A fetal heart monitoring image classification method and system based on neural network

Abstract

The present invention relates to the field of image classification technology, and in particular to a fetal heart monitoring image classification method and system based on neural network. The method comprises the following steps: collecting fetal heart monitoring images, and generating standard monitoring images by Gaussian noise elimination and spatial domain normalization, then performing multi-channel convolution on the standard image to extract image convolution features, and obtaining standard fetal heart features by compression regularization optimization, performing morphological reconstruction to generate a fetal heart dynamic model, and identifying baseline features based on historical data, and then performing periodic variation quantification and safe floating range derivation on the fusion features, and finally constructing a multi-scale discrimination framework to realize intelligent classification of fetal heart monitoring images. The present invention realizes a more stable and accurate fetal heart monitoring image classification method.

Description

Technical Field

[0001] The present invention relates to the technical field of image classification, and in particular to a fetal heart monitoring image classification method and system based on neural network. Background Art

[0002] The analysis of fetal heart rate monitoring images is of great significance for ensuring the health of mothers and infants. With the continuous advancement of obstetric technology, the application of fetal heart rate monitoring has become more and more common. However, traditional image processing methods rely on manual judgment and are easily affected by the subjective experience of medical staff. This dependence limits the accuracy and consistency of monitoring results, which in turn affects clinical decision-making and maternal and infant safety. Therefore, a more intelligent and automated analysis method is urgently needed to improve the classification accuracy and reliability of fetal heart rate monitoring images. Gaussian noise and other interferences often exist in fetal heart rate monitoring images, which will interfere with the effect of feature extraction and classification. In addition, there are significant differences in fetal heart rate monitoring images of different patients, resulting in insufficient generalization ability of existing models and difficulty in adapting to diverse clinical scenarios. These problems make it difficult for existing image classification methods to achieve good results in practical applications. Summary of the invention

[0003] Based on this, it is necessary to provide a fetal heart monitoring image classification method and system based on a neural network to solve at least one of the above technical problems.

[0004] To achieve the above object, a fetal heart monitoring image classification method based on a neural network comprises the following steps:

[0005] Step S1: Acquire a fetal heart monitoring image; perform Gaussian noise elimination on the fetal heart monitoring image to obtain a noise-reduced monitoring image; perform spatial domain normalization on the noise-reduced monitoring image to generate a standard monitoring image;

[0006] Step S2: performing multi-channel convolution on the standard monitoring image to generate image convolution features; performing compression regularization optimization on the image convolution features to obtain standardized fetal heart features;

[0007] Step S3: morphologically reconstructing the standard fetal heart characteristics to generate a fetal heart dynamic model; performing baseline feature recognition on the fetal heart dynamic model based on preset historical fetal heart monitoring data to obtain a fusion feature baseline;

[0008] Step S4: performing period variation quantification processing on the fusion feature baseline based on the standard fetal heart characteristics to obtain period variation data; deriving a safe floating range of the fusion feature baseline based on the period variation data to generate a baseline floating range;

[0009] Step S5: construct a multi-scale discrimination framework for the standard fetal heart characteristics according to the baseline floating range to obtain a classification discrimination framework; classify the fetal heart monitoring image based on the classification discrimination framework to realize an intelligent classification method for fetal heart monitoring images based on a neural network.

[0010] The present invention significantly improves the image quality and ensures the accuracy of subsequent processing by eliminating Gaussian noise and normalizing the spatial domain during the classification process of fetal heart monitoring images. The multi-channel convolution is used to enhance the feature extraction capability so that important information in the image can be effectively retained. The compression regularization optimization optimizes the feature dimension and reduces the computational complexity. The morphological reconstruction provides richer structural information for the fetal heart dynamic model. The extraction of the fusion feature baseline provides a more reliable reference basis for subsequent analysis. The periodic variation quantization processing makes the feature analysis more refined. The derivation of the safe floating range provides an effective boundary condition for classification. The constructed multi-scale discrimination framework enhances the adaptability of the model to different features, and finally realizes the efficient classification of fetal heart monitoring images, improves the accuracy and stability of classification, promotes the application of intelligent analysis technology in the monitoring system, and helps to promote research progress and technological innovation in related fields.

[0011] Preferably, step S1 comprises the following steps:

[0012] Step S11: Acquire a fetal heart monitoring image; perform multi-resolution decomposition on the fetal heart monitoring image to obtain multi-layer decomposition data;

[0013] Step S12: performing Gaussian kernel optimization processing on the multi-layer decomposition data to generate kernel function parameters;

[0014] Step S13: reconstructing the kernel function parameters by noise suppression to obtain a noise-reduced monitoring image;

[0015] Step S14: performing local contrast stretching on the noise reduction monitoring image to generate a local enhanced monitoring image;

[0016] Step S15: performing spatial domain normalization processing on the local enhanced monitoring image to generate a standard monitoring image.

[0017] The present invention realizes deep information extraction of fetal heart monitoring images through multi-resolution decomposition technology, kernel optimization processing improves the clarity and distinguishability of image features, noise suppression reconstruction significantly reduces interference signals in the image and ensures data quality, local contrast stretching enhances the detail display of the image and makes important features more prominent, and spatial domain normalization provides a standardized data basis for subsequent processing, which improves the image processing efficiency and analysis accuracy as a whole, promotes the application and development of intelligent classification systems in the field of fetal heart monitoring, enhances the robustness and adaptability of the system, and promotes the continuous progress and innovation of related technologies.

[0018] Preferably, step S2 comprises the following steps:

[0019] Step S21: performing bidirectional recursive convolution on the standard monitoring image to generate image convolution features;

[0020] Step S22: performing dimensionality reduction compression processing on the image convolution features to obtain compressed feature data;

[0021] Step S23: performing regularization optimization on the compressed feature data to generate optimized feature data;

[0022] Step S24: performing normalization conversion on the optimized characteristic data to obtain normalized fetal heart characteristics.

[0023] The present invention improves the ability to extract image features and enhances the capture of complex patterns through bidirectional recursive convolution technology. Dimensionality reduction and compression processing reduces data redundancy and improves computational efficiency. Regularization optimization makes feature data more consistent and reduces the risk of overfitting in model training. Normalization transformation further unifies the scale of feature data and ensures the stability of subsequent analysis. It improves the accuracy and efficiency of the system in processing fetal heart monitoring images as a whole, promotes the reliability and effectiveness of intelligent classification, and promotes the continuous evolution and innovation of related technologies.

[0024] Preferably, step S21 includes the following steps:

[0025] Step S211: performing forward recursive convolution on the standard monitoring image to obtain weighted forward data; performing reverse recursive convolution on the standard monitoring image to generate modulated reverse data;

[0026] Step S212: performing feature alignment processing on the weighted forward data and the modulated reverse data to generate feature alignment data;

[0027] Step S213: performing cross-time step correlation on the feature alignment data to obtain correlated feature data; performing time series dependency modeling on the correlated feature data to generate a correlated dependency model;

[0028] Step S214: performing a deformable convolution transformation based on the association dependency model to obtain deformed feature data; performing a dilated convolution processing on the deformed feature data to generate dilated feature data;

[0029] Step S215: performing group convolution on the expanded feature data to obtain group feature data; performing separable convolution on the group feature data to generate image separation features;

[0030] Step S216: performing residual connection enhancement on the image separation feature to generate an image residual feature; and performing feature recalibration on the image separation feature based on the image residual feature to obtain an image convolution feature.

[0031] The present invention provides multi-dimensional feature information by combining forward recursive convolution with backward recursive convolution. The feature alignment of weighted forward data and modulated reverse data improves consistency and accuracy. The processing of cross-time step association enhances the coherence of time series features. Time series dependency modeling provides a dynamic perspective for data analysis. The deformable convolution transformation flexibly adapts to the morphology of different features. The dilated convolution expansion increases the receptive field and enriches the expressiveness of features. The group convolution distinction improves the selectivity and effectiveness of features. The separable convolution significantly reduces the computational complexity. The residual connection enhances the transmission efficiency and stability of features. The feature recalibration ensures the accuracy of the final convolution features. The overall performance and robustness of the system in image classification tasks are improved, which promotes the innovation and application development of intelligent classification technology.

[0032] Preferably, step S211 includes the following steps:

[0033] Perform basic frame segmentation on the standard monitoring image to obtain time-series image frames;

[0034] Perform forward recursive convolution on the time-series image frames to generate forward feature map data;

[0035] Perform weighted superposition processing on the forward feature map data to obtain weighted forward data;

[0036] Perform mirror flip simulation on the standard monitoring image to obtain a mirror monitoring image;

[0037] Perform time-reversal evolution on standard monitoring images to generate time-reversed data;

[0038] Performing reverse recursive convolution on the mirror monitoring image based on the time-reversed data to obtain reverse feature mapping data;

[0039] The inverse feature map data is modulated and transformed to generate modulated inverse data.

[0040] The present invention realizes the effective extraction of time-series image frames through basic frame segmentation technology. The forward feature mapping data generated by forward recursive convolution enhances the feature capture capability. The weight superposition processing improves the representativeness and effectiveness of feature data. The mirror flip simulation provides diversity for feature extraction. The time reverse data generated by time reversal evolution enriches the dynamic information. The reverse recursive convolution processing further captures the potential features in the image. The transformation processing improves the flexibility and adaptability of the features. The overall accuracy and efficiency of image classification are improved, which promotes the development and application of intelligent classification technology based on neural networks.

[0041] Preferably, step S3 comprises the following steps:

[0042] Step S31: performing peak detection on the standard fetal heart characteristics to obtain a peak point set; performing valley value recognition on the peak point set to generate a valley point set;

[0043] Step S32: reconstructing the fluctuation contour based on the peak point set and the trough point set to generate a fetal heart dynamic model;

[0044] Step S33: Perform historical baseline statistics on the preset historical fetal heart monitoring data to obtain a historical fetal heart dynamic baseline;

[0045] Step S34: performing baseline feature matching on the fetal heart rate dynamics model according to the historical fetal heart rate dynamics baseline to obtain a fused feature baseline.

[0046] The present invention accurately identifies important features in fetal heart signals through peak detection technology. The generation of peak point sets and trough point sets provides a basis for the construction of dynamic models. Fluctuation contour reconstruction realizes the visualization and analysis of fetal heart dynamic models. The baseline statistics of historical fetal heart monitoring data provide a historical basis for feature matching. Baseline feature matching ensures the consistency of dynamic models with historical data. The fusion feature baseline improves the understanding and analysis depth of fetal heart signals, optimizes the accuracy and reliability of image classification as a whole, and promotes the application and development of intelligent classification technology in the field of fetal heart monitoring.

[0047] Preferably, step S4 comprises the following steps:

[0048] Step S41: performing time series calibration mapping on the standard fetal heart characteristics and the fusion characteristic baseline to obtain calibration synchronization data;

[0049] Step S42: performing cycle feature recognition on the calibration synchronization data to generate periodic variation data;

[0050] Step S43: reconstructing the trend of the periodic variation data by quantitative analysis to obtain the variation characteristic curvature; deriving the steady-state threshold value based on the variation characteristic curvature to generate threshold prediction data;

[0051] Step S44: Perform dynamic floating calibration processing on the fused feature baseline based on the threshold prediction data to produce a baseline floating range.

[0052] The present invention realizes the synchronous alignment of the standardized fetal heart characteristics and the fusion feature baseline through the time series calibration mapping technology. The cyclic feature recognition effectively captures the periodic changes of the fetal heart signal. The generation of periodic variation data enriches the understanding of dynamic features. The trend quantification reconstruction provides the clarity of the change features. The curvature of the change features lays the foundation for subsequent analysis. The steady-state threshold derivation ensures the accuracy and reliability of the dynamic analysis. The threshold prediction data provides a reference for the baseline floating calibration. The dynamic floating calibration processing improves the adaptability and stability of the fusion feature baseline, which improves the accuracy and effectiveness of image classification as a whole and promotes the continuous development and application of intelligent classification technology.

[0053] Preferably, step S43 includes the following steps:

[0054] Rearranging the periodic variation data into a phase sequence to obtain rearranged sequence data; and performing curvature sampling marking based on the rearranged sequence data to obtain curvature marking data;

[0055] Perform high-dimensional mapping transformation on the curvature labeled data to obtain a high-dimensional mapping matrix;

[0056] Perform feature reconstruction on the high-dimensional mapping matrix to generate reconstructed feature data; perform curvature normalization on the reconstructed feature data to obtain the change feature curvature;

[0057] The extreme point of the changing characteristic curvature is located to generate extreme value sequence data; the extreme value sequence data is decomposed into stable intervals to generate a steady-state interval sequence;

[0058] Steady-state thresholds are derived based on steady-state interval sequences to generate threshold prediction data.

[0059] The present invention improves the structuring of periodic variation data through phase sequence rearrangement technology, curvature sampling labeling provides a clear reference for feature extraction, high-dimensional mapping transformation enhances the expressiveness of features, the generation of reconstructed feature data ensures the integrity of feature information, curvature normalization processing improves the consistency and comparability of features, extreme point positioning accurately identifies key feature changes, stable interval decomposition provides a reliable data basis for subsequent analysis, steady-state threshold derivation ensures the accuracy and practicality of feature dynamic analysis, and overall optimizes the accuracy and stability of image classification, promoting the application and development of intelligent classification technology in related fields.

[0060] Preferably, step S5 comprises the following steps:

[0061] Step S51: performing time-sharing window segmentation according to the baseline floating range to obtain time multi-scale features; performing multi-band filtering on the time multi-scale features to generate frequency multi-scale features;

[0062] Step S52: performing multi-amplitude interval statistics on the frequency multi-scale feature to obtain the amplitude multi-scale feature;

[0063] Step S53: setting decision thresholds for the time multi-scale features, the frequency multi-scale features and the amplitude multi-scale features based on the preset threshold interval to obtain an initial discrimination rule;

[0064] Step S54: Perform multi-level combination modeling on the initial discrimination rules to obtain a classification discrimination framework;

[0065] Step S55: Classify the fetal heart monitoring image based on the classification and discrimination framework to implement an intelligent classification method for fetal heart monitoring images based on a neural network.

[0066] The present invention realizes multi-scale analysis of time data through time-sharing window segmentation technology, frequency band filtering processing improves the frequency resolution of the signal, the generation of amplitude multi-scale features enhances the understanding of the amplitude of data changes, the decision threshold setting provides a clear standard for feature classification, the construction of initial discrimination rules ensures the scientificity and rationality of classification, multi-level combined modeling improves the flexibility and adaptability of the classification framework, and the establishment of a classification discrimination framework optimizes the efficiency and accuracy of image classification, which overall promotes the application and progress of intelligent classification technology in fetal heart monitoring image analysis.

[0067] The present invention also provides a fetal heart monitoring image classification system based on a neural network, which is used to execute the fetal heart monitoring image classification method based on a neural network as described above. The fetal heart monitoring image classification system based on a neural network includes:

[0068] An image acquisition module is used to acquire fetal heart monitoring images; perform Gaussian noise elimination on the fetal heart monitoring images to obtain noise-reduced monitoring images; perform spatial domain normalization on the noise-reduced monitoring images to generate standard monitoring images;

[0069] The image convolution module is used to perform multi-channel convolution on the standard monitoring image to generate image convolution features; the image convolution features are compressed and regularized to optimize to obtain standardized fetal heart features;

[0070] The fusion baseline recognition module is used to perform morphological reconstruction of the standard fetal heart characteristics and generate a fetal heart dynamic model; perform baseline feature recognition on the fetal heart dynamic model based on the preset historical fetal heart monitoring data to obtain a fusion feature baseline;

[0071] The floating range derivation module is used to perform period variation quantification processing on the fusion feature baseline based on the standard fetal heart characteristics to obtain period variation data; the safe floating range of the fusion feature baseline is deduced according to the period variation data to generate the baseline floating range;

[0072] The discrimination framework building module is used to build a multi-scale discrimination framework for standard fetal heart characteristics according to the baseline floating range to obtain a classification discrimination framework; based on the classification discrimination framework, the fetal heart monitoring image is classified and processed to realize an intelligent classification method for fetal heart monitoring images based on a neural network.

[0073] The present invention improves the image quality through efficient image processing technology through the image acquisition module, laying a solid foundation for subsequent analysis, the noise reduction processing effectively removes the interference signal and enhances the clarity of the image, the spatial domain normalization standardizes the image data, making the images comparable, the image convolution module extracts rich feature information using multi-channel convolution technology, the compression regularization optimization reduces redundancy and improves the processing efficiency of features, the fusion baseline recognition module provides an in-depth understanding of the dynamic model through morphological reconstruction, the baseline feature recognition based on historical data improves the accuracy and consistency of feature recognition, the floating range derivation module enhances the sensitivity to data changes through periodic variation quantitative analysis, and the generated safe floating range provides a clear reference for subsequent discrimination, the discrimination framework building module improves the flexibility and adaptability of the model through multi-scale analysis, and finally realizes the efficient classification of fetal heart monitoring images, promotes the technological progress and expansion of the application scope of the intelligent monitoring system, and improves the performance and reliability of the system as a whole. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 A schematic diagram of the steps of a fetal heart monitoring image classification method based on a neural network;

[0075] Figure 2 Detailed implementation flow chart of step S2;

[0076] Figure 3 is a schematic diagram of a detailed implementation process of step S3;

[0077] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0078] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0079] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0080] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0081] To achieve this, please refer to Figures 1 to 3 , a fetal heart monitoring image classification method based on neural network, comprising the following steps:

[0082] Step S1: Acquire a fetal heart monitoring image; perform Gaussian noise elimination on the fetal heart monitoring image to obtain a noise-reduced monitoring image; perform spatial domain normalization on the noise-reduced monitoring image to generate a standard monitoring image;

[0083] Step S2: performing multi-channel convolution on the standard monitoring image to generate image convolution features; performing compression regularization optimization on the image convolution features to obtain standardized fetal heart features;

[0084] Step S3: morphologically reconstructing the standard fetal heart characteristics to generate a fetal heart dynamic model; performing baseline feature recognition on the fetal heart dynamic model based on preset historical fetal heart monitoring data to obtain a fusion feature baseline;

[0085] Step S4: performing period variation quantification processing on the fusion feature baseline based on the standard fetal heart characteristics to obtain period variation data; deriving a safe floating range of the fusion feature baseline based on the period variation data to generate a baseline floating range;

[0086] Step S5: construct a multi-scale discrimination framework for the standard fetal heart characteristics according to the baseline floating range to obtain a classification discrimination framework; classify the fetal heart monitoring image based on the classification discrimination framework to realize an intelligent classification method for fetal heart monitoring images based on a neural network.

[0087] The present invention significantly improves the image quality and ensures the accuracy of subsequent processing by eliminating Gaussian noise and normalizing the spatial domain during the classification process of fetal heart monitoring images. The multi-channel convolution is used to enhance the feature extraction capability so that important information in the image can be effectively retained. The compression regularization optimization optimizes the feature dimension and reduces the computational complexity. The morphological reconstruction provides richer structural information for the fetal heart dynamic model. The extraction of the fusion feature baseline provides a more reliable reference basis for subsequent analysis. The periodic variation quantization processing makes the feature analysis more refined. The derivation of the safe floating range provides an effective boundary condition for classification. The constructed multi-scale discrimination framework enhances the adaptability of the model to different features, and finally realizes the efficient classification of fetal heart monitoring images, improves the accuracy and stability of classification, promotes the application of intelligent analysis technology in the monitoring system, and helps to promote research progress and technological innovation in related fields.

[0088] In an embodiment of the present invention, the neural network-based fetal heart monitoring image classification method comprises the following steps:

[0089] Step S1: Acquire a fetal heart monitoring image; perform Gaussian noise elimination on the fetal heart monitoring image to obtain a noise-reduced monitoring image; perform spatial domain normalization on the noise-reduced monitoring image to generate a standard monitoring image;

[0090] In this embodiment, a professional device that supports fetal heart monitoring signal conversion to image is used, such as a signal collected by a fetal heart monitoring instrument, to convert it into image data with a standard resolution of 1024×1024 pixels. A Gaussian filtering method is used to eliminate Gaussian noise from the fetal heart monitoring image. The standard deviation of the Gaussian kernel is set to 2.0, and a 5×5 convolution kernel is used for convolution operation to filter out random noise interference in the image. After noise reduction is completed, the image pixel intensity distribution is adjusted by histogram equalization to evenly distribute its grayscale value in the range of 0 to 255. At the same time, all pixel values ​​of the image are adjusted to the range of 0 to 1 through the maximum and minimum normalization method of pixel intensity, thereby generating a standardized fetal heart monitoring image as input for subsequent processing.

[0091] Step S2: performing multi-channel convolution on the standard monitoring image to generate image convolution features; performing compression regularization optimization on the image convolution features to obtain standardized fetal heart features;

[0092] In this embodiment, the standard monitoring image is input into a convolutional neural network (CNN) for multi-channel convolution processing. A pre-trained ResNet50 network model is selected, and its convolution kernel size is set to 3×3, the convolution step size is 1, and the number of channels is 64. During the processing, ReLU (Rectified Linear Unit) is used as the activation function to extract deep image texture and spatial features. The convolution features are compressed by a specific regularization method, such as using L2 regularization to constrain the weight of the feature, and the regularization factor is set to 0.01. At the same time, all convolution features are standardized pixel by pixel to remove redundant features and generate standardized fetal heart image features.

[0093] Step S3: morphologically reconstructing the standard fetal heart characteristics to generate a fetal heart dynamic model; performing baseline feature recognition on the fetal heart dynamic model based on preset historical fetal heart monitoring data to obtain a fusion feature baseline;

[0094] In this embodiment, the standardized fetal heart features are taken as input, and the features are reconstructed using morphological operators. Specifically, a combination of expansion and erosion operations is used. The size of the structural element is set to 5×5 to capture the local dynamic changes of the fetal heart features. After reconstruction, the sliding window technology (the window length is 30 frames, and the sliding step is 10 frames) is used to extract the dynamic change model of the fetal heart. This model is compared and matched with the historical monitoring data. The historical data is obtained by querying the database, and the content includes the fetal heart rate and baseline data under the same gestational age. The baseline features are identified and matched by the dynamic time warping (Dynamic TimeWarping, DTW) algorithm. The similarity threshold is set to 0.9 in the identification process to obtain the feature baseline generated by the fusion of the current features and the historical data.

[0095] Step S4: performing period variation quantification processing on the fusion feature baseline based on the standard fetal heart characteristics to obtain period variation data; deriving a safe floating range of the fusion feature baseline based on the period variation data to generate a baseline floating range;

[0096] In this embodiment, the standardized fetal heart characteristics and the fused characteristic baseline are input into the periodic analysis module, the periodicity of the fetal heart characteristics is analyzed by Fourier transform, the fundamental frequency and the main periodic components are extracted, the sampling frequency is set to 4 Hz, and the periodic variation is quantified by calculating the peak spacing and standard deviation of the characteristic signal to generate periodic variation data. The quantified periodic variation data is input into the floating range derivation module, and the safe floating range is calculated based on the maximum value, minimum value and mean standard deviation. The floating range is limited by a threshold standard of no more than 15% up and down, and finally a baseline floating range is generated for subsequent operations.

[0097] Step S5: construct a multi-scale discrimination framework for the standard fetal heart characteristics according to the baseline floating range to obtain a classification discrimination framework; classify the fetal heart monitoring image based on the classification discrimination framework to realize an intelligent classification method for fetal heart monitoring images based on a neural network.

[0098] In this embodiment, the baseline floating range and the standard fetal heart features are used as input to construct a multi-scale discrimination framework. The framework is built through a pyramid feature network (FPN) to fuse and classify fetal heart features at different scales. The basic resolution of the feature pyramid is set to 128×128 pixels, and it is upsampled to 512×512 pixels layer by layer. At the same time, the cross entropy loss function is used as the classification standard. The classification head part in the discrimination framework is implemented through a fully connected layer and a Softmax (normalized exponential function) layer. Finally, the classification framework is applied to the classification task of fetal heart monitoring images. According to the probability distribution results output by the model, the input fetal heart monitoring images are intelligently classified.

[0099] Preferably, step S1 comprises the following steps:

[0100] Step S11: Acquire a fetal heart monitoring image; perform multi-resolution decomposition on the fetal heart monitoring image to obtain multi-layer decomposition data;

[0101] Step S12: performing Gaussian kernel optimization processing on the multi-layer decomposition data to generate kernel function parameters;

[0102] Step S13: reconstructing the kernel function parameters by noise suppression to obtain a noise-reduced monitoring image;

[0103] Step S14: performing local contrast stretching on the noise reduction monitoring image to generate a local enhanced monitoring image;

[0104] Step S15: performing spatial domain normalization processing on the local enhanced monitoring image to generate a standard monitoring image.

[0105] In this embodiment, a fetal heart monitoring device is used to collect fetal heart monitoring images, and the original image data is converted into a standard digital image format. The resolution is set to 1920×1080 pixels. The fetal heart monitoring image is decomposed into multiple resolutions using wavelet transform. Daubechies wavelet (db4) is selected as the mother wavelet function, and the number of decomposition layers is set to 3. The decomposition result includes three detail components and one low-frequency component, wherein the detail component contains high-frequency information such as image edges, and the low-frequency component retains the main contour and background information. Each component is stored in a matrix format for subsequent processing. Gaussian kernel optimization processing is introduced into the multi-resolution decomposition result, and support vector machine (SVM) kernel function optimization technology is used. The bandwidth parameter of the Gaussian kernel function is initialized to 0.5, and particle swarm optimization (Particle Swarm Optimization, The PSO algorithm was used to optimize the Gaussian kernel parameters, set the particle population size to 20, the number of iterations to 50, and minimize the classification error as the objective function. The optimal kernel parameters were obtained through iteration, and the optimized kernel function parameters were finally used for subsequent noise suppression processing. Combining the optimized Gaussian kernel parameters and the multi-layer decomposition results, Gaussian kernel filtering was used to suppress the image noise. First, Gaussian smoothing filtering was applied to the detail components of each layer to eliminate high-frequency noise. The filter bandwidth parameter was set to the optimized value of 0.42. Then, the wavelet coefficients were denoised by combining the soft threshold method. The threshold was dynamically calculated based on the standard deviation of the noise. After filtering and denoising, the components of each layer were reconstructed by wavelet. The low-frequency component and the detail component were synthesized into a complete image. The image size was kept at 1920×1080 pixels to obtain the denoised fetal heart monitoring image. The local histogram equalization was used The technology enhances the contrast of the fetal heart monitoring image after denoising. The image is divided into local blocks of 32×32 pixels. The grayscale histogram is calculated for each local block. The grayscale value distribution range is adjusted according to the histogram equalization principle, and the grayscale value is mapped to between 0 and 255. The processed local blocks are processed by the edge stitching algorithm to ensure the smooth transition of the overall image, and the fetal heart monitoring image after local enhancement is generated. Pixel intensity normalization processing is applied to the local enhanced monitoring image, and the pixel grayscale value range is calculated. The grayscale value of each pixel is subtracted from the minimum value and divided by the grayscale value range to achieve the grayscale value mapping to the standardized range of 0 to 1. Subsequently, the geometric space of the image is normalized to ensure that the pixel arrangement of the image conforms to the standard grid ratio. Finally, a fetal heart monitoring image with consistent resolution and standardized contrast is generated as input for subsequent processing.

[0106] Preferably, step S2 comprises the following steps:

[0107] Step S21: performing bidirectional recursive convolution on the standard monitoring image to generate image convolution features;

[0108] Step S22: performing dimensionality reduction compression processing on the image convolution features to obtain compressed feature data;

[0109] Step S23: performing regularization optimization on the compressed feature data to generate optimized feature data;

[0110] Step S24: performing normalization conversion on the optimized characteristic data to obtain normalized fetal heart characteristics.

[0111] In this embodiment, a bidirectional recursive convolutional network model is constructed by inputting a standard monitoring image. The model design includes two symmetrical convolution paths. The forward path includes three layers of convolution operations and a maximum pooling layer. The convolution kernel size is set to 3×3, the step size is 1, the number of output channels of the convolution layer is 64, 128 and 256 respectively, the window size of the maximum pooling layer is 2×2, and the step size is 2. The reverse path adopts a deconvolution structure, and the original resolution of the feature map is restored by upsampling operation. The upsampling adopts a bilinear interpolation method. The convolution kernel and step size parameters are consistent with the forward path. The recursive structure is set to three loop iterations. The recursive part performs a nonlinear transformation on the basis of the feature map generated by the forward path, and a jump connection is added to retain multi-scale information. The convolution operation adopts a ReLU activation function, and the output feature is saved in the form of a tensor (Tensor, tensor) with a size of 1080×1920×512. The convolution feature is input into the dimensionality reduction module, and the principal component analysis (Principal The principal component analysis (Principal Component Analysis) algorithm is used to reduce the dimensionality of high-dimensional features. First, the covariance matrix of the convolution feature is calculated, with a dimension of 512×512. The eigenvectors of the first 100 main components are selected according to the size of the eigenvalue to construct a dimensionality reduction transformation matrix. The convolution feature is subjected to matrix multiplication to complete the dimensionality reduction. The output compressed feature data has a dimension of 1080×1920×100. During the dimensionality reduction process, the cumulative variance contribution rate is monitored in real time to ensure that the amount of retained information is greater than 90%. After the processing is completed, the compressed features are stored as floating-point matrix files for subsequent optimization operations. Regularization optimization operations are performed on the compressed feature data. The sparsity and expression ability of the features are enhanced through the Laplace regularization method. First, a graph Laplacian matrix is ​​constructed for the compressed feature data. Matrix, graph Laplacian matrix), the graph Laplacian matrix obtains the weight matrix by calculating the Euclidean distance between sample points in the feature space, and adds a regularization factor on the diagonal line. The factor value is set to 0.01, and the weight distribution of the feature data is updated by iterative optimization. The optimization process adopts the gradient descent method (Gradient Descent), the initial learning rate is set to 0.001, the number of iterations is 200, and the dimension of the optimized feature data remains unchanged. The optimized feature data is normalized and transformed. First, the minimum-maximum normalization (Min-MaxNormalization) is performed to map the feature data to the range of [0,1]. After normalization, the feature vector is normalized by L2 norm so that the sum of the squares of each feature vector is equal to 1. The normalization transformation is completed through matrix operation. The size of the converted feature data is 1080×1920×100, ensuring that the data range and numerical distribution meet the input requirements of the subsequent classification model.

[0112] Preferably, step S21 includes the following steps:

[0113] Step S211: performing forward recursive convolution on the standard monitoring image to obtain weighted forward data; performing reverse recursive convolution on the standard monitoring image to generate modulated reverse data;

[0114] Step S212: performing feature alignment processing on the weighted forward data and the modulated reverse data to generate feature alignment data;

[0115] Step S213: performing cross-time step correlation on the feature alignment data to obtain correlated feature data; performing time series dependency modeling on the correlated feature data to generate a correlated dependency model;

[0116] Step S214: performing a deformable convolution transformation based on the association dependency model to obtain deformed feature data; performing a dilated convolution processing on the deformed feature data to generate dilated feature data;

[0117] Step S215: performing group convolution on the expanded feature data to obtain group feature data; performing separable convolution on the group feature data to generate image separation features;

[0118] Step S216: performing residual connection enhancement on the image separation feature to generate an image residual feature; and performing feature recalibration on the image separation feature based on the image residual feature to obtain an image convolution feature.

[0119] In this embodiment, the standard monitoring image is input into the bidirectional recursive convolution module. The forward recursive convolution path includes three layers of convolution operations. The size of each convolution kernel is set to 3×3, the step size is 1, and the padding method is to keep the input and output sizes consistent. The number of channels increases layer by layer to 32, 64, and 128. The activation function uses ReLU, and the number of forward recursions is set to 3. The context information of the original input image is retained in the output of each round through loop iteration to enhance the feature extraction ability of the convolution layer. The reverse recursive convolution path adopts a structure symmetrical to the forward path, and the spatial resolution is restored layer by layer through upsampling operations. The upsampling method is bilinear interpolation. The output modulated reverse data is consistent with the input image size and has a dimension of 128 channels. The weighted forward data and modulated reverse data finally obtained are The forward and reverse data are stored as independent multi-dimensional tensors (Tensor, tensor), and the feature alignment module is used to perform pixel-level feature alignment on the weighted forward data and the modulated reverse data. The feature alignment method uses bilinear interpolation combined with a cross-scale attention mechanism. First, the two sets of data are positionally encoded. The position encoding is completed by superimposing spatial coordinate information on each feature map channel. Then, the similarity matrix is ​​calculated using the point-to-point Euclidean distance, and the eigenvalues ​​in the modulated reverse data are mapped to the coordinate system of the weighted forward data through linear interpolation. The adjusted data is fused through the attention weight matrix. During the fusion process, the channel attention mechanism is added to further optimize the weight distribution. The size of the feature alignment data finally generated is consistent with the input image and contains 512 Channel, send the feature alignment data to the cross-time step association module, first slice the input data according to the time step, the slice size of each time step is 16 frames, and the inter-frame step is 4 frames. After slicing, the sequence features are processed using the long short-term memory network (LSTM, long short-term memory network). The number of hidden units of the LSTM network is set to 128. The output time series features are further mapped to a fixed-dimensional feature space through a fully connected layer. The temporal dependency modeling is completed by constructing a temporal dependency network based on Transformer. The dependency modeling network includes a multi-head attention layer and a feedforward network. The dimension of each attention head is 64, and a total of 8 attention heads are included. The final generated association dependency model is saved in matrix form for In the subsequent operation, the associated dependency model and feature alignment data are input into the deformable convolution module. The deformable convolution operation spatially transforms the feature map by dynamically generating an offset. The convolution kernel size is set to 3×3. The offset calculation is generated by a lightweight convolution network. The network structure contains two layers of 1×1 convolution operations and a ReLU activation layer. The deformed feature data size output by the deformable convolution is consistent with the input image. The hole convolution expansion processing is achieved by inserting a fixed number of holes in the convolution operation. The expansion rate is set to 2 and 4, and the convolution kernel size is 3×3. The expanded feature data generated after processing contains 768 channels. The expanded feature data is input into the group convolution module. The group convolution divides the channels of the feature map into 8 groups, with 96 channels in each group.The standard convolution operation is performed independently in the group, with a convolution kernel size of 5×5 and a stride of 1. The output of the group convolution is directly input into the separable convolution module. The separable convolution consists of depth convolution and point convolution. The depth convolution processes the data of each channel in the spatial dimension. The point convolution fuses the multi-channel data through a 1×1 convolution operation. The image separation feature generated finally contains 256 channels. The image separation feature and the feature data of the previous layer are residually connected. The connection process is realized by simple element-by-element addition. The output after the residual connection is input into the feature recalibration module. The feature recalibration adopts the self-attention mechanism. The importance of the feature map is redistributed by calculating the channel weight and the spatial weight. The size of the recalibrated feature map remains unchanged. The image convolution feature generated finally contains high-resolution context information and multi-scale features. ,

[0120] Preferably, step S211 includes the following steps:

[0121] Perform basic frame segmentation on the standard monitoring image to obtain time-series image frames;

[0122] Perform forward recursive convolution on the time-series image frames to generate forward feature map data;

[0123] Perform weighted superposition processing on the forward feature map data to obtain weighted forward data;

[0124] Perform mirror flip simulation on the standard monitoring image to obtain a mirror monitoring image;

[0125] Perform time-reversal evolution on standard monitoring images to generate time-reversed data;

[0126] Performing reverse recursive convolution on the mirror monitoring image based on the time-reversed data to obtain reverse feature mapping data;

[0127] The inverse feature map data is modulated and transformed to generate modulated inverse data.

[0128] In this embodiment, the frame segmentation algorithm is used to segment the standard monitoring image. The frame segmentation tool adopts a frame extraction module based on a fixed time interval. After the standard monitoring image is input, the continuous image sequence is evenly sliced ​​according to the set time interval parameters. Each slice contains a static image frame. The time interval is set to 1 second. The time sequence image frames obtained after segmentation are arranged in chronological order. The image resolution of each frame is 640×480 pixels. The time sequence image frames are input into the forward recursive convolution module for frame-by-frame processing. The recursive convolution module includes three continuous convolution layers. The convolution kernel size of each layer is set to 3×3, the convolution step size is 1, and the filling method is full zero filling to keep the feature map size unchanged. The number of recursive operations is set to 2, and the convolution output is accumulated to 1 in each iteration. The output of the previous round increases the feature expression ability through recursion. The convolution activation function uses ReLU (rectified linear unit). The forward feature map data finally generated is a multi-channel tensor. The feature data of each channel retains the local spatial features and time step related information. The weighted fusion algorithm is used to superimpose the channel weights of the forward feature map data. First, the feature response value of each channel is calculated. The calculation method is to globally pool the feature map through the average pooling layer. The pooling result is used as the basis for weight generation. The weight data is processed by the normalization function so that its value range is 0 to 1. The generated weight is multiplied element by element with the corresponding channel of the forward feature map data, and then the weighted results of all channels are summed pixel by pixel to finally generate the weighted forward data. , the standard monitoring image is mirrored by spatial flipping operation. The mirror flipping method is horizontally symmetrical. The pixel values ​​of the left half of the image are mapped to the right half position, and the pixel values ​​of the right half are mapped to the left half position. The flipped mirror monitoring image maintains the original resolution. The mirror data is used to enhance the symmetry and diversity of feature learning. The standard monitoring image is reversed frame by frame by calling the time series flipping function. The original image sequence is rearranged from back to front in time order. The rearranged image sequence is output as time reverse data. The number of frames of the time reverse data is consistent with the standard monitoring image sequence, and the time interval between frames remains unchanged. The storage format of the time reverse data is the same as the time series image frame. The reverse recursion is used. The reverse convolution module processes the mirror monitoring image. The structure of the reverse recursive convolution operation is the same as the forward recursive convolution, but the recursive direction is to iterate from the last frame to the first frame in reverse. The reverse convolution kernel size is also 3×3, the step size is 1, the activation function is ReLU, and the number of recursions is set to 2. Finally, the reverse feature map data is output, and the reverse feature map data is input into the modulation transformation module for channel feature enhancement. The modulation transformation module adjusts the channel weight distribution by calculating the correlation matrix between channels. First, the dot product similarity between each channel and the remaining channels is calculated, and then the channel attention weight is generated through the normalization operation. The generated attention weight is multiplied by the reverse feature map data channel by channel, and the output result is stored as modulated reverse data. The data structure is a multidimensional tensor.Contains 128 channels, and the resolution remains consistent with the input image.

[0129] Preferably, step S3 comprises the following steps:

[0130] Step S31: performing peak detection on the standard fetal heart characteristics to obtain a peak point set; performing valley value recognition on the peak point set to generate a valley point set;

[0131] Step S32: reconstructing the fluctuation contour based on the peak point set and the trough point set to generate a fetal heart dynamic model;

[0132] Step S33: Perform historical baseline statistics on the preset historical fetal heart monitoring data to obtain a historical fetal heart dynamic baseline;

[0133] Step S34: performing baseline feature matching on the fetal heart rate dynamics model according to the historical fetal heart rate dynamics baseline to obtain a fused feature baseline.

[0134] In this embodiment, the peak point identification is performed on the fetal heart characteristic signal after normalization by using a peak detection method based on a differential algorithm. First, a first-order differential operation is performed on the fetal heart signal to enhance the variation characteristics of the peak in the signal. The differential signal is searched for local maximum values ​​by a sliding window method. The window size is set to 50 sampling points. Each movement covers 75% of the window overlap area. The maximum value of the differential signal is detected in each window and its corresponding position is marked. The maximum value point positions in all windows are aggregated to form an initial peak point set. Then, the amplitude threshold method is used to eliminate pseudo peak points. The threshold is set to the 80% quantile of the overall signal amplitude. The peak points retained after elimination are expressed in time. The final peak point set is formed by arranging the peak points in sequence. Based on the detected peak point set, the trough point is identified by using the inter-peak area. First, the signal data interval between each pair of adjacent peak points is determined, and the lowest value and its position in each interval are taken as candidate trough points. Then, the candidate trough points are smoothed. The smoothing process is realized by using the Gaussian filtering method. The filter window size is set to 5 sampling points. The amplitude of the smoothed candidate trough points is screened. The screening condition is that the amplitude of the candidate trough point is lower than the 20% quantile of the signal amplitude. The screened trough point positions are arranged in chronological order to form the final trough point set. The peak point set and the trough point set are input into the contour reconstruction module. The sample The fluctuation contour of the fetal heart signal is reconstructed by an interpolation method. First, the initial contour curve of the signal is generated according to the positions of the peak point set and the trough point set. The interpolation point interval is set to 1 millisecond to ensure the smoothness and continuity of the curve. Then the initial contour curve is optimized twice. The optimization process is achieved by the least squares curve fitting method. The fitting goal is to reduce the error between the reconstructed curve and the original signal. Finally, the complete fluctuation contour of the fetal heart signal is obtained. The reconstructed contour data is stored in the form of a dynamic frame sequence as the input basis of the fetal heart dynamic model. Multiple groups of fetal heart dynamic signals are extracted from the historical monitoring data set. The data time span is nearly 6 months of monitoring records. First, each The historical fetal heart signals of the group are normalized. The normalization process includes amplitude normalization and time scale alignment. After normalization, the baseline features of all signals are extracted using statistical analysis tools. The baseline features include average amplitude, peak-to-valley distance and fluctuation frequency. The mean and standard deviation of all features are calculated, and the calculation results are stored as a baseline feature set. Then, the baseline feature set is used to generate a historical fetal heart dynamic baseline. The generation method is to fit the historical baseline features based on the multivariate normal distribution. The generated dynamic baseline is used for subsequent matching operations. The fetal heart dynamic model and the historical fetal heart dynamic baseline are input into the baseline feature matching module. The matching method is implemented using an algorithm based on dynamic time warping (DTW). First, the time axis of the fetal heart dynamic model and the historical baseline are dynamically aligned, and the minimum cumulative distance between the two is calculated. Then, the matching area is screened by the distance matching rule. The matching rule is that the cumulative distance is less than the set threshold, and the threshold is taken as 2 times the standard deviation of the historical baseline dynamic feature.The weighted fusion features of the fetal heart rate dynamic model and the historical baseline are calculated for the selected matching area. The fusion features are completed through the channel superposition method, and the fusion feature baseline is finally output.

[0135] Preferably, step S4 comprises the following steps:

[0136] Step S41: performing time series calibration mapping on the standard fetal heart characteristics and the fusion characteristic baseline to obtain calibration synchronization data;

[0137] Step S42: performing cycle feature recognition on the calibration synchronization data to generate periodic variation data;

[0138] Step S43: reconstructing the trend of the periodic variation data by quantitative analysis to obtain the variation characteristic curvature; deriving the steady-state threshold value based on the variation characteristic curvature to generate threshold prediction data;

[0139] Step S44: Perform dynamic floating calibration processing on the fused feature baseline based on the threshold prediction data to produce a baseline floating range.

[0140] In this embodiment, the two sets of data are normalized on the time axis. The normalization method uses piecewise linear interpolation to align the fetal heart characteristics and baseline data at the same time resolution. The aligned data are calibrated for time series characteristics. The calibration method is based on cross-correlation analysis. The time lag point of the characteristic signal is determined by calculating the cross-correlation function between the fetal heart characteristics and the fusion baseline. The fetal heart characteristics are time-shifted based on this, so that the two sets of signals are completely aligned on the time axis. Subsequently, the aligned data are locally synchronized using a sliding window method. The sliding window size is set to 2 seconds, and the window interval is 1 second. The synchronized data output after calibration are arranged in time order to form calibrated synchronized data. Fast Fourier Transform (FFT) is performed on the calibrated synchronized data to calculate the energy distribution of the data in different frequency bands. The frequency range is set to 0.1 Hz to 3 Hz, identify the energy peak position in the main frequency band, extract the corresponding frequency point as the initial reference point of the cyclic feature, then segment and calibrate the synchronous data through the sliding window, perform spectrum analysis on the local signal of each window, and perform statistical processing on the change trend of the frequency point, calculate the frequency offset and cycle length, and the output periodic variation data includes the frequency change and period fluctuation of each time window. Select the cubic polynomial as the fitting function, and perform fitting calculation on the amplitude change of the periodic variation data. The fitting result is used to generate the trend quantization curve, and the first-order derivative of the curve is used as the quantitative expression of the curvature of the changing characteristic. Then, the steady-state interval analysis of the curvature of the changing characteristic is performed. The analysis process is implemented based on the sliding mean method. The stable area is identified by calculating the local mean and variance of the curvature. The upper and lower bounds of the stable area are used as the initial threshold. The initial threshold is optimized twice in combination with the historical monitoring data. The optimization method is dynamic adjustment based on the mean shift model. Finally, the threshold prediction data is generated, and the baseline data and the threshold prediction data are synchronously aligned. The alignment method adopts dynamic time warping (DTW). Warping), after alignment, the peak points and valley points in the baseline data are expanded. The expansion range is based on the upper and lower bounds in the threshold prediction data. The weighted smoothing method is used to adjust the expansion result. The weighting factor is the square root of the standard deviation of the historical monitoring data. After adjustment, the expansion area of ​​the baseline data is dynamically marked, and the marked area is defined as the baseline floating range, which is output in the form of a time series.

[0141] Preferably, step S43 includes the following steps:

[0142] Rearranging the periodic variation data into a phase sequence to obtain rearranged sequence data; and performing curvature sampling marking based on the rearranged sequence data to obtain curvature marking data;

[0143] Perform high-dimensional mapping transformation on the curvature labeled data to obtain a high-dimensional mapping matrix;

[0144] Perform feature reconstruction on the high-dimensional mapping matrix to generate reconstructed feature data; perform curvature normalization on the reconstructed feature data to obtain the change feature curvature;

[0145] The extreme point of the changing characteristic curvature is located to generate extreme value sequence data; the extreme value sequence data is decomposed into stable intervals to generate a steady-state interval sequence;

[0146] Steady-state thresholds are derived based on steady-state interval sequences to generate threshold prediction data.

[0147] In this embodiment, the periodic variation data is input into the phase sequence rearrangement module, and the phase difference is calculated for each period segment in the input data. The phase difference is calculated by dividing the time interval between adjacent peak points and valley points by the complete period time. The calculation result is expressed as a standardized phase difference value. Then, all period segments are sorted from small to large according to the phase difference value, and the period segments are rearranged according to the sorting result. The output rearranged sequence data generates a new time series in chronological order. The local area of ​​the rearranged sequence data is segmented, and the length of each segment is set to 0.5 seconds. After segmentation, the curvature is calculated for the value change of each time period. The curvature calculation method is to calculate the first-order derivative and the second-order derivative through discrete points, and the curvature size is expressed by the absolute value. Then, the curvature of each time period is marked according to the change trend of the curvature size, and the area with obvious curvature change is marked as a feature point. The generated curvature marking data includes the time period number, the marking type and the curvature value. The curvature marking data is input into the high-dimensional mapping module. The mapping operation is based on multidimensional scaling analysis (MDS, Multidimensional The Scaling algorithm is implemented, which takes the time series characteristics, label type and curvature value of the labeled data as input dimensions, calculates the optimal embedding representation of the data in three-dimensional space, and uses the Euclidean distance as the similarity metric in the embedding calculation process. The converted high-dimensional mapping matrix contains three coordinate axis data, which represent time, curvature intensity and label classification respectively. The high-dimensional mapping matrix is ​​input into the feature reconstruction module, and the principal component analysis (PCA) is used to analyze the optimal embedding representation of the data in three dimensions. The main characteristic components of high-dimensional data are extracted by the method of MATLAB. The number of principal components is set to two during the extraction process. The data vectors corresponding to the two principal components are reconstructed into a two-dimensional characteristic data table. Then, the relationship between the columns of the two-dimensional characteristic data table is subjected to multivariate regression analysis to generate reconstructed characteristic data containing the weight relationship between the principal components. The reconstructed characteristic data are imported into the curvature normalization processing module. The maximum and minimum value normalization method is used to limit the curvature value range in the characteristic data to between 0 and 1. The minimum and maximum values ​​of the data are calculated during the normalization process, and the curvature value is scaled proportionally. The change characteristic curvature after normalization is expressed as a standardized value. The change characteristic curvature is input into the extreme point positioning module. The sliding window method is used to detect local extreme values ​​of the curvature data. The sliding window size is set to 20 data points. During the detection process, the maximum and minimum values ​​in each window are compared, and these extreme values ​​are recorded. The generated extreme value sequence data contains timestamps and extreme value sizes. The extreme value sequence data is input into the stable interval decomposition module. First, the difference between adjacent extreme value points is calculated, and the extreme value points with a difference less than the preset threshold are divided into the same interval. The preset threshold is set to 0.1. After the partitioning is completed, the time range and extreme value range of each stable interval are statistically calculated. The output steady-state interval sequence contains the interval number, time range and extreme value fluctuation range. The steady-state interval sequence is input into the steady-state threshold derivation module, and the mean and standard deviation of the extreme values ​​in the interval are used to calculate the steady-state threshold. The upper threshold is the mean plus two times the standard deviation, and the lower threshold is the mean minus two times the standard deviation. After the calculation is completed, the upper and lower limits of all intervals are arranged in chronological order, and the generated threshold prediction data is expressed and output in the form of a time series. .

[0148] Preferably, step S5 comprises the following steps:

[0149] Step S51: performing time-sharing window segmentation according to the baseline floating range to obtain time multi-scale features; performing multi-band filtering on the time multi-scale features to generate frequency multi-scale features;

[0150] Step S52: performing multi-amplitude interval statistics on the frequency multi-scale feature to obtain the amplitude multi-scale feature;

[0151] Step S53: setting decision thresholds for the time multi-scale features, the frequency multi-scale features and the amplitude multi-scale features based on the preset threshold interval to obtain an initial discrimination rule;

[0152] Step S54: Perform multi-level combination modeling on the initial discrimination rules to obtain a classification discrimination framework;

[0153] Step S55: Classify the fetal heart monitoring image based on the classification and discrimination framework to implement an intelligent classification method for fetal heart monitoring images based on a neural network.

[0154] In this embodiment, according to the data of the baseline floating range, time windows are divided on the original time series signal, the time length of each window is set to 5 seconds, and the sliding step of the window is 1 second. In the process of window division, the overlapping part of each window is adjusted according to the baseline floating range, and the overlapping ratio is set to 50%. Features are extracted in each window, and the calculated features include statistics such as mean value, standard deviation, kurtosis and skewness. These features are used to represent time multi-scale information. Finally, time multi-scale feature data including the time position and corresponding features of each window are output, and high-pass filtering and low-pass filtering are performed on the time multi-scale feature data. The filtering frequency band is set to 0.1 Hz to 30 Hz to remove high-frequency noise and low-frequency drift in the signal. After filtering, low-frequency and high-frequency partial signals are obtained, and the low-frequency signal is further band-pass filtered, and the filtering bandwidth is set to 0.5 Hz to 5 Hz, the filtered signal is regarded as a frequency multi-scale feature. The frequency multi-scale feature data contains different frequency band features from low frequency to high frequency. The amplitude interval is set as the mean range of the difference between the maximum and minimum amplitudes in each frequency band. The frequency multi-scale feature is divided into several amplitude intervals according to the amplitude interval. Each amplitude interval represents a signal strength range. By statistically analyzing the frequency multi-scale feature data in each amplitude interval, the mean, variance, peak value, and maximum amplitude in the interval are extracted to generate amplitude multi-scale feature data. The threshold range of each feature is set. The threshold range of the time multi-scale feature is set according to the average value and standard deviation in the window. The threshold range of the frequency multi-scale feature is set according to the difference between the high-frequency energy and the low-frequency energy of the signal. The threshold range of the amplitude multi-scale feature is set according to the statistical interval of the amplitude. Based on these threshold ranges, the preliminary judgment rules for each feature are obtained. Each rule contains specific features. The mapping relationship between the threshold range of the value and the corresponding category is calculated. The decision tree algorithm is used to hierarchically model the initial rules. Each layer of the model is trained by gradually selecting features. After constructing the feature combination, the data is recursively divided to obtain the classification label of each data segment. The multi-layer decision tree model is integrated by hierarchical combination to finally generate a classification discrimination framework. The classification discrimination framework includes multiple discrimination levels and weight coefficients corresponding to each level. The fetal heart monitoring image is normalized by the preprocessing module to ensure that the input size of the image is 256x256 pixels. Then the normalized image is input into the neural network model, and the image features are extracted by the convolution layer. Then, multi-level decisions are made according to the image features and the hierarchical rules in the classification discrimination framework, and the classification results are output. The classification results include the category label and classification confidence of each fetal heart monitoring image. The weight of the classification discrimination framework is adjusted through the training process of the neural network to improve the classification accuracy.

[0155] The present invention also provides a fetal heart monitoring image classification system based on a neural network, which is used to execute the fetal heart monitoring image classification method based on a neural network as described above. The fetal heart monitoring image classification system based on a neural network includes:

[0156] An image acquisition module is used to acquire fetal heart monitoring images; perform Gaussian noise elimination on the fetal heart monitoring images to obtain noise-reduced monitoring images; perform spatial domain normalization on the noise-reduced monitoring images to generate standard monitoring images;

[0157] The image convolution module is used to perform multi-channel convolution on the standard monitoring image to generate image convolution features; the image convolution features are compressed and regularized to optimize to obtain standardized fetal heart features;

[0158] The fusion baseline recognition module is used to perform morphological reconstruction of the standard fetal heart characteristics and generate a fetal heart dynamic model; perform baseline feature recognition on the fetal heart dynamic model based on the preset historical fetal heart monitoring data to obtain a fusion feature baseline;

[0159] The floating range derivation module is used to perform period variation quantification processing on the fusion feature baseline based on the standard fetal heart characteristics to obtain period variation data; the safe floating range of the fusion feature baseline is deduced according to the period variation data to generate the baseline floating range;

[0160] The discrimination framework building module is used to build a multi-scale discrimination framework for standard fetal heart characteristics according to the baseline floating range to obtain a classification discrimination framework; based on the classification discrimination framework, the fetal heart monitoring image is classified and processed to realize an intelligent classification method for fetal heart monitoring images based on a neural network.

[0161] The present invention improves the image quality through efficient image processing technology through the image acquisition module, laying a solid foundation for subsequent analysis, the noise reduction processing effectively removes the interference signal and enhances the clarity of the image, the spatial domain normalization standardizes the image data, making the images comparable, the image convolution module extracts rich feature information using multi-channel convolution technology, the compression regularization optimization reduces redundancy and improves the processing efficiency of features, the fusion baseline recognition module provides an in-depth understanding of the dynamic model through morphological reconstruction, the baseline feature recognition based on historical data improves the accuracy and consistency of feature recognition, the floating range derivation module enhances the sensitivity to data changes through periodic variation quantitative analysis, and the generated safe floating range provides a clear reference for subsequent discrimination, the discrimination framework building module improves the flexibility and adaptability of the model through multi-scale analysis, and finally realizes the efficient classification of fetal heart monitoring images, promotes the technological progress and expansion of the application scope of the intelligent monitoring system, and improves the performance and reliability of the system as a whole.

[0162] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0163] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A fetal heart monitoring image classification method based on neural network, characterized in that: The following steps are involved: Step S1: Acquire a fetal heart monitoring image; perform Gaussian noise elimination on the fetal heart monitoring image to obtain a noise-reduced monitoring image; perform spatial domain normalization on the noise-reduced monitoring image to generate a standard monitoring image; Step S2: performing multi-channel convolution on the standard monitoring image to generate image convolution features; performing compression regularization optimization on the image convolution features to obtain standardized fetal heart features; Step S3: morphologically reconstructing the standardized fetal heart characteristics to generate a fetal heart dynamic model; Perform baseline feature recognition on the fetal heart dynamic model based on preset historical fetal heart monitoring data to obtain a fusion feature baseline; Step S4: performing period variation quantification processing on the fusion feature baseline based on the standard fetal heart characteristics to obtain period variation data; The safe floating range of the fusion feature baseline is derived based on the periodic variation data to generate the baseline floating range; Step S5: constructing a multi-scale discrimination framework for the standard fetal heart characteristics according to the baseline floating range to obtain a classification discrimination framework; Fetal heart rate monitoring images are classified and processed based on a classification and discrimination framework to realize an intelligent classification method for fetal heart rate monitoring images based on a neural network.

2. The neural network-based fetal heart rate monitoring image classification method according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: Acquire a fetal heart monitoring image; perform multi-resolution decomposition on the fetal heart monitoring image to obtain multi-layer decomposition data; Step S12: performing Gaussian kernel optimization processing on the multi-layer decomposition data to generate kernel function parameters; Step S13: reconstructing the kernel function parameters by noise suppression to obtain a noise-reduced monitoring image; Step S14: performing local contrast stretching on the noise reduction monitoring image to generate a local enhanced monitoring image; Step S15: performing spatial domain normalization processing on the local enhanced monitoring image to generate a standard monitoring image.

3. The neural network-based fetal heart monitoring image classification method according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: performing bidirectional recursive convolution on the standard monitoring image to generate image convolution features; Step S22: performing dimensionality reduction compression processing on the image convolution features to obtain compressed feature data; Step S23: performing regularization optimization on the compressed feature data to generate optimized feature data; Step S24: performing normalization conversion on the optimized characteristic data to obtain normalized fetal heart characteristics.

4. The neural network-based fetal heart monitoring image classification method according to claim 3, characterized in that: Step S21 includes the following steps: Step S211: performing forward recursive convolution on the standard monitoring image to obtain weighted forward data; performing reverse recursive convolution on the standard monitoring image to generate modulated reverse data; Step S212: performing feature alignment processing on the weighted forward data and the modulated reverse data to generate feature alignment data; Step S213: performing cross-time step correlation on the feature alignment data to obtain correlated feature data; performing time series dependency modeling on the correlated feature data to generate a correlated dependency model; Step S214: performing a deformable convolution transformation based on the association dependency model to obtain deformed feature data; performing a dilated convolution processing on the deformed feature data to generate dilated feature data; Step S215: performing group convolution on the expanded feature data to obtain group feature data; performing separable convolution on the group feature data to generate image separation features; Step S216: performing residual connection enhancement on the image separation feature to generate an image residual feature; and performing feature recalibration on the image separation feature based on the image residual feature to obtain an image convolution feature.

5. The neural network-based fetal heart monitoring image classification method according to claim 4, characterized in that: Step S211 includes the following steps: Perform basic frame segmentation on the standard monitoring image to obtain time-series image frames; Perform forward recursive convolution on the time-series image frames to generate forward feature map data; Perform weighted superposition processing on the forward feature map data to obtain weighted forward data; Perform mirror flip simulation on the standard monitoring image to obtain a mirror monitoring image; Perform time-reversal evolution on standard monitoring images to generate time-reversed data; Performing reverse recursive convolution on the mirror monitoring image based on the time-reversed data to obtain reverse feature mapping data; The inverse feature map data is modulated and transformed to generate modulated inverse data.

6. The neural network-based fetal heart monitoring image classification method according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: performing peak detection on the standard fetal heart characteristics to obtain a peak point set; performing valley value recognition on the peak point set to generate a valley point set; Step S32: reconstructing the fluctuation contour based on the peak point set and the trough point set to generate a fetal heart dynamic model; Step S33: Perform historical baseline statistics on the preset historical fetal heart monitoring data to obtain a historical fetal heart dynamic baseline; Step S34: performing baseline feature matching on the fetal heart rate dynamics model according to the historical fetal heart rate dynamics baseline to obtain a fused feature baseline.

7. The neural network-based fetal heart monitoring image classification method according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: performing time series calibration mapping on the standard fetal heart characteristics and the fusion characteristic baseline to obtain calibration synchronization data; Step S42: performing cycle feature recognition on the calibration synchronization data to generate periodic variation data; Step S43: reconstructing the trend of the periodic variation data by quantitative analysis to obtain the variation characteristic curvature; deriving the steady-state threshold value based on the variation characteristic curvature to generate threshold prediction data; Step S44: Perform dynamic floating calibration processing on the fused feature baseline based on the threshold prediction data to produce a baseline floating range.

8. The neural network-based fetal heart monitoring image classification method according to claim 7, characterized in that: Step S43 includes the following steps: Performing phase sequence rearrangement on the periodic variation data to obtain rearranged sequence data; performing curvature sampling marking based on the rearranged sequence data to obtain curvature marking data; Perform high-dimensional mapping transformation on the curvature labeled data to obtain a high-dimensional mapping matrix; Perform feature reconstruction on the high-dimensional mapping matrix to generate reconstructed feature data; perform curvature normalization on the reconstructed feature data to obtain the change feature curvature; The extreme point of the changing characteristic curvature is located to generate extreme value sequence data; the extreme value sequence data is decomposed into stable intervals to generate a steady-state interval sequence; Steady-state thresholds are derived based on steady-state interval sequences to generate threshold prediction data.

9. The neural network-based fetal heart monitoring image classification method according to claim 1, characterized in that: Step S5 includes the following steps: Step S51: performing time-sharing window segmentation according to the baseline floating range to obtain time multi-scale features; performing multi-band filtering on the time multi-scale features to generate frequency multi-scale features; Step S52: performing multi-amplitude interval statistics on the frequency multi-scale feature to obtain the amplitude multi-scale feature; Step S53: setting decision thresholds for the time multi-scale features, the frequency multi-scale features and the amplitude multi-scale features based on the preset threshold interval to obtain an initial discrimination rule; Step S54: Perform multi-level combination modeling on the initial discrimination rules to obtain a classification discrimination framework; Step S55: Classify the fetal heart monitoring image based on the classification and discrimination framework to implement an intelligent classification method for fetal heart monitoring images based on a neural network.

10. A fetal heart monitoring image classification system based on neural network, characterized in that: For executing the neural network-based fetal heart monitoring image classification method as claimed in claim 1, the neural network-based fetal heart monitoring image classification system comprises: An image acquisition module is used to acquire fetal heart monitoring images; perform Gaussian noise elimination on the fetal heart monitoring images to obtain noise-reduced monitoring images; perform spatial domain normalization on the noise-reduced monitoring images to generate standard monitoring images; The image convolution module is used to perform multi-channel convolution on the standard monitoring image to generate image convolution features; the image convolution features are compressed and regularized to optimize to obtain standardized fetal heart features; The fusion baseline recognition module is used to perform morphological reconstruction of the standard fetal heart characteristics and generate a fetal heart dynamic model; perform baseline feature recognition on the fetal heart dynamic model based on the preset historical fetal heart monitoring data to obtain a fusion feature baseline; The floating range derivation module is used to perform period variation quantification processing on the fusion feature baseline based on the standard fetal heart characteristics to obtain period variation data; the safe floating range of the fusion feature baseline is deduced according to the period variation data to generate the baseline floating range; The discrimination framework building module is used to build a multi-scale discrimination framework for standard fetal heart characteristics according to the baseline floating range to obtain a classification discrimination framework; based on the classification discrimination framework, the fetal heart monitoring image is classified and processed to realize an intelligent classification method for fetal heart monitoring images based on a neural network.

Images