Endometrial receptivity detection system and method based on machine learning and medium

Abstract

The invention discloses an endometrial receptivity detection system and method based on machine learning and a medium, and the method comprises the following steps: collecting a continuous multi-frame uterus ultrasonic image sequence, and executing space registration and time standardization processing; segmenting an endometrial region of interest, and processing to obtain an endometrial form dynamic signal sequence; executing time-frequency transformation and local morphological analysis; generating a space-time resonance feature vector through the space-time form resonance network STMR-Net; performing feature fusion to generate an endometrial receptivity embedding vector; calculating receptivity characteristic distribution, generating an endometrial receptivity index, and judging a receptivity grade result; executing frequency spectrum inversion analysis and key feature mapping to generate a receptivity interpretation graph; and outputting a detection result. According to the method, intelligent detection of dynamic oscillation characteristics of the endometrium is realized by utilizing a machine learning and space-time form resonance analysis method, and the method has the advantages of high quantification, high interpretability and high diagnosis precision.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis of medical images, and in particular to a machine learning-based system, method and medium for detecting endometrial receptivity. Background Technology

[0002] Existing endometrial receptivity testing methods mostly rely on two-dimensional ultrasound imaging or hemodynamic static indicators. Empirical analysis is conducted based on features such as grayscale distribution of single-frame images, blood flow signal intensity, and tissue thickness. However, traditional methods cannot capture the dynamic changes of the endometrium during a complete physiological cycle, nor can they reflect the complex coupling relationship between tissue morphology, microblood perfusion, and phase oscillation. This results in highly subjective test results, poor repeatability, and insufficient quantitative accuracy.

[0003] In recent years, although some studies have attempted to introduce deep learning models to automatically extract features from ultrasound images, most of them are limited to static image classification or simple time-series modeling. They have failed to achieve the integration of physical correlation and physiological interpretability between multidimensional features. Existing models lack time-frequency analysis mechanisms for endometrial morphological oscillation characteristics, cannot distinguish resonance modes and tolerance change trends in different frequency ranges, and are difficult to generate quantitative indicators and interpretive maps with clinical guidance significance, thus limiting the practical application value of artificial intelligence technology in assisted reproductive medicine testing. Summary of the Invention

[0004] One objective of this invention is to propose a machine learning-based endometrial receptivity detection system, method, and medium. This invention utilizes machine learning and spatiotemporal morphological resonance analysis to achieve intelligent detection of dynamic oscillation characteristics of the endometrium, possessing the advantages of high quantification, strong interpretability, and high diagnostic accuracy.

[0005] A machine learning-based method for detecting endometrial receptivity according to an embodiment of the present invention includes the following steps: A series of consecutive uterine ultrasound images were acquired, and spatial registration and temporal normalization were performed to obtain a standardized ultrasound sequence with spatiotemporal registration. The region of interest of the endometrium was segmented from the standardized ultrasound sequence, and the grayscale changes, texture gradients, blood flow velocity and direction features of each frame were extracted to construct a set of morphological signals that change over time. Denoising and normalization were then performed to obtain a dynamic signal sequence of endometrial morphology. Based on the dynamic signal sequence of endometrial morphology, time-frequency transformation and local morphological analysis are performed, and amplitude and phase information are separated to generate a set of morphological resonance spectra. The morphological resonance spectrum set is input into the spatiotemporal morphological resonance network STMR-Net, and multi-scale spatial features and multi-frequency band phase consistency features are extracted by spatial convolution units and resonant phase convolution units to generate spatiotemporal resonance feature vectors. Feature fusion is performed based on spatiotemporal resonance feature vectors, and spatial structural features, time series features and phase coupling features are jointly encoded to generate endometrial receptivity embedding vectors. Based on the endometrial receptivity embedding vector, the receptivity feature distribution is calculated, the endometrial receptivity index is generated, and the receptivity level is determined according to the multi-level threshold rule. Based on the tolerance index and tolerance level results, perform spectral inversion analysis and key feature mapping to generate a tolerance interpretation map; The output includes the detection results, including the acceptability index, acceptability level, and acceptability interpretation map.

[0006] Optionally, the uterine ultrasound imaging sequence includes three-dimensional grayscale signals, color Doppler blood flow signals, and time-series identification information.

[0007] Optionally, the generation of the endometrial morphology dynamic signal sequence specifically includes: The endometrial region is located from the spatiotemporally registered standardized ultrasound sequence. An adaptive region segmentation method is used to extract the region of interest containing the endometrial boundary to generate an initial endometrial mask. Morphological closing operations and smoothing filtering are performed on the initial endometrial mask to obtain a region of interest mask with continuous boundaries and complete structure. This mask is then frame-by-frame matched with a standardized ultrasound sequence to form an endometrial region of interest sequence. Using the region of interest in the endometrium as the analysis scope, grayscale changes, texture gradients, blood flow velocity, and blood flow direction features are extracted from each frame of standardized ultrasound images. The extracted features are then combined to form a set of frame-level morphological feature vectors. Perform time series reconstruction on the set of frame-level morphological feature vectors to generate a set of time series signals; Perform denoising and normalization processing on the time series signal set; The normalized time series signal set is reorganized into a standardized feature matrix according to the order of gray level change, texture gradient, blood flow velocity, and blood flow direction; The standardized feature matrix is ​​arranged in a time sequence to form a dynamic signal sequence of endometrial morphology.

[0008] Optionally, the generation of the morphological resonance spectrum set specifically includes: Based on the dynamic signal sequence of endometrial morphology, a time-frequency transformation is performed, and a short-time Fourier transform is performed on the time signal of each pixel in the region of interest of the endometrium to generate a morphological energy distribution map. Based on the morphological energy distribution map, local morphological analysis is performed. By using the local structure operator to slide in the time and frequency domains, the local time-frequency patterns, energy accumulation regions and energy boundary morphology of the morphological energy distribution map are detected. The local morphological parameters, energy concentration region boundaries and time-frequency connectivity features of the endometrial signal on the time-frequency plane are extracted to form a local time-frequency morphological feature matrix. The amplitude and phase information of the local time-frequency morphological feature matrix are separated to generate an amplitude information matrix and a phase information matrix; Local smoothing and noise suppression are performed on the amplitude information matrix to obtain a smoothed amplitude information matrix. Phase continuity processing is performed on the phase information matrix to correct the phase difference between adjacent sampling points in the time series, eliminate period jumps and maintain the smoothness and continuity of phase changes, and calculate the phase consistency matrix. Within the region of interest in the endometrium, for each spatial location, the smoothed amplitude information matrix and phase consistency matrix are combined to calculate the local average oscillation energy and phase synchronization index. The two are then weighted and combined to generate the local morphological resonance intensity distribution matrix of the endometrium. A sliding window accumulation operation is performed on the morphological resonance intensity distribution matrix along the frequency dimension to obtain the endometrial comprehensive oscillation energy distribution matrix; Based on the morphological energy distribution map, a corresponding index between time-frequency coordinates and spatial coordinates is established. The weighted phase continuity values ​​in the phase consistency matrix are accumulated in the spatial domain according to the corresponding index to generate a spatial projection phase consistency matrix. The spatial projection phase consistency matrix and the endometrial comprehensive oscillation energy distribution matrix are weighted element-wise to generate a set of morphological resonance spectra.

[0009] Optionally, the generation of the spatiotemporal resonance feature vector specifically includes: The set of morphological resonance spectra is input into the spatiotemporal morphological resonance network STMR-Net. The input data is standardized, and the resonance intensity values ​​corresponding to each spatial location in the set of morphological resonance spectra are rearranged according to the spatial coordinate index and frequency index to construct the input tensor. The spatiotemporal morphological resonance network STMR-Net includes spatial convolution units, resonant phase convolution units, spatiotemporal coupling units, and fully connected layers; In the spatial convolutional unit of STMR-Net, multi-scale convolution operations are performed on the input tensor along the spatial dimension to obtain multi-scale spatial feature maps. In the resonant phase convolution unit of STMR-Net, a phase convolution operation is performed on the input tensor along the frequency dimension to obtain the phase feature map; Spatial feature maps and phase feature maps are spliced ​​and fused in the frequency dimension to form a joint feature map, and batch normalization is performed on the joint feature map to obtain a joint time-frequency feature map. In the spatiotemporal coupling unit of STMR-Net, the joint time-frequency feature map is spatiotemporally convolved along the time series dimension to generate a spatiotemporal feature tensor. Perform pooling operation on the spatiotemporal feature tensor to obtain the resonant feature matrix after spatial dimensionality reduction; In the fully connected layer, the resonant feature matrix is ​​expanded in the frequency dimension and a weighted linear transformation is performed. The feature values ​​at each spatial location are weighted and summed and a bias is superimposed to output a local spatiotemporal resonant feature vector. A feature integration operation is performed on the local spatiotemporal resonance feature vectors corresponding to all spatial locations, and a spatiotemporal resonance feature vector is generated by a feature concatenation and weighted averaging strategy.

[0010] Optionally, the generation of the endometrial receptivity embedding vector specifically includes: Normalize the spatiotemporal resonance feature vector to obtain a standardized spatiotemporal resonance feature vector; The standardized spatiotemporal resonance feature vectors are classified into spatial structure feature vectors, time series feature vectors, and phase coupling feature vectors according to feature categories. Perform local attention weighting on the spatial structure feature vectors to generate weighted spatial structure feature vectors; Perform a sliding weighting operation on the time series feature vectors to generate dynamic time feature vectors; Perform band fusion on the phase-coupled feature vector to generate a phase-combined feature vector; The weighted spatial structure feature vector, the time dynamic feature vector, and the phase comprehensive feature vector are concatenated sequentially according to the feature dimension to form a fusion feature matrix, and channel normalization is performed within the fusion feature matrix; A multilayer perceptron is used to perform an encoding operation on the fused feature matrix to generate an encoded result vector; The encoded vector is subjected to feature compression to extract the principal component features with the highest energy percentage, and an endometrial receptivity embedding vector is generated.

[0011] Optionally, the generation of the tolerance level result specifically includes: The receptivity feature distribution of the endometrial receptivity embedding vector is calculated, and all components of the endometrial receptivity embedding vector are normalized to generate a standardized receptivity embedding vector. Statistical calculations are performed on all components in the standardized receptive embedding vector. Local feature response intensities are calculated based on spatial distribution dimension, time series dimension, and phase coupling dimension, respectively. Weighted summation is then applied to the feature response intensities of each dimension to generate a receptive feature distribution. The receptivity feature distribution is normalized as a whole, and the weighted average of all elements in the receptivity feature distribution is calculated to obtain the endometrial receptivity index. The endometrial receptivity index is compared with preset multi-level thresholds. When the receptivity index is in the first threshold range, it is determined to be a low receptivity level; when the receptivity index is in the second threshold range, it is determined to be a medium receptivity level; and when the receptivity index is in the third threshold range, it is determined to be a high receptivity level.

[0012] Optionally, the generation of the receptivity interpretation map specifically includes: The endometrial receptivity index and receptivity grade results were obtained. The receptivity index was used as the input for spectral inversion, and the receptivity grade results were used as the frequency resolution control parameter. The frequency inversion interval is determined based on the numerical range of the tolerance index, and the frequency inversion interval is divided according to the preset frequency interval to obtain the frequency component sequence; Perform energy backtracking calculations on each frequency component in the frequency component sequence, distribute the comprehensive energy value corresponding to the tolerance index along the frequency dimension, calculate the energy contribution intensity of each frequency component, and generate a frequency energy distribution sequence. In the frequency energy distribution sequence, the location of the energy peak is identified, the frequency interval corresponding to the energy peak is marked as the dominant resonance frequency interval, and different intervals are color-coded and classified according to the tolerance level results to generate the spectrum inversion results. The frequency index of the spatiotemporal resonance feature vector is matched with the spectrum inversion results to determine the corresponding region of each dominant resonance frequency interval in the spatial dimension, and the corresponding spatial location and time series fragment are extracted to generate a key feature set. For each spatial region in the key feature set, a weighting is performed, using the energy contribution intensity of each region as the weighting coefficient, and the feature values ​​of the regions are summed to generate a key feature mapping matrix; The key feature mapping matrix is ​​arranged in spatial coordinate order and normalized to form an acceptable interpretation map.

[0013] An endometrial receptivity detection system based on machine learning according to an embodiment of the present invention includes: The ultrasound image acquisition module is used to acquire multiple consecutive frames of uterine ultrasound images and perform spatial registration and temporal standardization processing to obtain a standardized ultrasound sequence with spatiotemporal registration. The feature extraction module is used to segment the region of interest in the endometrium from the standardized ultrasound sequence, extract features, and generate a dynamic signal sequence of endometrial morphology. The time-frequency feature analysis module is used to perform time-frequency transformation and local morphological analysis on the dynamic signal sequence of endometrial morphology, separate amplitude information and phase information, and generate a set of morphological resonance spectra. The spatiotemporal morphological resonance feature extraction module is used to extract multi-scale spatial features and multi-frequency band phase consistency features to generate spatiotemporal resonance feature vectors. The feature fusion module is used to fuse and jointly encode the spatiotemporal resonance feature vectors to generate endometrial receptivity embedding vectors. The receptivity calculation module is used to calculate the receptivity feature distribution based on the endometrial receptivity embedding vector, and generate the endometrial receptivity index and receptivity level results. The spectrum inversion module is used to perform spectrum inversion analysis and key feature mapping based on the tolerance index and tolerance level results, and generate a tolerance interpretation map. The results output module is used to output the detection results, which include the acceptability index, acceptability level, and acceptability interpretation map.

[0014] According to an embodiment of the present invention, a computer-readable storage medium stores a computer program that, when executed by a processor, enables the processor to perform a machine learning-based endometrial receptivity detection method.

[0015] The beneficial effects of this invention are: This invention overcomes the shortcomings of traditional ultrasound imaging, which relies on human experience and lacks dynamic cycle analysis and interpretable quantitative indicators, by introducing a machine learning-based endometrial receptivity detection method and system. It achieves multi-dimensional intelligent modeling and accurate assessment of the physiological state of the endometrium. Starting from continuous ultrasound image sequences, this invention constructs a multi-frame dynamic signal sequence that includes spatial structural features, time series features, and hemodynamic features. It can completely capture the dynamic changes in the morphological structure and microblood perfusion of the endometrium during the menstrual cycle. By performing time-frequency transformation and local morphological analysis on the morphological signals, amplitude information and phase information are separated, and a set of morphological resonance spectra is generated, thereby achieving a joint description of the local oscillation energy and phase stability of the endometrium.

[0016] At the deep learning modeling level, this invention proposes the Spatiotemporal Morphological Resonance Network (STMR-Net), which combines spatial convolutional units and resonant phase convolutional units to perform multi-scale spatial feature extraction and multi-band phase consistency analysis on the input morphological resonance spectrum set, thereby obtaining a spatiotemporal resonance feature vector that can reflect the tissue resonance characteristics. By further jointly encoding and fusing spatial structural features, time series features, and phase coupling features, an endometrial receptivity embedding vector is generated, realizing a high-dimensional representation of the physiological state of the endometrium. After the receptivity feature distribution is calculated and multi-level threshold is determined, the embedding vector can output the endometrial receptivity index and corresponding level results, effectively improving the quantification and stability of the detection.

[0017] Furthermore, this invention proposes an interpretability mechanism based on spectral inversion, which uses the tolerance index and grading results to infer the frequency energy distribution characteristics, identify the dominant resonance frequency range, and generate a tolerance interpretation map. This enables visualization of the test results and explanation of the physiological mechanism. This technology allows doctors to intuitively understand the role of different spatial regions and frequency oscillations in the tolerance formation process, significantly enhancing the transparency and clinical usability of the model. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a machine learning-based endometrial receptivity detection method proposed in this invention; Figure 2 This is a schematic diagram of the spatiotemporal morphological resonance network STMR-Net, which is a machine learning-based endometrial receptivity detection method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 and Figure 2 A machine learning-based method for detecting endometrial receptivity includes the following steps: A series of consecutive uterine ultrasound images were acquired, and spatial registration and temporal normalization were performed to obtain a standardized ultrasound sequence with spatiotemporal registration. The region of interest of the endometrium was segmented from the standardized ultrasound sequence, and the grayscale changes, texture gradients, blood flow velocity and direction features of each frame were extracted to construct a set of morphological signals that change over time. Denoising and normalization were then performed to obtain a dynamic signal sequence of endometrial morphology. Based on the dynamic signal sequence of endometrial morphology, time-frequency transformation and local morphological analysis are performed, and amplitude and phase information are separated to generate a set of morphological resonance spectra. The morphological resonance spectrum set is input into the spatiotemporal morphological resonance network STMR-Net, and multi-scale spatial features and multi-frequency band phase consistency features are extracted by spatial convolution units and resonant phase convolution units to generate spatiotemporal resonance feature vectors. Feature fusion is performed based on spatiotemporal resonance feature vectors, and spatial structural features, time series features and phase coupling features are jointly encoded to generate endometrial receptivity embedding vectors. Based on the endometrial receptivity embedding vector, the receptivity feature distribution is calculated, the endometrial receptivity index is generated, and the receptivity level is determined according to the multi-level threshold rule. Based on the tolerance index and tolerance level results, perform spectral inversion analysis and key feature mapping to generate a tolerance interpretation map; The output includes the acceptance index, acceptance level, and acceptance interpretation map, enabling dynamic detection and interpretable quantitative analysis of endometrial microblood perfusion, tissue morphology changes, and their periodic oscillation patterns.

[0021] In this embodiment, the uterine ultrasound imaging sequence includes three-dimensional grayscale signals, color Doppler blood flow signals, and time series identification information, which are used to characterize the spatial structural features, hemodynamic features, and temporal evolution features of the endometrium.

[0022] In this embodiment, the generation of the dynamic signal sequence of endometrial morphology specifically includes: The endometrial region is located from the spatiotemporally registered standardized ultrasound sequence. An adaptive region segmentation method is used to extract the region of interest containing the endometrial boundary to generate an initial endometrial mask. Morphological closing operations and smoothing filtering were performed on the initial endometrial mask to remove boundary artifacts and isolated noise points, resulting in a continuous and structurally complete endometrial region of interest mask. This mask was then frame-by-frame matched with a standardized ultrasound sequence to form an endometrial region of interest sequence. Using the region of interest in the endometrium as the analysis scope, grayscale variation, texture gradient, blood flow velocity, and blood flow direction features are extracted from each frame of standardized ultrasound image. The extracted features are combined to form a set of frame-level morphological feature vectors. The grayscale variation is used to reflect tissue density variation, the texture gradient is used to describe the distribution of microstructure orientation, and the blood flow velocity and direction are used to characterize local microblood flow perfusion characteristics. Perform time series reconstruction on the frame-level morphological feature vector set, connect the features at the same spatial location in consecutive frames in time order, and generate a time series signal set that reflects the dynamic changes of the endometrium. Perform denoising and normalization processing on the time series signal set; The normalized time series signal set is reorganized into a standardized feature matrix according to the order of gray level change, texture gradient, blood flow velocity, and blood flow direction; The standardized feature matrix is ​​arranged temporally to form a dynamic signal sequence of endometrial morphology. The dynamic signal sequence of endometrial morphology includes spatial structural features, blood perfusion features and temporal change information, which is used to describe the structural changes and microblood perfusion dynamic behavior of the endometrium in the time dimension.

[0023] In this embodiment, the generation of the morphological resonance spectrum set specifically includes: Based on the dynamic signal sequence of endometrial morphology, a time-frequency transformation is performed. Short-time Fourier transform is performed on the time signal of each pixel in the region of interest of the endometrium to generate a morphological energy distribution map with time as the horizontal axis and frequency as the vertical axis. This map is used to characterize the oscillation energy distribution of endometrial grayscale change features, texture gradient features and blood flow signal features at different frequencies. Based on the morphological energy distribution map, local morphological analysis is performed. By using the local structure operator to slide in the time and frequency domains, the local time-frequency patterns, energy accumulation regions and energy boundary morphology of the morphological energy distribution map are detected. The local morphological parameters, energy concentration region boundaries and time-frequency connectivity features of the endometrial signal on the time-frequency plane are extracted to form a local time-frequency morphological feature matrix. The generation of the local time-frequency morphological feature matrix specifically includes: setting the time window length and frequency window width of the local structure operator; sliding the morphological energy distribution map point by point along the time and frequency directions with a fixed step size; within each window, performing a summation operation on all energy values ​​and dividing by the number of pixels in the window to obtain the local average energy value of that window; subtracting the minimum energy value from the maximum energy value in the window to obtain the energy fluctuation amplitude value, which is used to reflect the intensity of local energy change; performing a division calculation on the energy value difference between adjacent time points in the window to obtain the energy change rate per unit time, which is recorded as the local time gradient value; performing a division calculation on the energy value difference between adjacent frequency points to obtain the energy change rate per unit frequency, which is recorded as the local frequency gradient value; squaring the time gradient value and the frequency gradient value respectively, adding them together, and taking the square root to obtain the energy gradient amplitude, which is used to characterize the energy boundary intensity; setting an energy threshold, and filtering values ​​higher than the energy threshold... Energy points are marked as valid oscillation points. Adjacent valid oscillation points are detected one by one. When the time interval and frequency interval between adjacent points are both less than a preset interval threshold, they are determined to be connected. The average energy value of the two points is divided by the sum of the energy in the window as the weighted connectivity weight. The weighted connectivity weights of all connected pairs are accumulated, and the accumulated result is the time-frequency connectivity value of the local energy accumulation area, which is used to represent the continuity of the high-energy region in the time and frequency dimensions. Then, the rate of change of the energy boundary line morphology is calculated in the same window. The sum of the energy gradient amplitudes of continuous boundary segments is divided by the number of boundary segments to obtain the regional morphological parameter value, which is used to describe the stability of the local oscillation morphology. The local average energy value, energy fluctuation amplitude value, energy gradient amplitude, time-frequency connectivity value and regional morphological parameter value obtained in each sliding window are arranged in time and frequency position order to generate a local time-frequency morphological feature matrix. The amplitude information and phase information of the local time-frequency morphological feature matrix are separated. The amplitude information is used to represent the oscillation intensity, and the phase information is used to represent the relationship between the oscillation rhythm direction and time displacement, thereby generating an amplitude information matrix and a phase information matrix. The generation of the amplitude information matrix and phase information matrix specifically includes: weighting and summing the local average energy value and energy fluctuation amplitude value corresponding to each time point and frequency point in the local time-frequency morphological feature matrix, with the weight coefficient being the proportion of the two in the total energy of the window, to obtain the amplitude value at that position; performing a division operation between the energy gradient amplitude value and the regional morphological parameter value at the same time point and frequency point to obtain the local phase change rate; using the time-frequency connectivity value as a smoothing weight, performing a weighted average operation on the phase change rate of adjacent time points to obtain the weighted phase change value; then accumulating the weighted phase change value along the time direction to obtain the cumulative phase change total value, and determining the positive or negative phase state identifier according to the energy gradient direction relationship between adjacent frequency points to form a phase state set; arranging the amplitude value of each time point in frequency order to form an amplitude information matrix, and arranging the phase state value corresponding to each time point in frequency order to form a phase information matrix; Local smoothing and noise suppression are performed on the amplitude information matrix. A sliding window filtering method is used in the time and frequency domains to reduce high-frequency random noise and retain stable oscillation components in the mid-to-low frequency range, resulting in a smoothed amplitude information matrix. Phase continuity processing is performed on the phase information matrix to correct the phase difference between adjacent sampling points in the time series, eliminate period jumps and maintain the smoothness and continuity of phase changes, and calculate the phase consistency matrix. The generation of the phase consistency matrix specifically includes: sequentially pairing and calculating the phase state values ​​of adjacent time points in the phase information matrix to obtain a phase difference set; performing a correction operation of adding or subtracting period values ​​on elements in the phase difference set whose absolute value is greater than half a period threshold, so that the corrected phase difference value is within the continuous period range, to obtain a preliminary corrected phase difference set; performing an accumulation operation on the time series phase difference value corresponding to each frequency point in chronological order to obtain a phase cumulative value sequence, which is used to reflect the continuous trend of phase change; then performing a moving average processing on the phase cumulative value sequence to obtain a smoothed phase sequence value; then using the time-frequency connectivity value as a weighting coefficient, performing a weighted average on the smoothed phase sequence corresponding to each frequency point to obtain a weighted phase continuity value; arranging the weighted phase continuity value of each frequency point according to the time index to generate a phase consistency matrix, which is used to characterize the phase stability and continuity of the endometrial signal in the time dimension. The period threshold is used to determine whether the phase change has a jump across the period, and its value is half of the complete oscillation period range of the signal. Within the region of interest in the endometrium, for each spatial location, the smoothed amplitude information matrix and phase consistency matrix are combined to calculate the local average oscillation energy and phase synchronization index. The two are then weighted and combined to generate the local morphological resonance intensity distribution matrix of the endometrium, which is used to reflect the coordinated oscillation characteristics of the tissue structure in the time and frequency dimensions. The generation of the endometrial local morphological resonance intensity distribution matrix specifically includes: at each time point and frequency point, extracting a set of local amplitude values ​​from the smoothed amplitude information matrix, summing all amplitude values ​​in the set and dividing by the number of elements in the set to obtain the local average oscillation energy value; extracting a set of weighted phase continuous values ​​corresponding to the same spatial location from the phase consistency matrix, performing weighted averaging on the weighted phase continuous values ​​of adjacent time points, and taking the time-frequency connectivity value as the weighting coefficient to obtain the local phase synchronization index value; weighting and summing the local average oscillation energy value and the phase synchronization index value according to a preset proportional coefficient to obtain the local morphological resonance intensity value at that spatial location; then arranging all local morphological resonance intensity values ​​sequentially according to the coordinate index of the spatial location to construct the endometrial local morphological resonance intensity distribution matrix, and normalizing the endometrial local morphological resonance intensity distribution matrix; A sliding window accumulation operation is performed on the morphological resonance intensity distribution matrix along the frequency dimension to obtain the endometrial comprehensive oscillation energy distribution matrix, which is used to represent the global oscillation energy characteristics over the entire cycle. Based on the morphological energy distribution map, a corresponding index between time-frequency coordinates and spatial coordinates is established. The weighted phase continuity values ​​in the phase consistency matrix are accumulated in the spatial domain according to the corresponding index to generate a spatial projection phase consistency matrix. The spatial projection phase consistency matrix and the endometrial comprehensive oscillation energy distribution matrix are weighted element-wise to generate a set of morphological resonance spectra, which are used to comprehensively characterize the resonance intensity distribution of the endometrium under the dual dimensions of energy intensity and phase stability.

[0024] In this embodiment, the generation of the spatiotemporal resonance feature vector specifically includes: The set of morphological resonance spectra is input into the spatiotemporal morphological resonance network STMR-Net. The input data is standardized, and the resonance intensity values ​​corresponding to each spatial location in the set of morphological resonance spectra are rearranged according to the spatial coordinate index and frequency index to construct the input tensor. The spatiotemporal morphological resonance network STMR-Net includes spatial convolution units, resonant phase convolution units, spatiotemporal coupling units, and fully connected layers; In the spatial convolution unit of STMR-Net, multi-scale convolution operation is performed on the input tensor along the spatial dimension. Different convolution kernel sizes are used to slide on multiple spatial scales. Weighted summation and bias accumulation operations are performed on the input values ​​in each local region to obtain multi-scale spatial feature mapping, which is used to characterize the morphological changes of endometrial tissue at different spatial scales. The different convolutional kernel sizes include three types: 3×3, 5×5, and 7×7. In the resonant phase convolution unit of STMR-Net, a phase convolution operation is performed on the input tensor along the frequency dimension. The input values ​​of each frequency component are periodically weighted and summed using the resonant phase weight and phase offset as parameters to extract the phase consistency features between different frequency components and obtain the phase feature mapping. Spatial feature maps and phase feature maps are spliced ​​and fused in the frequency dimension to form a joint feature map, and batch normalization is performed on the joint feature map to obtain a joint time-frequency feature map. In the spatiotemporal coupling unit of STMR-Net, the joint time-frequency feature map is spatiotemporally convolved along the time series dimension, and the feature values ​​between adjacent time frames are weighted smoothed and locally summed to generate a spatiotemporal feature tensor, which is used to describe the propagation characteristics of endometrial morphological oscillations in the time dimension. A pooling operation is performed on the spatiotemporal feature tensor to obtain the spatially reduced resonance feature matrix, which is used to compress spatial redundancy and retain the average response features of local significant oscillation regions. In the fully connected layer, the resonance feature matrix is ​​expanded in the frequency dimension and a weighted linear transformation is performed. The feature values ​​at each spatial location are weighted and summed and the bias is superimposed to output a local spatiotemporal resonance feature vector, which is used to represent the comprehensive oscillation characteristics of the local region of the endometrium under multi-scale spatial structure and multi-frequency phase consistency. A feature integration operation is performed on the local spatiotemporal resonance feature vectors corresponding to all spatial locations. A spatiotemporal resonance feature vector is generated by a feature splicing and weighted averaging strategy. The spatiotemporal resonance feature vector is used as the spatiotemporal resonance expression result of the endometrium as a whole.

[0025] In this embodiment, the generation of the endometrial receptivity embedding vector specifically includes: Normalize the spatiotemporal resonance feature vector to obtain a standardized spatiotemporal resonance feature vector; The standardized spatiotemporal resonance feature vectors are classified into spatial structure feature vectors, time series feature vectors, and phase coupling feature vectors according to feature categories. The spatial structure feature vectors are used to characterize the spatial morphological distribution characteristics of the endometrium, the time series feature vectors are used to characterize the temporal variation law of the endometrial oscillation characteristics, and the phase coupling feature vectors are used to characterize the phase synchronization relationship between multi-frequency resonances. The specific steps of classifying by feature category include: sorting all components of the standardized spatiotemporal resonance feature vector and establishing a feature index table according to the encoding order of spatial dimension, time dimension, and phase dimension; extracting a set of spatially distributed feature values ​​from the spatial dimension index, and performing a weighted average of the feature values ​​at the same spatial location based on the neighborhood response weight to obtain a spatial structure feature vector; then extracting a set of time-varying feature values ​​from the time dimension index, and performing a weighted average of the feature values ​​at adjacent time points based on the time window weight to obtain a time series feature vector; and finally extracting a set of multi-frequency resonance feature values ​​from the phase dimension index, and performing a weighted average of the feature values ​​of different frequency components based on the frequency coupling weight to obtain a phase coupling feature vector. Local attention weighting is performed on the spatial structure feature vector. The correlation weight between the feature value and the feature value of the adjacent position is calculated at each spatial position. The correlation weight is multiplied by the original feature value and summed to generate a weighted spatial structure feature vector, which is used to enhance the saliency response of the structure. The generation of the relevant weights specifically includes: within the two-dimensional spatial range corresponding to the spatial structure feature vector, setting a fixed neighborhood window centered on the current spatial position, and extracting the feature value set of adjacent positions within the neighborhood; calculating the difference index between the feature values ​​of adjacent positions and the feature values ​​of the center position, and using the inverse ratio of the difference as the correlation measure; normalizing all correlation measures within the neighborhood; and using the normalized correlation measure as the relevant weight of the corresponding position within the neighborhood. A sliding weighting operation is performed on the time series feature vector. A sliding window of fixed length is set in the time dimension, and the feature values ​​within the window are weighted according to their time position weights to generate a time dynamic feature vector, which is used to reflect the continuous characteristics of the endometrial morphology evolution over time. Frequency band fusion is performed on the phase coupling feature vector. The feature values ​​corresponding to different frequency components are weighted and averaged according to the frequency coupling weight. The weighted result is normalized to generate a phase comprehensive feature vector, which is used to characterize the overall coherence among multi-frequency oscillations. The weighted spatial structure feature vector, the time dynamic feature vector, and the phase comprehensive feature vector are concatenated sequentially according to the feature dimension to form a fusion feature matrix, and channel normalization is performed within the fusion feature matrix; A multilayer perceptron is used to perform encoding operations on the fused feature matrix to generate an encoding result vector, which is used to establish a mapping between spatial structure features, time series features and phase coupling features in a unified vector space. Feature compression is performed on the encoded result vector to extract the principal component features with the highest energy proportion, generating an endometrial receptivity embedding vector, which is used to comprehensively characterize the fusion state of spatial structural features, time series features and phase coupling features.

[0026] In this embodiment, the generation of the tolerance level result specifically includes: The receptivity feature distribution of the endometrial receptivity embedding vector is calculated, and all components of the endometrial receptivity embedding vector are normalized to generate a standardized receptivity embedding vector. Statistical calculations are performed on all components in the standardized receptive embedding vector. Local feature response intensities are calculated based on spatial distribution dimension, time series dimension, and phase coupling dimension, respectively. Weighted summation is then applied to the feature response intensities of each dimension to generate a receptive feature distribution. The generation of the receptivity feature distribution specifically includes: dividing the standardized receptivity embedding vector into dimensions according to spatial distribution dimension, time series dimension, and phase coupling dimension, and extracting the corresponding feature component sets for each dimension; within the spatial distribution dimension, establishing a fixed neighborhood window centered on each spatial location, calculating the feature difference degree between adjacent locations for all feature components within the window, and using the inverse ratio of the feature difference degree as the local spatial response weight, performing a weighted average on the feature components within the neighborhood to obtain the spatial local feature response intensity; within the time series dimension, using a fixed time window as a unit, calculating the feature components at adjacent time points within the window according to time... The adjacency weight is used to perform a weighted average to obtain the temporal local feature response intensity. Within the phase coupling dimension, the feature components of different frequency components are weighted according to the frequency coupling weight, and the phase continuity consistency is calculated in the frequency neighborhood. The phase continuity consistency is used as a weight adjustment factor to correct the weighting result to obtain the phase local feature response intensity. The spatial local feature response intensity, temporal local feature response intensity, and phase local feature response intensity are normalized respectively, and weighted according to preset spatial, temporal, and phase weight coefficients to generate a receptivity feature distribution, which is used to characterize the overall response state of the endometrium under multi-dimensional features. The receptivity feature distribution is normalized as a whole, and the weighted average of all elements in the receptivity feature distribution is calculated to obtain the endometrial receptivity index. The endometrial receptivity index is compared with preset multi-level thresholds. When the receptivity index is in the first threshold range, it is determined to be a low receptivity level; when the receptivity index is in the second threshold range, it is determined to be a medium receptivity level; and when the receptivity index is in the third threshold range, it is determined to be a high receptivity level.

[0027] In this embodiment, the generation of the receptivity interpretation map specifically includes: The endometrial receptivity index and receptivity grade results were obtained. The receptivity index was used as the input for spectral inversion, and the receptivity grade results were used as the frequency resolution control parameter. The frequency inversion interval is determined based on the numerical range of the tolerance index, and the frequency inversion interval is divided according to the preset frequency interval to obtain the frequency component sequence; The generation of the frequency component sequence specifically includes: reading the value of the endometrial receptivity index, comparing the value with a preset receptivity threshold range, and determining the start and end frequencies of the target frequency inversion range based on the threshold range in which the receptivity index is located; setting a fixed frequency step size within the determined frequency inversion range, and dividing the interval from the start frequency to the end frequency into several consecutive frequency intervals according to the frequency step size, with each interval corresponding to an independent frequency component; assigning an index identifier to each frequency component and recording its corresponding frequency value, start and end boundaries, and positional order within the overall inversion range to form a frequency component index table; extracting the center frequency values ​​of all frequency components sequentially according to the index order to construct an ordered frequency component sequence; performing an integrity check on the frequency component sequence to ensure that the frequency intervals are consistent and cover the entire frequency inversion range, and outputting the checked frequency component sequence. Perform energy backtracking calculations on each frequency component in the frequency component sequence, distribute the comprehensive energy value corresponding to the tolerance index along the frequency dimension, calculate the energy contribution intensity of each frequency component, and generate a frequency energy distribution sequence. The generation of the frequency energy distribution sequence specifically includes: reading the frequency component sequence and its corresponding index information, treating each frequency component as an independent calculation unit; calling the value of the endometrial receptivity index as a global energy benchmark, mapping this value to a comprehensive energy value, representing the overall receptivity's total energy level in the frequency domain; smoothly initializing the energy proportion of adjacent frequency components according to their position order in the sequence, and proportionally distributing the comprehensive energy value to all frequency components to generate an initial energy allocation set; then calculating the energy difference value of each frequency component relative to its adjacent components based on the frequency interval and energy gradient relationship between adjacent frequency components, and superimposing the energy difference value with the initial energy value to obtain a corrected energy contribution value; normalizing the energy contribution values ​​of all frequency components; and arranging the normalized energy contribution values ​​sequentially according to the frequency index order to form a frequency energy distribution sequence. In the frequency energy distribution sequence, the location of the energy peak is identified, the frequency interval corresponding to the energy peak is marked as the dominant resonance frequency interval, and different intervals are color-coded and classified according to the tolerance level results to generate the spectrum inversion results. The generation of the spectrum inversion results specifically includes: reading the energy contribution values ​​of all frequency components in the frequency energy distribution sequence; comparing the energy contribution values ​​of three adjacent frequency components as a group using a sliding window; calculating the energy contribution difference between the middle frequency component and the two side components; if the difference is positive and exceeds a set energy threshold, then the frequency component is determined to be an energy peak point; recording the frequency index and corresponding energy contribution value of each energy peak point; grouping adjacent peak points into several continuous dominant resonance frequency intervals according to the frequency interval between them, and assigning a unique number to each interval; matching the color labels corresponding to different levels with the threshold levels according to the tolerance level results, color-coding and classifying each dominant resonance frequency interval to form a frequency level mapping table; then combining the frequency level mapping table with the dominant resonance frequency interval index table, synchronously recording the frequency range, energy peak, and color level of each interval to generate spectrum labeling data; arranging all dominant resonance frequency intervals in sequence according to the frequency index order to construct a spectrum inversion result matrix, and outputting the spectrum inversion results after color labeling and level classification; The frequency index of the spatiotemporal resonance feature vector is matched with the spectrum inversion results to determine the corresponding region of each dominant resonance frequency interval in the spatial dimension, and the corresponding spatial location and time series fragment are extracted to generate a key feature set. For each spatial region in the key feature set, a weighting is performed, using the energy contribution intensity of each region as the weighting coefficient, and the feature values ​​of the regions are summed to generate a key feature mapping matrix; The key feature mapping matrix is ​​arranged in spatial coordinate order and normalized to form an acceptability interpretation map, which is used to show the main roles and response distributions of different frequency ranges and spatial regions in the acceptability formation process.

[0028] A machine learning-based endometrial receptivity detection system includes: The ultrasound image acquisition module is used to acquire multiple consecutive frames of uterine ultrasound images and perform spatial registration and temporal standardization processing to obtain a standardized ultrasound sequence with spatiotemporal registration. The feature extraction module is used to segment the region of interest in the endometrium from the standardized ultrasound sequence, extract features, and generate a dynamic signal sequence of endometrial morphology. The time-frequency feature analysis module is used to perform time-frequency transformation and local morphological analysis on the dynamic signal sequence of endometrial morphology, separate amplitude information and phase information, and generate a set of morphological resonance spectra. The spatiotemporal morphological resonance feature extraction module is used to extract multi-scale spatial features and multi-frequency band phase consistency features to generate spatiotemporal resonance feature vectors. The feature fusion module is used to fuse and jointly encode the spatiotemporal resonance feature vectors to generate endometrial receptivity embedding vectors. The receptivity calculation module is used to calculate the receptivity feature distribution based on the endometrial receptivity embedding vector, and generate the endometrial receptivity index and receptivity level results. The spectrum inversion module is used to perform spectrum inversion analysis and key feature mapping based on the tolerance index and tolerance level results, and generate a tolerance interpretation map. The results output module is used to output the detection results, which include the acceptability index, acceptability level, and acceptability interpretation map.

[0029] A computer-readable storage medium storing a computer program that, when executed by a processor, enables the processor to perform a machine learning-based endometrial receptivity detection method. Example

[0030] To verify the feasibility of this invention in actual medical testing, it was applied to the reproductive medicine center of a tertiary hospital for quantitative analysis and intelligent assessment of endometrial receptivity. This hospital requires precise assessment of endometrial receptivity levels during assisted reproductive treatment to guide embryo transfer timing. However, traditional two-dimensional ultrasound and manual judgment methods rely primarily on doctors' subjective interpretation of grayscale images, lacking quantitative expression of temporal changes and failing to fully reflect the dynamic oscillation patterns of the endometrium during the cycle. This is particularly prone to misjudgment regarding blood perfusion and subtle changes in tissue morphology. This invention, by introducing machine learning and spatiotemporal morphological resonance analysis, achieves multidimensional joint modeling of endometrial morphology, blood flow signals, and their temporal oscillations, thus overcoming the shortcomings of traditional detection methods in dynamic quantification and time-frequency correlation analysis.

[0031] During implementation, a series of three-dimensional ultrasound images are acquired from the hospital's existing color Doppler ultrasound equipment. These image data include grayscale signals and Doppler blood flow signals. Spatial registration and temporal standardization are performed by the ultrasound image acquisition module of this invention to ensure that different frames are strictly aligned in space and maintain continuity in the temporal dimension. Based on this, the system automatically identifies the endometrial region and extracts the grayscale changes, texture gradients, blood flow velocity and direction features of each frame through the feature extraction module. After normalization and noise removal, a dynamic morphological signal sequence reflecting the changes of the endometrium over time is formed, which can capture the rhythmic changes of microblood flow perfusion and the fluctuation characteristics of tissue microstructure at different cyclic stages.

[0032] After the morphological dynamic signal sequence is generated, the system automatically executes the time-frequency feature analysis module to perform a short-time Fourier transform on the signal, mapping the time series to the time-frequency plane to obtain the energy distribution of the endometrium at different frequencies. Subsequently, local morphological features, energy accumulation regions, and time-frequency connectivity features are extracted using local morphological analysis methods. The local oscillation energy and morphological stability are calculated through a sliding window time-frequency scanning process. After separating the amplitude information and phase information, the system generates a set of morphological resonance spectra for subsequent modeling. The amplitude information represents the energy intensity, and the phase information represents the oscillation direction and rhythm synchronization state. The phase consistency matrix obtained after phase continuity processing can accurately describe the continuous trend of oscillation in the time dimension, thereby avoiding the phase instability problem caused by periodic jumps in traditional methods.

[0033] In the feature extraction stage, the spatiotemporal morphological resonance network STMR-Net designed in this invention fully leverages the multi-scale feature capture capability of deep learning. The network uses a spatial convolution module to extract morphological change features at different spatial scales, captures the phase consistency between multi-band signals through a resonant phase convolution module, and realizes continuous feature aggregation in the time dimension in the spatiotemporal coupling unit. After fully connected layers and feature integration operations, the spatiotemporal resonance feature vector generated by the system can completely characterize the oscillation behavior of the endometrium in the three-dimensional domains of time, space and frequency, providing high-precision input for acceptability analysis.

[0034] In the receptivity analysis phase, the system uses a feature fusion module to divide the spatiotemporal resonance feature vector into three parts: spatial structural features, time series features, and phase coupling features. After joint encoding by a multilayer perceptron, an endometrial receptivity embedding vector is generated. Subsequently, the receptivity calculation module performs distribution analysis on the features of each dimension of the embedding vector, calculates the receptivity feature distribution, and generates a receptivity index. By setting multi-level threshold intervals, the receptivity index is divided into three levels: low, medium, and high, enabling automatic determination of endometrial receptivity. The trend of this index can intuitively reflect the coordinated state of tissue structure and microcirculation during the cycle, providing doctors with quantifiable diagnostic evidence.

[0035] In the results visualization stage, the spectrum inversion module of this invention generates a frequency energy distribution sequence by performing frequency inversion calculation on the receptivity index and identifies the dominant resonance frequency range. Combined with phase consistency analysis, the system generates a receptivity interpretation map, enabling doctors to intuitively see the key feature areas that have the greatest impact on endometrial receptivity at different frequencies and spatial locations. Through this interpretable visualization method, not only can the overall grade results be obtained, but it can also be clearly identified which part of the tissue structure or blood flow signal changes most significantly, thereby improving the reliability of clinical decision-making.

[0036] To verify the performance of the present invention, it was compared with the traditional method. The comparison results are shown in Table 1.

[0037] As can be seen from Table 1, the endometrial receptivity detection method based on machine learning proposed in this invention is superior to traditional methods in all key indicators.

[0038] In terms of the accuracy of tolerance level determination, the present invention achieves 93.8%, which is 14.5 percentage points higher than the traditional manual evaluation method and more than 7 percentage points higher than the two-dimensional gray-scale statistical method. This shows that after the STMR-Net network integrates multi-dimensional spatiotemporal features, it can more accurately distinguish different tolerance level stages and avoid errors caused by subjective differences in images in manual judgment.

[0039] In terms of consistency determination, traditional methods rely on doctors' subjective judgment, which can easily lead to inconsistent results. However, this invention uses a deep learning feature fusion mechanism to stabilize the feature distribution among different samples, increasing the Kappa value to 0.91, indicating that the algorithm has significant advantages in repeatability and reliability.

[0040] In terms of detection time comparison, the method of the present invention can complete the entire detection process in only about 4.3 minutes, saving about 70% of the time compared with the manual evaluation method. This is due to the efficient feature extraction and automated reasoning mechanism of the spatiotemporal morphological resonance network STMR-Net, which can output the tolerance index and grade results without manual intervention, thereby improving the efficiency of clinical detection.

[0041] Regarding the feature extraction error rate and temporal stability index, this invention reduces the error rate to 2.6% and the temporal stability index to 0.08. The core reason is that it adopts a feature expression method that combines time-frequency transformation and phase consistency matrix, which makes the response to local morphological changes and periodic oscillations smoother and more stable, avoiding noise interference and instantaneous signal drift.

[0042] Traditional methods struggle to capture synchronized rhythms under multi-frequency resonance in terms of phase synchronization recognition rate. However, this invention introduces a resonant phase convolution structure in the frequency domain, which can effectively identify the synergistic changes between high-frequency and low-frequency oscillations, increasing the recognition rate to 95.2%. This improvement enhances the ability to characterize microvascular perfusion rhythms and tissue response rhythms, providing a solid basis for judging the periodic receptivity of the intima.

[0043] Furthermore, at the system operation level, this invention excels in average processing frame rate and comprehensive clinical usability score. The system can achieve a processing speed of 72 frames per second and can respond to operation input in real time.

[0044] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based method for detecting endometrial receptivity, characterized in that, Includes the following steps: A series of consecutive uterine ultrasound images were acquired, and spatial registration and temporal normalization were performed to obtain a standardized ultrasound sequence with spatiotemporal registration. The region of interest of the endometrium was segmented from the standardized ultrasound sequence, and the grayscale changes, texture gradients, blood flow velocity and direction features of each frame were extracted to construct a set of morphological signals that change over time. Denoising and normalization were then performed to obtain a dynamic signal sequence of endometrial morphology. Based on the dynamic signal sequence of endometrial morphology, time-frequency transformation and local morphological analysis are performed, and amplitude and phase information are separated to generate a set of morphological resonance spectra. The morphological resonance spectrum set is input into the spatiotemporal morphological resonance network STMR-Net, and multi-scale spatial features and multi-frequency band phase consistency features are extracted by spatial convolution units and resonant phase convolution units to generate spatiotemporal resonance feature vectors. Feature fusion is performed based on spatiotemporal resonance feature vectors, and spatial structural features, time series features and phase coupling features are jointly encoded to generate endometrial receptivity embedding vectors. Based on the endometrial receptivity embedding vector, the receptivity feature distribution is calculated, the endometrial receptivity index is generated, and the receptivity level is determined according to the multi-level threshold rule. Based on the tolerance index and tolerance level results, perform spectral inversion analysis and key feature mapping to generate a tolerance interpretation map; The output includes the detection results, including the acceptability index, acceptability level, and acceptability interpretation map.

2. The method for detecting endometrial receptivity based on machine learning according to claim 1, characterized in that, The uterine ultrasound imaging sequence includes three-dimensional grayscale signals, color Doppler blood flow signals, and time-series identification information.

3. The method for detecting endometrial receptivity based on machine learning according to claim 1, characterized in that, The generation of the dynamic signal sequence of endometrial morphology specifically includes: The endometrial region is located from the spatiotemporally registered standardized ultrasound sequence. An adaptive region segmentation method is used to extract the region of interest containing the endometrial boundary to generate an initial endometrial mask. Morphological closing operations and smoothing filtering are performed on the initial endometrial mask to obtain a region of interest mask with continuous boundaries and complete structure. This mask is then frame-by-frame matched with a standardized ultrasound sequence to form an endometrial region of interest sequence. Using the region of interest in the endometrium as the analysis scope, grayscale changes, texture gradients, blood flow velocity, and blood flow direction features are extracted from each frame of standardized ultrasound images. The extracted features are then combined to form a set of frame-level morphological feature vectors. Perform time series reconstruction on the set of frame-level morphological feature vectors to generate a set of time series signals; Perform denoising and normalization processing on the time series signal set; The normalized time series signal set is reorganized into a standardized feature matrix according to the order of gray level change, texture gradient, blood flow velocity, and blood flow direction; The standardized feature matrix is ​​arranged in a time sequence to form a dynamic signal sequence of endometrial morphology.

4. The method for detecting endometrial receptivity based on machine learning according to claim 1, characterized in that, The generation of the morphological resonance spectrum set specifically includes: Based on the dynamic signal sequence of endometrial morphology, a time-frequency transformation is performed, and a short-time Fourier transform is performed on the time signal of each pixel in the region of interest of the endometrium to generate a morphological energy distribution map. Based on the morphological energy distribution map, local morphological analysis is performed. By using the local structure operator to slide in the time and frequency domains, the local time-frequency patterns, energy accumulation regions and energy boundary morphology of the morphological energy distribution map are detected. The local morphological parameters, energy concentration region boundaries and time-frequency connectivity features of the endometrial signal on the time-frequency plane are extracted to form a local time-frequency morphological feature matrix. The amplitude and phase information of the local time-frequency morphological feature matrix are separated to generate an amplitude information matrix and a phase information matrix; Local smoothing and noise suppression are performed on the amplitude information matrix to obtain a smoothed amplitude information matrix. Phase continuity processing is performed on the phase information matrix to correct the phase difference between adjacent sampling points in the time series, eliminate period jumps and maintain the smoothness and continuity of phase changes, and calculate the phase consistency matrix. Within the region of interest in the endometrium, for each spatial location, the smoothed amplitude information matrix and phase consistency matrix are combined to calculate the local average oscillation energy and phase synchronization index. The two are then weighted and combined to generate the local morphological resonance intensity distribution matrix of the endometrium. A sliding window accumulation operation is performed on the morphological resonance intensity distribution matrix along the frequency dimension to obtain the endometrial comprehensive oscillation energy distribution matrix; Based on the morphological energy distribution map, a corresponding index between time-frequency coordinates and spatial coordinates is established. The weighted phase continuity values ​​in the phase consistency matrix are accumulated in the spatial domain according to the corresponding index to generate a spatial projection phase consistency matrix. The spatial projection phase consistency matrix and the endometrial comprehensive oscillation energy distribution matrix are weighted element-wise to generate a set of morphological resonance spectra.

5. The method for detecting endometrial receptivity based on machine learning according to claim 1, characterized in that, The generation of the spatiotemporal resonance feature vector specifically includes: The morphological resonance spectrum set is input into the spatiotemporal morphological resonance network STMR-Net. The input data is standardized, and the resonance intensity values ​​corresponding to each spatial location in the morphological resonance spectrum set are rearranged according to the spatial coordinate index and frequency index to construct the input tensor. The spatiotemporal morphological resonance network STMR-Net includes spatial convolution units, resonant phase convolution units, spatiotemporal coupling units, and fully connected layers; In the spatial convolutional unit of STMR-Net, multi-scale convolution operations are performed on the input tensor along the spatial dimension to obtain multi-scale spatial feature maps. In the resonant phase convolution unit of STMR-Net, a phase convolution operation is performed on the input tensor along the frequency dimension to obtain the phase feature map; Spatial feature maps and phase feature maps are spliced ​​and fused in the frequency dimension to form a joint feature map, and batch normalization is performed on the joint feature map to obtain a joint time-frequency feature map. In the spatiotemporal coupling unit of STMR-Net, the joint time-frequency feature map is spatiotemporally convolved along the time series dimension to generate a spatiotemporal feature tensor. Perform pooling operation on the spatiotemporal feature tensor to obtain the resonant feature matrix after spatial dimensionality reduction; In the fully connected layer, the resonant feature matrix is ​​expanded in the frequency dimension and a weighted linear transformation is performed. The feature values ​​at each spatial location are weighted and summed and a bias is superimposed to output a local spatiotemporal resonant feature vector. A feature integration operation is performed on the local spatiotemporal resonance feature vectors corresponding to all spatial locations, and a spatiotemporal resonance feature vector is generated by a feature concatenation and weighted averaging strategy.

6. The method for detecting endometrial receptivity based on machine learning according to claim 1, characterized in that, The generation of the endometrial receptivity embedding vector specifically includes: Normalize the spatiotemporal resonance feature vector to obtain a standardized spatiotemporal resonance feature vector; The standardized spatiotemporal resonance feature vectors are classified into spatial structure feature vectors, time series feature vectors, and phase coupling feature vectors according to feature categories. Perform local attention weighting on the spatial structure feature vectors to generate weighted spatial structure feature vectors; Perform a sliding weighting operation on the time series feature vectors to generate dynamic time feature vectors; Perform band fusion on the phase-coupled feature vector to generate a phase-combined feature vector; The weighted spatial structure feature vector, the time dynamic feature vector, and the phase comprehensive feature vector are concatenated sequentially according to the feature dimension to form a fusion feature matrix, and channel normalization is performed within the fusion feature matrix; A multilayer perceptron is used to perform an encoding operation on the fused feature matrix to generate an encoded result vector; The encoded vector is subjected to feature compression to extract the principal component features with the highest energy percentage, and an endometrial receptivity embedding vector is generated.

7. The method for detecting endometrial receptivity based on machine learning according to claim 1, characterized in that, The generation of the tolerance level result specifically includes: The receptivity feature distribution of the endometrial receptivity embedding vector is calculated, and all components of the endometrial receptivity embedding vector are normalized to generate a standardized receptivity embedding vector. Statistical calculations are performed on all components in the standardized receptive embedding vector. Local feature response intensities are calculated based on spatial distribution dimension, time series dimension, and phase coupling dimension, respectively. Weighted summation is then applied to the feature response intensities of each dimension to generate a receptive feature distribution. The receptivity feature distribution is normalized as a whole, and the weighted average of all elements in the receptivity feature distribution is calculated to obtain the endometrial receptivity index. The endometrial receptivity index is compared with preset multi-level thresholds. When the receptivity index is in the first threshold range, it is determined to be a low receptivity level; when the receptivity index is in the second threshold range, it is determined to be a medium receptivity level; and when the receptivity index is in the third threshold range, it is determined to be a high receptivity level.

8. The method for detecting endometrial receptivity based on machine learning according to claim 1, characterized in that, The generation of the receptivity interpretation map specifically includes: The endometrial receptivity index and receptivity grade results were obtained. The receptivity index was used as the input for spectral inversion, and the receptivity grade results were used as the frequency resolution control parameter. The frequency inversion interval is determined based on the numerical range of the tolerance index, and the frequency inversion interval is divided according to the preset frequency interval to obtain the frequency component sequence; Perform energy backtracking calculations on each frequency component in the frequency component sequence, distribute the comprehensive energy value corresponding to the tolerance index along the frequency dimension, calculate the energy contribution intensity of each frequency component, and generate a frequency energy distribution sequence. In the frequency energy distribution sequence, the location of the energy peak is identified, the frequency interval corresponding to the energy peak is marked as the dominant resonance frequency interval, and different intervals are color-coded and classified according to the tolerance level results to generate the spectrum inversion results. The frequency index of the spatiotemporal resonance feature vector is matched with the spectrum inversion results to determine the corresponding region of each dominant resonance frequency interval in the spatial dimension, and the corresponding spatial location and time series fragment are extracted to generate a key feature set. For each spatial region in the key feature set, a weighting is performed, using the energy contribution intensity of each region as the weighting coefficient, and the feature values ​​of the regions are summed to generate a key feature mapping matrix; The key feature mapping matrix is ​​arranged in spatial coordinate order and normalized to form an acceptable interpretation map.

9. A machine learning-based endometrial receptivity detection system, comprising the machine learning-based endometrial receptivity detection method according to any one of claims 1 to 8, characterized in that, include: The ultrasound image acquisition module is used to acquire multiple consecutive frames of uterine ultrasound images and perform spatial registration and temporal standardization processing to obtain a standardized ultrasound sequence with spatiotemporal registration. The feature extraction module is used to segment the region of interest in the endometrium from the standardized ultrasound sequence, extract features, and generate a dynamic signal sequence of endometrial morphology. The time-frequency feature analysis module is used to perform time-frequency transformation and local morphological analysis on the dynamic signal sequence of endometrial morphology, separate amplitude information and phase information, and generate a set of morphological resonance spectra. The spatiotemporal morphological resonance feature extraction module is used to extract multi-scale spatial features and multi-frequency band phase consistency features to generate spatiotemporal resonance feature vectors. The feature fusion module is used to fuse and jointly encode the spatiotemporal resonance feature vectors to generate endometrial receptivity embedding vectors. The receptivity calculation module is used to calculate the receptivity feature distribution based on the endometrial receptivity embedding vector, and generate the endometrial receptivity index and receptivity level results. The spectrum inversion module is used to perform spectrum inversion analysis and key feature mapping based on the tolerance index and tolerance level results, and generate a tolerance interpretation map. The results output module is used to output the detection results, which include the acceptability index, acceptability level, and acceptability interpretation map.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the processor to perform a machine learning-based endometrial receptivity detection method as described in any one of claims 1 to 8.

Images