A real-time evaluation method for sheep follicle development state based on fusion of ultrasonic image and pressure sensing

Abstract

The application discloses a kind of based on ultrasound image and pressure sensing fusion's sheep follicle development state real-time evaluation method, it is related to the intelligent evaluation technical field of animal reproductive traits.The method is by obtaining the ultrasound image and internal pressure data of target follicle, the ultrasound image is segmented using WIFE-Net deep learning network to extract follicle maximum diameter, minimum diameter, wall thickness and sphericity index, simultaneously, stable pressure waveform is intercepted from continuous pressure value and the rest pressure in cavity and pressure fluctuation amplitude are calculated, the above six features are fused into six-dimensional feature vector input random forest model, and the attachment tightness level between oocyte-oocyte complex and follicle wall is output, divided into high, medium, low three levels;The present application expands the detection dimension from macroscopic morphology to internal mechanical environment by multi-modal fusion perception, realizes the quantitative discrimination of the inside-out state that follicle external morphological maturity and internal tight attachment are different.

Description

Technical Field

[0001] This invention relates to the field of intelligent assessment technology for animal reproductive traits, specifically a real-time assessment method for sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing. Background Technology

[0002] In vitro embryo production in sheep plays a crucial role in the breeding of superior breeds. Accurate assessment of follicular development is a key prerequisite for determining the quality of oocyte collection and the success rate of subsequent embryo development. Traditional assessment of follicular development relies primarily on the operator's observation of the external outline, diameter, and wall thickness of sheep follicles using ultrasound imaging. The main drawback of this method is that external morphological characteristics alone cannot accurately reflect the complex physical changes within the follicle: a sheep follicle that appears mature externally may still have its oocyte-cumulus complex tightly attached to the follicle wall. This discrepancy between appearance and reality makes conventional ultrasound morphological assessment highly misleading.

[0003] The lack of perception of the internal physical state of follicles prevents operators from accurately determining the tightness of attachment between the oocyte-cumulus complex and the follicle wall, forcing them to rely on subjective experience or general morphological indicators for inference. The internal mechanical environment of sheep follicles differs significantly at different developmental stages: mature follicles exhibit lower and more stable intracavitary resting pressure, with a looser bond between the cumulus cell layer and the follicle wall; while immature follicles show higher intracavitary resting pressure and greater pressure fluctuations, with a tighter bond between the cumulus cell layer and the follicle wall. These inherent mechanical characteristics cannot be directly observed through conventional ultrasound imaging. Relying solely on morphological parameters from imaging for follicle development assessment not only makes it difficult to distinguish follicles with similar external morphology but different internal attachment states but may also lead to misjudgment of the true developmental stage of the follicle, thus affecting the formulation and implementation of subsequent oocyte retrieval strategies. Therefore, how to effectively integrate the external morphological characteristics and internal mechanical characteristics of follicles to construct an assessment method that can objectively quantify the degree of adhesion between the oocyte-cumulus complex and the follicle wall has become a key technical issue for improving the accuracy of sheep follicle development status assessment. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time assessment method for the developmental status of sheep follicles based on the fusion of ultrasound imaging and pressure sensing. Through multimodal data fusion and machine learning model inference, it can achieve an objective quantitative assessment of the degree of adhesion between the oocyte-cumulus complex and the follicle wall inside the sheep follicle.

[0005] The objective of this invention can be achieved through the following technical solutions: This application provides a method for real-time assessment of sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing, including the following steps: S1. Acquire real-time ultrasound image data of the target follicle, as well as corresponding continuous pressure data inside the cavity of the target follicle; S2. Input each frame of ultrasound image data into the pre-trained WIFE-Net, perform parallel operations of conventional convolution and dilated convolution, and aggregate multi-scale spatial information using double pyramid pooling to output a pixel-level segmentation mask of the target follicle region; calculate the maximum diameter, minimum diameter, follicle wall thickness, and follicle sphericity index of the target follicle based on the segmentation mask. S3. Extract a segment of pressure waveform representing the stable mechanical environment within the follicular cavity from the continuous pressure numerical data, calculate the time average value of the pressure waveform as the resting pressure within the follicular cavity, and calculate the standard deviation of the pressure waveform as the pressure fluctuation amplitude. S4. The maximum diameter, the minimum diameter, the follicle wall thickness, the follicle sphericity index, the follicle cavity resting pressure, and the pressure fluctuation amplitude are combined into a six-dimensional feature vector, which is then input into a pre-trained random forest model. The random forest model outputs the degree of adhesion between the oocyte-cumulus complex and the follicle wall.

[0006] The beneficial effects of this invention are as follows: This invention constructs a six-dimensional feature vector by integrating morphological features of ultrasound imaging with mechanical features of pressure sensing, and inputs it into a random forest model to output the degree of adhesion between the oocyte-cumulus complex and the follicle wall. This breaks through the limitations of traditional ultrasound, which can only assess external morphology such as follicle size and wall thickness. It expands the detection dimension from macroscopic morphology to the internal mechanical environment, and realizes the quantitative discrimination of the different states of the follicle surface and interior. This allows the assessor to accurately grasp the binding state between the cumulus cell layer and the follicle wall with clear high, medium and low three-level classifications. It effectively solves the technical problem that it is impossible to perceive the internal mechanical environment and objectively judge the attachment state by simply relying on external morphology. This invention establishes a quantitative correlation model between external geometric morphology and internal physical state by multimodal fusion of ultrasound imaging morphological parameters such as maximum and minimum follicle diameter, wall thickness, and sphericity index with mechanical parameters such as resting pressure and pressure fluctuation amplitude within the follicular cavity. This method does not rely on the operator's subjective experience and judgment, but instead uses a random forest model to automatically infer and output attachment levels from six-dimensional feature vectors, achieving objectivity and standardization in the assessment of follicle development status. This significantly improves the consistency and repeatability of the assessment results, providing reliable data support for the formulation of subsequent oocyte retrieval strategies. This invention pre-trains a random forest model using in vitro experimental data and solidifies the evaluation criteria, enabling the evaluation process to be completed without any additional intervention or manipulation of live sheep follicles, requiring only the acquisition of ultrasound images and pressure sensing data. The method can be executed offline or in real-time, exhibiting good portability and applicability. By accurately extracting follicle morphology parameters through the multi-scale image segmentation network WIFE-Net and adaptively extracting stable pressure waveforms using autocorrelation functions, the invention calculates resting pressure and fluctuation amplitude. This invention effectively identifies the special condition of "externally mature but internally tightly attached" follicle development in sheep, filling a technological gap in sensing the internal mechanical environment of sheep follicles and improving the evaluation of follicle development quality in sheep reproduction. Attached Figure Description

[0007] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0008] Figure 1 A flowchart illustrating a real-time assessment method for sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing, provided in Embodiment 1 of this application; Figure 2 This is a flowchart illustrating step S3 in a real-time assessment method for sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing, as provided in Embodiment 1 of this application. Figure 3 This is a flowchart illustrating step S4 of a real-time assessment method for sheep follicle development based on the fusion of ultrasound imaging and pressure sensing, as provided in Embodiment 1 of this application. Detailed Implementation

[0009] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0010] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0011] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0012] Example 1, please refer to Figures 1-3 This embodiment provides a real-time assessment method for the developmental status of sheep follicles based on the fusion of ultrasound imaging and pressure sensing, including the following steps: S1. Acquire real-time ultrasound image data of the target follicle, as well as corresponding continuous pressure data inside the cavity of the target follicle; Further, step S1 specifically includes: S11. Real-time ultrasound image data of the target follicle is acquired by ultrasound equipment, wherein the ultrasound image data includes at least imaging information of the anechoic dark area of ​​the follicle region and its boundary contour. The ultrasound equipment uses a high-frequency vaginal probe with a center frequency of no less than 7.5MHz, covering the area of ​​the sheep ovary and clearly displaying the anechoic dark area of ​​the follicle and its boundary contour. By adjusting the gain and depth parameters of the ultrasound equipment, the follicle area is made to present clear anechoic features in the image, so that the subsequent image segmentation network can accurately identify the follicle boundary. The acquired ultrasound image data is stored in DICOM format or lossless compression format to retain the original pixel information and avoid edge blurring caused by lossy compression.

[0013] S12. Collect continuous pressure data inside the target follicle cavity using a pressure sensor. The pressure data is continuously recorded at a preset sampling frequency to form time-series pressure data. The pressure sensor employs a MEMS pressure sensing chip with a range of -100 mmHg to +200 mmHg and a resolution of 0.1 mmHg. Through miniaturized packaging, it can be implanted into the target follicle cavity for pressure detection. The preset sampling frequency is determined based on the Nyquist sampling theorem and the dominant frequency of pressure fluctuations within the sheep follicle: by performing Fourier analysis on the pressure waveform within the sheep follicle cavity, the highest frequency component of the pressure fluctuation energy concentration is obtained, and the sampling frequency is set to 5 to 10 times this highest frequency, i.e., 250 Hz to 500 Hz. The analog voltage signal output by the pressure sensor is converted into a digital quantity by an analog-to-digital converter and stored in the data buffer along with the sampling timestamp, forming continuous time-series pressure data. S13. The ultrasound image data and the pressure data are synchronized using a unified timestamp to establish a time correspondence between each frame of ultrasound image and the corresponding pressure value. The synchronization is achieved through a hardware synchronization signal: when the ultrasound device acquires a frame of ultrasound image, it synchronously outputs a pulse signal to the data acquisition card. The data acquisition card simultaneously records the latest sampling value of the current pressure sensor and its corresponding timestamp, thereby ensuring that each ultrasound image frame and pressure sampling point have a one-to-one correspondence in time. The synchronization accuracy is controlled within ±1ms to ensure the consistency of data in each dimension in time during subsequent multimodal feature fusion.

[0014] S14. Use the synchronized ultrasound image data and pressure numerical data as input data for subsequent steps of image segmentation and pressure feature extraction.

[0015] Specifically, S1 acquires real-time ultrasound image data of the target follicle using ultrasound equipment, and simultaneously acquires continuous pressure data inside the target follicle cavity using a pressure sensor. The two types of data are synchronized in time, achieving accurate alignment of multimodal data, which provides a data foundation for subsequent morphological feature extraction and mechanical feature extraction in S2 and S3, respectively.

[0016] S2. Input each frame of ultrasound image into the pre-trained WIFE-Net, use conventional convolution and dilated convolution in parallel and use double pyramid pooling to aggregate multi-scale spatial information, and output a pixel-level segmentation mask of the follicle region; calculate the maximum diameter, minimum diameter, follicle wall thickness and follicle sphericity index of the target follicle based on the segmentation mask. Further, step S2 specifically includes: S21. Input each frame of ultrasound image into the encoder of WIFE-Net. The encoder is set to run in parallel with a regular convolution branch and a dilated convolution branch. The dilated convolution branch uses multiple parallel sub-branches with different expansion rates to simultaneously capture the fine boundary of the follicle wall and the contextual information of the tissue surrounding the follicle. Among them, WIFE-Net stands for Widely Integrated Follicle Extraction Network, which is a deep learning model specifically designed for automatically identifying and delineating the precise contours of follicles from ovarian ultrasound images. In the segmentation of amniotic follicle ultrasound images, the conventional convolution branch and the dilated convolution branch run in parallel. The conventional convolution branch uses a 3×3 convolution kernel with a stride of 1 and a padding of 1, and its receptive field is 3×3. When the 3×3 convolution kernel slides through the feature map pixel by pixel with a stride of 1, the center of the window is located at the boundary pixel of the follicle wall, which simultaneously covers the low grayscale pixels inside the follicle, the high grayscale pixels of the follicle wall, and the medium and high grayscale pixels of the surrounding tissue. The weight parameters learned through training produce a strong response to this grayscale step change. The stride of 1 ensures that any small grayscale change at the boundary position can be completely captured without losing edge details due to downsampling or stride crossing. The padding of 1 ensures the consistency of the number of convolution calculations between the boundary pixels and the internal pixels, avoiding the weakening of the feature response at the edge, thereby achieving high-precision positioning of the fine boundary of the follicle wall. Meanwhile, the dilated convolution branch consists of three parallel sub-branches with dilation rates of 2, 4, and 8, respectively. Each branch, without downsampling, obtains effective receptive fields of 7×7, 15×15, and 31×31 through different dilation rates. The sub-branch with a dilation rate of 2 covers the follicle wall and the adjacent basal layer to observe the local tissue grayscale contrast on both sides of the follicle wall. The sub-branch with a dilation rate of 4 covers the surrounding stroma region to observe structural landmarks such as vascular cross-sections and connective tissue strands, thus inferring the location of missing boundaries when boundaries break. The sub-branch with a dilation rate of 8 covers the far-field ovarian background to distinguish the true follicle boundary from artifact boundaries generated by adjacent structures. The above conventional convolution branch and the three dilated convolution sub-branches are computed in parallel on the same input image. The convolutions of each branch are independent and do not wait for each other, all being completed synchronously in one forward propagation of the network. Subsequently, the feature maps output by the four branches are concatenated along the channel dimension, allowing the fine boundary response information of the follicle wall and multi-scale contextual information to exist side-by-side along the channel dimension, achieving simultaneous capture of the fine boundary of the follicle wall and the multi-scale contextual information of the surrounding tissue. Then, a 1×1 convolution is used to compress the number of channels in the stitched feature map back to the original number of channels. The weights of the 1×1 convolution are automatically learned during training to perform weighted fusion of features from each channel, enabling the network to adaptively select the most effective feature combination based on the specific situation of the current follicle. When the follicle boundary is clear, fine boundary information is used for accurate localization. When the follicle boundary is blurred or interfered with by noise, the contextual information of the large receptive field is used for auxiliary inference. This effectively addresses the problem of inaccurate segmentation of amniotic follicle ultrasound images caused by blurred boundaries, size differences, and uneven gray levels.

[0017] S22. The feature maps output by the regular convolutional branches and each dilated convolutional branch are concatenated in the channel dimension, and then the number of channels is compressed through the convolutional layer to fuse local details and multi-scale semantic features. Specifically, the channel-dimensional concatenation involves connecting the feature map output from the regular convolutional branch (with C channels) to the feature maps output from the three dilated convolutional branches (each with C channels) end-to-end along the channel direction to obtain a concatenated feature map with 4C channels. The channel compression uses a 1×1 convolutional layer to map the input channel number 4C to the output channel number C, which preserves the multi-scale information of each branch and reduces the subsequent computational load. The weights of the 1×1 convolutional layer are learned through training and automatically adjust the fusion weights of features at different scales, enabling the network to adaptively select the most effective receptive field combination according to the actual size of the amniotic follicle.

[0018] S23. Input the fused feature map into the dual pyramid pooling module, perform average pooling at multiple scales at the image level and pixel level respectively, upsample the pooling results at each scale to the original size and then stitch them with the original feature map to enhance the segmentation robustness for follicles of different sizes and shapes. The dual pyramid pooling module comprises two parallel sub-modules: an image-level pyramid pooling module and a pixel-level pyramid pooling module. The image-level pyramid pooling module performs average pooling on the input feature map at four scales: 1×1, 2×2, 4×4, and 8×8, respectively, to obtain four global and local contextual feature maps with different resolutions. Each feature map is then reduced in dimensionality by 1×1 convolution and upsampled to the original feature map size through bilinear interpolation. The pixel-level pyramid pooling module extracts the local average value within a 3×3, 5×5, or 7×7 window centered on each pixel position of the input feature map, forming three sets of pixel-level neighborhood features. The four upsampled feature maps from the image-level pyramid pooling module, the three sets of neighborhood feature maps from the pixel-level pyramid pooling module, and the original fused feature map are concatenated along the channel dimension to obtain an enhanced feature map with rich spatial contextual information, thereby solving the problem of inaccurate segmentation caused by uneven grayscale and blurred edges inside the follicle due to ultrasound scattering in amniotic follicle images.

[0019] S24. The enhanced feature map is upsampled step by step by the decoder and skipped to output a binary segmentation mask of the same size as the input image, where the foreground is the follicle region and the background is the non-follicle region. The follicle contour is extracted from the mask by the boundary tracking algorithm, and the major axis of the minimum circumscribed ellipse of the contour is used as the maximum diameter, the minor axis as the minimum diameter, the average gray-level gradient within the contour as the wall thickness, and the ratio of the major axis to the minor axis as the sphericity index. The decoder employs a four-layer upsampling structure. Each layer contains a 3×3 transposed convolution (with a stride of 2) to double the feature map size. The feature maps are then concatenated with the corresponding layers of the encoder via skip connections in the channel dimension. After two 3×3 convolutional layers, the features are refined. Finally, the output layer uses a 1×1 convolution to map the number of channels to 2. After passing through the Softmax activation function, a probability map is obtained for each pixel belonging to a follicle or the background. Pixels with a probability greater than 0.5 are selected as the foreground to generate a binary segmentation mask. The boundary tracking algorithm employs the Suzuki algorithm, traversing the binary segmentation mask image. A marker matrix of the same size as the input image records the access status of each pixel. Starting from the top left corner of the image, it scans pixel by pixel, first then column by column. When an unvisited pixel (follicle region) is encountered, it is used as the starting point of the contour, and a new contour list is created. Then, using this point as the current tracking point, the search proceeds in an eight-neighborhood clockwise order (right, top right, top, top left, left, bottom left, bottom, bottom right). Starting from the opposite direction of the previous search, it rotates 90° clockwise to begin the search, selecting the first neighboring pixel within the foreground range as the next boundary point, appending it to the contour list, and marking it as tracked. This process iterates until the next boundary point equals the starting point, completing the tracking of the closed contour and marking it as complete. The scanning continues until the entire image is traversed, acquiring all closed contours. Finally, by calculating the area enclosed by each contour, the contour with the largest area is selected as the target follicle boundary, and its pixel coordinate sequence arranged clockwise is output for subsequent minimum bounding ellipse fitting and grayscale gradient calculation. The minimum circumscribed ellipse is obtained by calculating the covariance matrix of the boundary point set to obtain the ellipse center, major and minor axis lengths, and rotation angle. The average gray-level gradient within the contour is calculated as follows: on the original ultrasound image, three pixels are taken outward and inward along the contour normal direction, the variance of the gray-level values ​​of these six pixels is calculated, and then the average value is taken along the entire contour. The larger the average value, the more significant the difference in gray level between the follicle wall and the surrounding tissue, that is, the thinner the follicle wall or the clearer the boundary. The sphericity index is defined as the ratio of the major axis to the minor axis of the minimum circumscribed ellipse. The closer the ratio is to 1, the closer the follicle is to a circle.

[0020] Specifically, S2 uses the parallel structure of regular and dilated convolutions and the dual pyramid pooling module of WIFE-Net to capture multi-scale contextual information while maintaining feature map resolution. Combined with Suzuki boundary tracking and covariance matrix fitting, it accurately extracts follicle contours and calculates morphological parameters, solving the technical problem of inaccurate segmentation of sheep follicle ultrasound images due to blurred boundaries, size differences, and uneven gray levels. It achieves stable and high-precision follicle morphological quantification, providing reliable geometric feature data for subsequent attachment status assessment.

[0021] S3. Extract a segment of pressure waveform representing the stable mechanical environment within the follicular cavity from the continuous pressure numerical data, calculate the time average value of the pressure waveform as the resting pressure within the follicular cavity, and calculate the standard deviation of the pressure waveform as the pressure fluctuation amplitude. Furthermore, step S3 specifically includes: S31. Identify the starting moment when the pressure value drops sharply and the pressure fluctuation amplitude gradually converges to a stable range from the pressure numerical data, and take this moment as the starting mark point of the stable pressure waveform in the follicular cavity. The pressure sensor continuously collects pressure data at a preset sampling frequency to obtain a discrete pressure value sequence that changes over time. The method for detecting a sudden pressure drop is as follows: the pressure difference between two adjacent sampling points is calculated sequentially. When three consecutive differences are negative and the absolute value of each difference exceeds a preset drop threshold, and the cumulative drop from the start of the pressure drop to the current sampling point exceeds a specified percentage of the baseline pressure, a sudden pressure drop event is determined. The method for determining the convergence of pressure fluctuation amplitude is as follows: starting from several sampling points after the above-mentioned drop event, a sliding window of fixed time length is used to calculate the standard deviation of the pressure values ​​within the window. When the standard deviation of three consecutive windows is less than a preset stability threshold, the starting time of the third window is marked as the pressure stabilization start time, which is the starting mark point of the stable pressure waveform in the follicular cavity.

[0022] S32. Extract a continuous pressure waveform from the marked point. The duration of the extracted waveform is determined based on the autocorrelation function of the pressure fluctuation in the amniotic follicle: calculate the autocorrelation function of the stable pressure data, find the delay time corresponding to the first decay of the autocorrelation function value to a specified proportion of its maximum value as the decorrelation time, and take the extraction duration as three times the decorrelation time. The autocorrelation function is used to measure the similarity of pressure signals under different time delays. It is calculated as follows: the stable pressure data is multiplied point-by-point by its own data after a delay time τ, the sum is obtained, and then divided by the total number of points to obtain the autocorrelation function value R(τ). As τ gradually increases from 0, R(τ) begins to decay from its maximum value (signal energy when τ=0). The value of τ corresponding to the first decay of R(τ) to a specified proportion of its maximum value (e.g., 1 / e, approximately 36.8%) is defined as the decorrelation time. The truncation time is set to three times the decorrelation time to ensure that the truncated pressure waveform contains at least three independent fluctuation cycles, fully reflecting the stable mechanical environment characteristics within the follicular cavity. If the autocorrelation function decays very slowly, a maximum truncation time is set to avoid errors introduced by slow pressure drift due to follicular wall leakage.

[0023] S33. Calculate the arithmetic mean of all sampling points within the truncated waveform as the resting pressure within the follicular cavity; The arithmetic mean is calculated as follows: the pressure values ​​of all sampling points within the captured stable pressure waveform are summed, and then divided by the total number of sampling points. Before calculating the arithmetic mean, to improve the anti-pulse interference capability, the pressure data is first subjected to median filtering to remove isolated outliers. Specifically, for each sampling point, the pressure values ​​of itself and an odd number of adjacent sampling points (e.g., two before and two after, for a total of five) are taken. These values ​​are sorted by size, and the value at the median position is taken as the new pressure value of that sampling point. After median filtering and smoothing, the average of the pressure values ​​of all sampling points is calculated according to the above arithmetic mean formula. This average is the resting pressure in the follicular cavity, and its unit is millimeters of mercury. This resting pressure reflects the static tension exerted by the follicular fluid on the follicular wall and is one of the important parameters for subsequent assessment of the follicular mechanical state.

[0024] S34. Calculate the arithmetic square root of the sum of squares of the differences between each sampling point within the waveform and the resting pressure, and use it as the pressure fluctuation amplitude. The formula for calculating the pressure fluctuation amplitude is: Where N is the total number of sampling points within the truncated waveform. The pressure value at the i-th sampling point after median filtering. The resting pressure within the follicular cavity calculated in step S33; the amplitude of this pressure fluctuation. It reflects the degree of dispersion of pressure fluctuation in the follicular cavity of sheep over time. When the cumulus cells are tightly attached to the follicular wall, regular small-amplitude vibrations keep the pressure at a low level. When the attachment tends to loosen or partially detach, irregular large-amplitude vibrations keep the pressure at a corresponding high level. This parameter is used as an independent dimension of the six-dimensional feature vector and input into the random forest model to help determine the degree of attachment between the oocyte-cumulus complex and the follicular wall.

[0025] Specifically, S3 accurately locates the starting point of the stable pressure waveform in the follicular cavity by detecting sudden drops based on the pressure difference sequence and using the sliding window standard deviation convergence criterion. It also adaptively determines the truncation time based on the autocorrelation function to obtain a stable pressure waveform, and then calculates the resting pressure and pressure fluctuation amplitude. This solves the problem of pressure feature extraction deviation caused by the inability to accurately identify the starting point of the stable pressure waveform, and realizes the objective and stable quantification of mechanical parameters in the follicular cavity, providing reliable mechanical dimension features for subsequent attachment level discrimination.

[0026] S4. The maximum diameter, the minimum diameter, the follicle wall thickness, the follicle sphericity index, the follicle cavity resting pressure, and the pressure fluctuation amplitude are combined into a six-dimensional feature vector, which is then input into a pre-trained random forest model; the random forest model outputs the degree of adhesion between the oocyte-cumulus complex and the follicle wall. Furthermore, step S4 specifically includes: S41. The maximum diameter, minimum diameter, wall thickness, sphericity index, intracavitary resting pressure, and pressure fluctuation amplitude of the follicle are spliced ​​together in a fixed order to form a six-dimensional fusion feature vector. The first four components are derived from the morphological quantification of ultrasound images, and the last two components are derived from the mechanical parameters of pressure sensing. This feature vector simultaneously encodes the external geometric morphology and internal mechanical environment of the follicle. The fixed order is as follows: the first dimension is the maximum diameter of the follicle, the second dimension is the minimum diameter of the follicle, the third dimension is the follicle wall thickness, the fourth dimension is the follicle sphericity index, the fifth dimension is the resting pressure inside the follicle cavity, and the sixth dimension is the pressure fluctuation amplitude. This order is kept consistent during training and use to ensure that the physical meaning of each dimension of the feature vector is clear. By integrating external morphological features with internal mechanical features, the technical problem that traditional methods cannot perceive the true attachment state inside the follicle by relying solely on the ultrasound appearance is solved, thus overcoming the misleading nature of discrepancies between appearance and reality.

[0027] S42. Input the six-dimensional feature vector into a pre-trained random forest model. Construct multiple decision trees from sheep ex vivo experimental samples through bootstrap aggregation sampling. When splitting at a node, each decision tree randomly selects some features from all features and chooses the splitting threshold that maximizes the reduction of impurity. The input feature vector is processed by each decision tree from the root node, judging layer by layer according to the feature index and splitting threshold stored in each node, and finally reaches the leaf node. The attachment level stored in the node is the prediction result of the tree. The pre-training process of the random forest model is as follows: First, a large number of samples from sheep in vitro follicle experiments are collected. Each sample contains a six-dimensional feature vector and the true attachment level (high, medium, and low) determined by anatomical microscopy. Then, a bootstrap aggregation sampling method is used to randomly extract multiple sub-training sets with replacement from the original training set. The size of each sub-training set is the same as that of the original training set, which are used to construct a decision tree. When constructing each decision tree, for each node split, a fixed number of feature subsets (e.g., two features) are randomly selected from all six features. All possible splitting thresholds for each feature in the subset are traversed, and the impurity after each candidate split is calculated (using Gini impurity or information gain). The feature that reduces the impurity the most and its splitting threshold are selected as the splitting rule for that node. The growth of the decision tree repeats the above process until all samples in each leaf node belong to the same attachment level or the preset maximum depth is reached. All decision trees together form a random forest model, which is frozen after training for real-time evaluation.

[0028] S43. The random forest model collects the predicted levels of all decision tree outputs, uses a majority voting mechanism to determine the final result, and counts the frequency of high, medium and low levels in all decision trees. The level with the highest frequency is taken as the attachment tightness level of the final output. If two levels have the same frequency, the average prediction probability corresponding to these two levels is compared. The proportion of samples stored in each leaf node is used as the probability estimate, and the level with the highest probability is selected as the output. Furthermore, the final result is determined by a majority voting mechanism, expressed as:

[0029] Where E represents the attachment tightness level of the output, and M represents the total number of decision trees in the random forest. Let be the set of leaf nodes of the i-th decision tree. This represents the number of leaf nodes in the tree. Let be the predicted category of leaf node j for the input six-dimensional feature vector v, and let {.} be the indicator function.

[0030] The actual meaning of this formula is as follows: Each decision tree in the random forest outputs an attachment level prediction for the input feature vector v. Then, the frequency of predictions as high, medium, and low levels in the M decision trees is counted, and the level with the highest frequency is taken as the final output. If two levels have the same frequency (for example, high and medium each receive M / 2 votes), the average prediction probability of these two levels is calculated by further utilizing the sample proportion stored in the corresponding leaf node of each decision tree, and the level with the higher average probability is selected as the final output. This voting mechanism fully leverages the advantages of ensemble learning, effectively avoids misjudgments caused by sample perturbation or feature noise in a single decision tree, and improves the accuracy and robustness of sheep follicle attachment status discrimination.

[0031] S44. Based on the output attachment tightness level, automatically determine the binding state between the oocyte-cumulus complex and the follicle wall. A high level indicates a large number of granulosa cell layers, dense intercellular connections, and strong binding force between the complex and the follicle wall. A low level indicates a small number of granulosa cell layers, loose intercellular connections, and weak binding force between the complex and the follicle wall. A medium level indicates a medium attachment state between the two. The correspondence between the degree of attachment tightness and the actual binding state of the oocyte-cumulus complex was established in advance through sheep in vitro follicle experiments: sheep follicles at different developmental stages were collected, and the level prediction output by the random forest model was obtained through the method described in this embodiment. Then, the follicles were dissected and the number of cumulus cell layers, the density of cell arrangement, and the separation resistance between the oocyte and the follicle wall (measured by a micro-force sensor) were observed and recorded under a stereomicroscope. The above experimental data were used as the standard to establish the level classification rules: the high level corresponds to no less than three cumulus cell layers with cells arranged tightly without gaps, the low level corresponds to no more than two cumulus cell layers or loose cell arrangement or even some cells have fallen off, and the medium level is between the two, with about two to three cumulus cell layers arranged in a moderately tight manner.

[0032] Specifically, S4 integrates morphological parameters from ultrasound images with mechanical parameters from pressure sensors into a six-dimensional feature vector, inputs it into a random forest model pre-trained from sheep in vitro experimental data and processed by majority voting and probability adjudication mechanisms, and outputs the degree of attachment tightness between the oocyte-cumulus complex and the follicle wall. This solves the technical problem of inaccurate assessment caused by the inability to perceive the actual attachment state inside the follicle by relying solely on ultrasound images in traditional methods, and achieves objective quantification and accurate identification of the oocyte attachment state.

[0033] Example 2 differs from Example 1 in that it abandons the deep learning segmentation network based on B-mode ultrasound images. Instead, it utilizes the attenuation coefficient and backscatter integral parameter of the original ultrasound radio frequency signal to evaluate the viscoelasticity of follicular fluid and the physical state of the follicular wall. Simultaneously, it extracts the peak time, peak overshoot, and damping ratio of the dynamic response spectrum from the pressure waveform. The attachment level is then output through a gradient boosting tree model, thereby reducing dependence on labeled image data and improving robustness to changes in ultrasound gain or operator technique. Specific differences are as follows: This embodiment directly acquires the raw radio frequency signal data without scanning conversion. The raw radio frequency signal data retains complete echo amplitude, phase, and frequency information without undergoing image post-processing such as logarithmic compression, gain adjustment, and scan transformation. During acquisition, the ultrasound equipment is switched to radio frequency mode, the center frequency of the high-frequency vaginal probe is not less than 10MHz, each radio frequency scan line contains complete depth direction echo sampling points, and the sampling rate is not less than 40MHz. The acquired raw radio frequency data is kept in polar coordinate format for the analysis of the physical parameters of the radio frequency signal.

[0034] Example 1 calculated the morphological parameters (maximum diameter, minimum diameter, wall thickness, and sphericity index) of follicles after segmentation using WIFE-Net; this example, however, directly calculates the viscoelastic parameters of the follicular fluid based on the original radio frequency signal, specifically: The radio frequency signal segment between the anterior and posterior walls of the target follicle is extracted from the raw ultrasound radio frequency signal data as the effective radio frequency signal region of the follicular fluid. The frequency-dependent ultrasound attenuation coefficient is calculated using the spectral shift method: the effective radio frequency signal region of the follicular fluid is divided into multiple overlapping analysis windows (window length 2 mm, overlap rate 50%) along the depth direction. The radio frequency signal in each window is subjected to fast Fourier transform and the centroid frequency of the power spectrum is calculated. The rate of change of the centroid frequency with depth is used as the local attenuation coefficient estimate. The average of the estimates of all windows is taken to obtain the ultrasound attenuation coefficient. The lower the attenuation coefficient, the less protein and cell components in the follicular fluid and the higher the maturity. Calculate the backscatter integral and its cyclic variability: Divide the effective radio frequency signal region of follicular fluid into multiple analysis windows (window length 1mm). After detrending and windowing processing of the radio frequency signal within the window, calculate the power spectrum integral value. Take the logarithm of all integral values ​​and average them to obtain the backscatter integral. Calculate the coefficient of variation between the means of each group of analysis windows to obtain the cyclic variability. The backscatter integral reflects the concentration of tiny particles in the follicular fluid, and the cyclic variability reflects the spatial uniformity of particle distribution. Extract the centroid frequency offset of the power spectrum and convert it into the viscosity coefficient of follicular fluid through offline calibration: calculate the offset of the centroid frequency of the power spectrum relative to the center frequency of the probe in each analysis window, perform linear fitting on the offset along the depth direction, and convert the measured offset rate of change into the viscosity coefficient through the pre-established correspondence between the centroid offset change rate and the standard liquid viscosity.

[0035] Example 1 calculates the resting pressure and pressure fluctuation amplitude (statistical domain characteristics) after extracting a stable pressure waveform from continuous pressure values; this example extracts dynamic response spectrum characteristics from the pressure waveform, specifically: Steady-state pressure waveform segments are extracted from numerical pressure data. These segments are then detrended (using polynomial fitting to eliminate low-frequency drift). A simulated step disturbance is applied to the detrended steady-state pressure waveform segments, and the step signal is convolved with the steady-state pressure fluctuation signal to obtain the system's simulated transient response curve. Three dynamic features are extracted from this response curve: Peak time: The delay time required for the pressure response to rise to its peak value from the moment the step signal is applied, reflecting the inertial characteristics of cumulus cell attachment; Peak overshoot: The percentage of the difference between the maximum peak value and the steady-state final value divided by the steady-state final value, reflecting the elastic modulus characteristics of the attachment structure between the cumulus cell layer and the follicle wall; Damping ratio: Calculated by dividing the natural logarithm of the ratio of the amplitudes of adjacent peaks in the decaying oscillation segment of the response curve by 2π using the logarithmic decay method, it reflects the frictional dissipation capacity between the cumulus cell layer and the follicle wall.

[0036] The above three dynamic characteristics are directly related to the adhesion mechanics of the oocyte-cumulus complex and are insensitive to low-frequency drift.

[0037] Example 1 uses a six-dimensional feature vector composed of the maximum and minimum diameters of follicles, wall thickness, sphericity index, intracavitary resting pressure, and pressure fluctuation amplitude to be input into a random forest model for attachment level classification. In this example, the ultrasound attenuation coefficient, backscatter integral, viscosity coefficient, peak time, peak overshoot, and damping ratio are concatenated in a fixed order to form a six-dimensional feature vector, which is then input into a pre-trained gradient boosting tree model for attachment level classification.

[0038] The gradient boosting tree model is trained using the XGBoost framework. During training, the maximum tree depth is set to 4, the learning rate is 0.1, and the number of iterations is 100. The multi-class log loss function is used as the optimization objective. In each iteration, the new decision tree fits the prediction residuals of all previous trees to gradually reduce the model's prediction error. An early stopping strategy is adopted during training. Training is stopped when the validation set loss function no longer decreases within 10 consecutive iterations to prevent overfitting.

[0039] The model outputs predicted probabilities for high, medium, and low adhesion levels for the input six-dimensional feature vector, and takes the level with the highest probability as the final output adhesion tightness level. The gradient boosting tree model can automatically handle nonlinear interactions between features (such as the coupling between viscosity coefficient and damping ratio) and provide an importance ranking of each feature after training.

[0040] Apart from the differences mentioned above, the synchronous acquisition of pressure numerical data, the temporal alignment of time-series pressure numerical data and ultrasound data, the input preparation after data synchronization, and the method of automatically determining the binding status of the oocyte-cumulus complex and follicle wall based on the attachment level in this embodiment are all the same as the corresponding descriptions in parts S1 and S4 of Embodiment 1, and will not be repeated here.

[0041] This embodiment uses the raw ultrasonic radio frequency signal instead of B-mode image, avoiding the alteration of the original physical information caused by image post-processing such as gain adjustment and logarithmic compression, thus improving the stability of parameter extraction. It also enriches the dynamic dimension of mechanical features by extracting the dynamic response spectrum features of the pressure waveform instead of resting pressure and fluctuation amplitude. Furthermore, it enhances the modeling ability for nonlinear interactions between features by using a gradient boosting tree model instead of a random forest. The two embodiments complement each other and are suitable for different data acquisition conditions and application scenarios.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any brief modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for real-time assessment of sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing, characterized in that: Includes the following steps: S1. Acquire real-time ultrasound image data of the target follicle, as well as corresponding continuous pressure data inside the cavity of the target follicle; S2. Input each frame of ultrasound image data into the pre-trained WIFE-Net, perform parallel operations of conventional convolution and dilated convolution, and aggregate multi-scale spatial information using double pyramid pooling to output a pixel-level segmentation mask of the target follicle region; calculate the maximum diameter, minimum diameter, follicle wall thickness, and follicle sphericity index of the target follicle based on the segmentation mask. S3. Extract a segment of pressure waveform representing the stable mechanical environment within the follicular cavity from the continuous pressure numerical data, calculate the time average value of the pressure waveform as the resting pressure within the follicular cavity, and calculate the standard deviation of the pressure waveform as the pressure fluctuation amplitude. S4. The maximum diameter, the minimum diameter, the follicle wall thickness, the follicle sphericity index, the follicle cavity resting pressure, and the pressure fluctuation amplitude are combined into a six-dimensional feature vector, which is then input into a pre-trained random forest model. The random forest model outputs the degree of adhesion between the oocyte-cumulus complex and the follicle wall.

2. The method for real-time assessment of sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing according to claim 1, characterized in that: Step S1 specifically includes: S11. Real-time ultrasound image data of the target follicle is acquired by ultrasound equipment, wherein the ultrasound image data includes at least imaging information of the anechoic dark area of ​​the follicle region and its boundary contour. S12. Collect continuous pressure data inside the target follicle cavity using a pressure sensor. The pressure data is continuously recorded at a preset sampling frequency to form time-series pressure data. S13. Synchronize the ultrasound image data and the pressure value data using a unified timestamp to establish a time correspondence between each frame of ultrasound image and the corresponding pressure value at that time. S14. Use the synchronized ultrasound image data and pressure value data as input data.

3. The method for real-time assessment of sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing according to claim 1, characterized in that: Step S2 specifically includes: S21. Input each frame of ultrasound image into the encoder of WIFE-Net. The encoder is set to run in parallel with a regular convolution branch and a dilated convolution branch. The dilated convolution branch uses multiple parallel sub-branches with different expansion rates to simultaneously capture the fine boundary of the follicle wall and the contextual information of the tissue surrounding the follicle. S22. The feature maps output by the regular convolutional branches and each dilated convolutional branch are concatenated in the channel dimension, and then the number of channels is compressed through the convolutional layer to fuse local details and multi-scale semantic features. S23. Input the fused feature map into the dual pyramid pooling module, perform average pooling at multiple scales at the image level and pixel level respectively, upsample the pooling results at each scale to the original size and then stitch them with the original feature map to enhance the segmentation robustness for follicles of different sizes and shapes. S24. The enhanced feature map is upsampled step by step by the decoder and then connected in a skip connection to output a binary segmentation mask of the same size as the input image. The follicle contour is extracted from the mask by the boundary tracking algorithm, and the major axis of the minimum circumscribed ellipse of the contour is used as the maximum diameter, the minor axis as the minimum diameter, the average gray-level gradient within the contour as the wall thickness, and the ratio of the major axis to the minor axis as the sphericity index.

4. The method for real-time assessment of sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing according to claim 3, characterized in that: The boundary tracing algorithm adopts the Suzuki algorithm, which traverses the binary segmentation mask image and scans pixel by pixel starting from the top left corner. When an unmarked foreground pixel is found, its eight neighboring boundaries are traced in a clockwise direction until it returns to the starting point, and the pixel coordinate sequence of the boundary point is recorded.

5. The method for real-time assessment of sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing according to claim 1, characterized in that: Step S3 specifically includes: S31. Identify the starting moment when the pressure value drops sharply and the pressure fluctuation amplitude gradually converges to a stable range from the pressure numerical data, and take this moment as the starting mark point of the stable pressure waveform in the follicular cavity. S32. Extract a continuous pressure waveform from the marked point. The duration of the extracted waveform is determined based on the autocorrelation function of the pressure fluctuation in the amniotic follicle: calculate the autocorrelation function of the stable pressure data, find the delay time corresponding to the first decay of the autocorrelation function value to a specified proportion of its maximum value as the decorrelation time, and take the extraction duration as three times the decorrelation time. S33. Calculate the arithmetic mean of all sampling points within the truncated waveform as the resting pressure within the follicular cavity; S34. Calculate the square root of the sum of squares of the differences between each sampling point within the waveform and the resting pressure, and use it as the pressure fluctuation amplitude.

6. The method for real-time assessment of sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing according to claim 5, characterized in that: The formula for calculating the pressure fluctuation amplitude is: Where N is the total number of sampling points within the truncated waveform. The pressure value at the i-th sampling point after median filtering. The resting pressure within the follicular cavity calculated in step S33; the amplitude of this pressure fluctuation. This reflects the degree of dispersion of pressure fluctuations within the amniotic follicle cavity over time.

7. The method for real-time assessment of sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing according to claim 1, characterized in that: Step S4 specifically includes: S41. The maximum diameter, minimum diameter, wall thickness, sphericity index, intracavitary resting pressure, and pressure fluctuation amplitude of the follicle are spliced ​​together in a fixed order to form a six-dimensional fusion feature vector. S42. Input the six-dimensional feature vector into a pre-trained random forest model, and construct multiple decision trees from sheep ex vivo experimental samples through bootstrap aggregation sampling. When splitting a node, each decision tree randomly selects some features from all features and selects the splitting threshold that maximizes the reduction of impurity. S43. The random forest model collects the predicted levels of all decision tree outputs, uses a majority voting mechanism to determine the final result, and counts the frequency of high, medium and low levels in all decision trees. The level with the highest frequency is taken as the attachment tightness level of the final output. S44. Based on the output level of attachment tightness, automatically determine the binding status of the oocyte-cumulus complex and the follicle wall.

8. The method for real-time assessment of sheep follicle development status based on the fusion of ultrasound imaging and pressure sensing according to claim 7, characterized in that: The final result is determined by a majority voting mechanism, expressed as: ； Where E represents the attachment tightness level of the output, and M represents the total number of decision trees in the random forest. Let be the set of leaf nodes of the i-th decision tree. This represents the number of leaf nodes in the tree. Let be the predicted category of leaf node j for the input six-dimensional feature vector v, and let {.} be the indicator function.

Images