliable. Conversely, if the ovary occupies a small proportion in the image, it indicates that most of the image consists of abdominal background or other irrelevant tissues, resulting in insufficient effective information, increased extraction error of morphological parameters, and reduced reliability of prediction results based on morphological parameters. Using the pixel ratio, an objective and quantifiable indicator, to characterize image reliability allows for the direct extraction of quality evaluation parameters from the image itself, without the need for additional manual evaluation or complex calculations. It has the advantages of strong objectivity, simple calculation, and clear physical meaning, and can provide an intuitive and reasonable quantitative basis for the weight allocation of imaging evidence in subsequent multi-source evidence fusion.

[0176] In S194, in at least one embodiment, the fusion method can be:

[0177] For each preset reproductive potential level, the corresponding values ​​of the first, second, and third mass values ​​are multiplied to obtain the product value for that level. Then, the products of the five levels are summed to obtain the total product. Finally, the product value of each level is divided by the total product for normalization, yielding the final probability for each preset reproductive potential level. The physical significance of this method is that evidence sources with higher reliability coefficients have larger mass values, exerting a greater influence on the product result and thus holding a higher weight in the final probability. Levels where all three evidence sources consider the probability high have larger product values ​​and higher final probabilities. Conversely, levels with significant disagreement among evidence sources have relatively smaller product values ​​and lower final probabilities, thus achieving the reinforcement of consistency and the weakening of inconsistency among multi-source evidence. This method originates from Bayesian inference theory, and taking the logarithm is equivalent to weighted log-probability fusion, possessing a rigorous mathematical foundation.

[0178] In this embodiment, from a probabilistic perspective, under the premise that multiple evidence sources are independent of each other, the joint probability that multiple pieces of evidence simultaneously support the same level is equal to the product of the probabilities of each piece of evidence. This is consistent with the calculation principle of posterior probability in Bayesian inference. Therefore, multiplying the mass values ​​of the corresponding levels of three independent evidence sources can reflect the degree of joint support of the three pieces of evidence pointing to that level. The reliability coefficient is incorporated into the product operation process through a discounted form of the mass value. The higher the reliability of the evidence source, the larger its mass value, the higher its weight in the product operation, and the greater its impact on the final fusion result. This achieves an adaptive weighting effect of high reliability evidence having a high weight and low reliability evidence having a low weight. From the perspective of the fusion effect, for levels that are consistently supported by the three pieces of evidence, the product value increases significantly after multiple multiplications, and the final probability increases accordingly, thus strengthening the consistency of multi-source evidence. Conversely, for levels where there is significant disagreement among the evidence, the product value decreases accordingly because the probability values ​​of some evidence are lower, and the final probability decreases accordingly, thus weakening the evidence disagreement effect. The aforementioned fusion mechanism can effectively aggregate information from multiple sources and reduce the impact of single evidence bias on the evaluation results. At the same time, this method is based on Bayesian inference theory, and taking the logarithm is equivalent to weighted log probability fusion, which has a rigorous mathematical foundation and theoretical basis, and can ensure the objectivity and rationality of the fusion results.

[0179] For example:

[0180] Taking gilts as an example, the first step is to calculate the product of each grade. That is, for each preset reproductive potential grade, multiply the corresponding values ​​of the first mass value, second mass value, and third mass value to obtain the product value of that grade. The specific calculation results are as follows: Grade S product is 0.014×0.276×0.020=0.000077, Grade A product is 0.245×0.499×0.060=0.007335, Grade B product is 0.350×0.065×0.070=0.001593, Grade C product is 0.084×0.009×0.030=0.000023, and Grade D product is 0.007×0.001×0.020=0.0000001. The physical significance of this step is that the product of levels supported by all three sources of evidence will be amplified, while the product of levels with significant disagreements among the sources of evidence will be reduced, thereby achieving a fusion effect of strengthening consistency and weakening inconsistency.

[0181] The second step is to calculate the sum of the products, which involves adding the product values ​​of the five levels together to obtain the sum of the products of all levels. The specific calculation result is: 0.000077 + 0.007335 + 0.001593 + 0.000023 + 0.0000001 = 0.009028. The purpose of this step is to provide a denominator for subsequent normalization, ensuring that the final output probability distribution satisfies the fundamental probability axiom that the sum of the probabilities of each level is 1.

[0182] The third step is normalization to obtain the final probability, which involves dividing the product of each level by the sum of the products to get the final probability for that level. The specific calculation results are as follows: Level S final probability is 0.000077 ÷ 0.009028 = 0.9%, Level A final probability is 0.007335 ÷ 0.009028 = 81.2%, Level B final probability is 0.001593 ÷ 0.009028 = 17.6%, Level C final probability is 0.000023 ÷ 0.009028 = 0.3%, and Level D final probability is approximately 0.0%. The sum of the probabilities for all five levels is 100%, verifying the calculation's correctness. The final classification result is Level A, which has the highest probability, indicating that the sow is most likely to belong to the Level A reproductive potential category.

[0183] In at least one embodiment, the method for fusing the first mass value, the second mass value, and the third mass value to obtain the final probability can employ the Dempster evidence synthesis rule, specifically:

[0184] First, the first mass value and the second mass value are used for the first evidence synthesis: First, the degree of conflict K between the two evidence sources is calculated, i.e., all combinations of inconsistent opinions in the two pieces of evidence are traversed, the mass values ​​corresponding to each combination are multiplied and summed to obtain the total degree of conflict. Then, for each preset reproductive potential level, all combinations where the intersection of the two pieces of evidence equals that level are found, including combinations where both pieces of evidence point to that level and combinations where one piece of evidence is uncertain. The mass values ​​of each combination are multiplied separately, summed, and then normalized by (1-K) to obtain the synthesized mass value for that level. Simultaneously, the uncertainty after synthesis is calculated, i.e., the product of the uncertainties of the two pieces of evidence divided by (1-K). Next, the fused mass distribution obtained from the first synthesis is used for a second Dempster synthesis with the third mass value, i.e., the prior evidence, using the same synthesis method as the first, to obtain the final fused mass distribution. Finally, the mass value of each level in the final fused result, i.e., the belief level, is used as the final probability for that level, and the level with the highest belief level is taken as the final reproductive potential level. The core of this synthesis rule is that the trust level of the consistent parts of the evidence is superimposed, while the conflicting parts are eliminated and re-normalized, thereby achieving effective fusion of multi-source evidence and decision-making.

[0185] This application provides an electronic device that may include a processing module and a memory. The memory stores a computer program, which, when executed by the processor, enables the electronic device to perform the corresponding steps in the above-described method for assessing the reproductive potential of sows.

[0186] In this embodiment, the processor can be an integrated circuit chip with signal processing capabilities. For example, the processor can be a central processing unit (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0187] The memory can be, but is not limited to, random access memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, etc.

[0188] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the electronic device described above can be referred to the corresponding steps in the aforementioned method, and will not be elaborated further here.

[0189] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the sow reproductive potential assessment method as described in the above embodiments.

[0190] Computer-readable storage media may be magnetic disks, optical disks, read-only memory, random access memory, flash memory, USB flash drives, hard disks, or solid-state drives, etc., and may also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implement the methods shown in the above embodiments.

[0191] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in the above-described method for assessing the reproductive potential of sows. The computer program product may exist in a computer-readable storage medium in forms including, but not limited to, source files, executable files, and installation package files.

[0192] Based on the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, electronic device, or network device, etc.) to execute the methods described in the various implementation scenarios of this application.

[0193] In the embodiments provided in this application, it should be understood that the disclosed methods can also be implemented in other ways. The method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, program segment, or part of code, which includes one or more executable instructions for implementing a specified logical function. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0194] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for assessing the reproductive potential of sows, characterized in that, The method includes: Obtain the raw AMH value, ovarian ultrasound image, physiological data, feeding environment data, environmental data, and health status data of the sows to be evaluated; Based on the original AMH value and physiological data, an original AMH feature sequence is obtained. The original AMH feature sequence includes the original AMH value and several original AMH feature values, which are calculated from the original AMH value and physiological data. Based on a preset image processing tool, the ovarian parenchymal area, number of follicles, and average diameter of follicles in the ovarian ultrasound image are extracted. Based on the original AMH value, ovarian parenchymal area, number of follicles and average diameter of follicles, the original AMH feature sequence is corrected to obtain the corrected AMH feature sequence, which includes the corrected AMH value and the corrected AMH feature value. Based on the corrected AMH value and the specified corrected AMH feature value, the ovarian ultrasound image is enhanced to obtain an enhanced image. Based on the corrected AMH feature sequence and the enhanced image, the initial predicted number of piglets to be evaluated and the initial probability of each preset reproductive potential level are obtained according to the pre-trained deep learning model. Based on the aforementioned rearing environment data, environmental data, and health status data, the initial predicted litter size is revised to obtain the final predicted litter size. The initial probability is corrected based on the final predicted number of offspring to obtain the first intermediate probability for each preset reproductive potential level; The final probability of the preset reproductive potential level is determined based on the first intermediate probability, the prior probability of the pre-established preset reproductive potential level, the ovarian parenchymal area, the number of follicles, and the average diameter of follicles.

2. The method according to claim 1, characterized in that, The aforementioned primitive AMH characteristic values ​​include intrinsic AMH values, which characterize the AMH values ​​after removing the influence of age and variety from the primitive AMH values; The physiological data includes the age and breed of the sows to be evaluated; The step of obtaining the original AMH feature sequence based on the original AMH value and physiological data includes: The age and variety are input into a pre-established first regression model, and the AMH benchmark value is output through the first regression model. The first regression model is used to characterize the correspondence between the age and variety and the AMH benchmark value. The intrinsic AMH value is obtained by the difference between the original AMH value and the AMH reference value.

3. The method according to claim 1, characterized in that, The step of correcting the original AMH feature sequence based on the original AMH value, ovarian parenchymal area, number of follicles, and average follicle diameter includes: The ovarian parenchymal area, number of follicles, and average follicle diameter are input into the second regression model, and the AMH predicted value is obtained through the second regression model. The second regression model represents the correspondence between the ovarian parenchymal area, number of follicles, and average follicle diameter and the AMH predicted value. Calculate the deviation rate between the original AMH value and the predicted AMH value; When the deviation rate is less than or equal to a preset value, the original AMH value and the AMH prediction value are weighted to obtain the corrected AMH value. The corrected AMH feature sequence is obtained based on the corrected AMH value.

4. The method according to claim 1, characterized in that, The step of enhancing the ovarian ultrasound image based on the corrected AMH value and a specified corrected AMH feature value to obtain an enhanced image includes: Based on the corrected AMH value and the specified corrected AMH feature value, enhancement parameters are obtained according to the specified enhancement rule. The specified enhancement rule includes a first range of the corrected AMH value and a second range of the specified corrected AMH feature value mapped to the enhancement parameters. The enhancement parameters include enhancement weight, a first Gaussian standard deviation and a second Gaussian standard deviation, wherein the first Gaussian standard deviation is less than the second Gaussian standard deviation. The ovarian ultrasound image is subjected to Gaussian blurring using a first Gaussian standard deviation as a parameter to obtain a first blurred image; the ovarian ultrasound image is then subjected to Gaussian blurring using a second Gaussian standard deviation as a parameter to obtain a second blurred image. The first blurred image is subtracted from the second blurred image pixel by pixel to obtain the difference of Gaussian image; The enhanced gray value of each pixel in the ovarian ultrasound image is obtained by multiplying the enhancement weight by the Gaussian difference value corresponding to the pixel. The Gaussian difference value is obtained based on the Gaussian difference image. The enhanced grayscale values ​​are truncated at upper and lower limits to ensure that the grayscale values ​​are within a pre-set valid range, and the enhanced image is finally obtained.

5. The method according to claim 1, characterized in that, The process of revising the initial predicted litter size based on the feeding environment data, environmental data, and health status data to obtain the final predicted litter size includes: Based on the feeding environment data, environmental data, and health status data, the current feeding level index, current environmental stress index, and current health index are obtained, and the current feeding level index, current environmental stress index, and current health index are all greater than 0 and less than 1. The current feeding level index, current environmental stress index, and current health index are input into the third regression model, and the current confounding effect size is calculated through the third regression model. The current confounding effect size represents the contribution of the current feeding conditions corresponding to the current feeding level index, current environmental stress index, and current health index to the initial predicted number of piglets. The pre-set baseline feeding level index, baseline environmental stress index, and baseline health index are input into the third regression model, and the baseline confounding effect size is calculated through the third regression model. Calculate the difference between the initial predicted number of offspring and the current confounding effect; The sum of the difference and the baseline confounding effect is calculated to obtain the final predicted litter size, thereby correcting the initial predicted litter size.

6. The method according to claim 1, characterized in that, The step of correcting the initial probability based on the final predicted number of offspring to obtain a first intermediate probability for each preset reproductive potential level includes: The final predicted number of offspring, the preset level threshold of reproductive potential, and the initial probability are input into the Newton iterative model, which is based on Newton's method and outputs the first prediction standard deviation. Based on the final predicted number of litters, the initial predicted number of litters, and the first predicted standard deviation, the first intermediate probability is determined using the cumulative distribution function of the standard normal distribution.

7. The method according to claim 1, characterized in that, The step of determining the final probability of the preset reproductive potential level based on the first intermediate probability, the prior probability of the pre-established preset reproductive potential level, the ovarian parenchymal area, the number of follicles, and the average diameter of follicles includes: The ovarian parenchymal area, number of follicles, and average follicle diameter are input into a fourth regression model, and the predicted number of litters is output by the fourth regression model. The fourth regression model represents the correspondence between the ovarian parenchymal area, number of follicles, and average follicle diameter and the predicted number of litters. The second intermediate probability of the preset reproductive potential level is calculated by using the predicted number of offspring, the first intermediate probability, and the pre-calibrated second prediction standard deviation. A first mass value is obtained based on the first intermediate probability and the pre-calculated numerical reliability coefficient; a second mass value is obtained based on the second intermediate probability and the image reliability coefficient; and a third mass value is obtained based on the prior probability and the preset prior reliability coefficient, wherein the image reliability coefficient is equal to the ratio of the number of pixels corresponding to the ovarian parenchymal area to the total number of pixels in the ovarian ultrasound image. The final probability is obtained by combining the first mass value, the second mass value, and the third mass value.

8. An electronic device, characterized in that, The electronic device includes a processor and a memory coupled together, the memory storing a computer program that, when executed by the processor, causes the electronic device to perform the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the method as described in any one of claims 1 to 7.

10. A program product, characterized in that, It includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images