Fetal heart rate pathology signal generation method and system based on progressive growth architecture

Abstract

The invention discloses a fetal heart rate pathology signal generation method and system based on a progressive growth architecture. The method comprises the following steps: acquiring a real fetal heart rate signal, and preprocessing the real fetal heart rate signal to form a data set; a progressive generative adversarial network model is constructed, and training is carried out by using the data set and adopting a progressive growth strategy; and inputting a random noise vector into the trained generator, and synthesizing a predicted fetal heart rate signal sample. According to the method, a complex long-time sequence generation task is decomposed into a plurality of simple sub-tasks through a progressive growth training strategy, the characteristics of signals from a coarse-grained structure to fine-grained details are learned from shallow to deep, and long-sequence, rich-detail and high-fidelity FHR signals are effectively generated.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence and biological signal processing technology, and relates to a method and system for generating fetal heart rate pathological signals based on a progressive growth architecture. It solves the problem of imbalance in fetal heart rate data categories and is suitable for synthesizing high-fidelity and diverse minority pathological fetal heart rate samples. Background Technology

[0002] Fetal heart rate monitoring is a core tool for assessing fetal health during prenatal and intrapartum periods. By analyzing changes in fetal heart rate (FHR) signals, dangerous conditions such as fetal distress can be detected early. In recent years, deep learning-based methods, such as convolutional neural networks and recurrent neural networks, have shown great potential in FHR signal classification tasks. However, these data-driven methods heavily rely on large amounts of high-quality, class-balanced labeled data for training.

[0003] In clinical practice, the vast majority of fetuses are healthy, while those with pathological conditions are a minority. This results in a significantly larger number of normal samples than pathological samples in the collected FHR dataset, leading to a severe class imbalance. Directly using such an imbalanced dataset to train a classification model will cause the model to be severely biased towards the majority class (normal samples), while its recognition rate for the minority class (pathological samples) will be extremely low, i.e., insufficient sensitivity. This is unacceptable in clinical practice because the consequences of missing pathological cases are very serious.

[0004] To address the data imbalance problem, traditional methods are mainly divided into two categories: oversampling and feature transformation. Oversampling methods, such as SMOTE and its derivatives, synthesize new samples by performing linear interpolation between the k nearest neighbors of minority class samples. However, when applied to high-dimensional, long-range dependent FHR time series, these methods are prone to generating clinically meaningless, overgeneralized samples, leading to blurred class boundaries and sample overlap. Feature transformation methods, such as converting a one-dimensional signal into a two-dimensional image, while increasing the data dimensionality, often ignore the temporal global features of the signal itself, resulting in poor model adaptability.

[0005] Generative Adversarial Networks (GANs), as a powerful deep generative model, offer another solution. Through adversarial games between the generator and the discriminator, they can learn the high-dimensional distribution of real data and generate new samples similar to it. However, traditional GANs and their variants (such as DCGAN and WGAN) face many challenges when generating long-term sequences: (1) Training instability: The training process of GANs is difficult to converge and is prone to mode collapse, that is, the generator can only produce a limited number or even one type of sample, lacking diversity. (2) Gradient vanishing / exploding: For long sequences, the gradient is prone to decay or surge during backpropagation, making it difficult for the model to learn effective long-term dependencies. (3) Contradiction between generation quality and length: Directly training the model to generate long-term sequences often results in blurred signal structure, missing details, and low fidelity.

[0006] Therefore, there is an urgent need in this field for a new method that can stably and efficiently generate high-quality, long-term FHR pathological samples, in order to fundamentally improve the training data environment for fetal health auxiliary diagnostic models and enhance the reliability and accuracy of their clinical applications. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for generating fetal heart rate pathological samples based on progressive generative adversarial networks. This method can effectively generate long-term sequence samples that are highly consistent with real pathological FHR signals in morphology and statistical distribution, thereby solving the class imbalance problem in fetal health classification and significantly improving the diagnostic performance of downstream classification models.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides a method for generating fetal heart rate pathological samples based on a progressive generative adversarial network, comprising the following steps:

[0010] Step S1: Obtain the actual fetal heart rate signal and preprocess it to form a dataset;

[0011] Step S2: Construct a progressive generative adversarial network model and train it using a progressive growth strategy with the dataset;

[0012] A trained progressive generative adversarial network consists of a generator G and a discriminator D.

[0013] The generator G employs a progressive growth structure, with its input being a normally distributed random noise vector z. The generator G includes a fully connected layer, a reshape layer, two cascaded first convolutional blocks, multiple stacked progressive growth blocks, and a convolutional output layer. The fully connected layer and the reshape layer map the input noise into a low-resolution, fixed-length initial sequence. The two cascaded first convolutional blocks perform convolution processing on the initial sequence. The multiple stacked progressive growth blocks perform successive learning and upsampling processing on the output after convolution, moving from coarse-grained structure to fine-grained details. The convolutional output layer generates a predicted fetal heart rate signal from the output of the final progressive growth block.

[0014] The discriminator D takes as input the real fetal heart rate signal samples from the dataset and the predicted fetal heart rate signal samples generated by the generator G, and outputs the discrimination result.

[0015] Step S3: By inputting a random noise vector into the trained generator, a sample of the predicted fetal heart rate signal is synthesized.

[0016] Preferably, each progressive growth block has the same structure, including a one-dimensional upsampling layer and two cascaded first convolutional blocks; wherein the one-dimensional upsampling layer is used to double the length of the input sequence; the first convolutional block includes a one-dimensional convolutional layer, a LeakyReLU activation function, and a normalization layer.

[0017] Preferably, the discriminator D includes a second convolutional block, multiple stacked progressive discriminant blocks, a mini-batch standard deviation layer, two cascaded second convolutional blocks, a flattening layer, and a fully connected layer. The second convolutional block is responsible for convolutional processing of the input real fetal heart rate signal samples and the predicted fetal heart rate signal samples. The multiple stacked progressive discriminant blocks are responsible for performing multi-level feature extraction and progressive downsampling on the output of the second convolutional blocks, from fine-grained to coarse-grained. The mini-batch standard deviation layer is responsible for calculating the standard deviation of all samples in the current mini-batch on each feature dimension, forming a standard deviation vector, and concatenating it to the output feature map of the last progressive discriminant block. The two cascaded second convolutional blocks convolve the output of the mini-batch standard deviation layer. The flattening layer and the fully connected layer sequentially flatten and combine the outputs of the two second convolutional blocks to output the discrimination result.

[0018] More preferably, each progressive discriminant block has the same structure, which includes two cascaded second convolutional blocks and a downsampling layer; the second convolutional block includes a one-dimensional convolutional layer and a LeakyReLU activation function; the downsampling layer is used to halve the length of the input sequence.

[0019] Preferably, during training, the loss function of the progressive generative adversarial network is optimized by introducing an L2 regularization term to achieve a loss function based on Wasserstein distance and gradient penalty; this loss function is defined as:

[0020]

[0021] in Represents the true data distribution. This indicates the distribution of the generated data output by the generator. This indicates the output of the discriminator. The generator outputs a sample of the predicted fetal heart rate signal. This is a sample of a real fetal heart rate signal. These are random sampling points on the line connecting the actual fetal heart rate signal sample points and the predicted fetal heart rate signal sample points. The gradient penalty coefficient is... The L2 regularization coefficient is... Expressing expectations, Represents the 2-norm; This indicates the discriminator's expected value for classifying all predicted fetal heart rate signal samples as true estimates. This indicates the discriminator's expected value in classifying all real fetal heart rate signal samples as real. This represents the expectation of the squared difference between the 2-norm of the gradient and 1 for the discriminator at random sampling points along the line connecting the true and predicted fetal heart rate signal sample points. This represents the L2 regularized expectation of the discriminator's assessment of all real fetal heart rate signal samples as real.

[0022] Secondly, the present invention provides a fetal heart rate pathological sample generation system for implementing the above method, comprising:

[0023] The data acquisition module is responsible for generating random noise vectors;

[0024] The sample generation module is responsible for inputting random noise vectors into the trained generator to generate sample fetal heart rate prediction signals.

[0025] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the first aspect above.

[0026] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect above.

[0027] The beneficial effects of this invention are:

[0028] 1. High-quality long-term sequence generation: Through a progressive growth training strategy, the complex task of generating long-term sequences is decomposed into multiple simple sub-tasks. The features of the signal from coarse-grained structure to fine-grained details are learned from shallow to deep, effectively generating long sequences, rich in details, and high-fidelity FHR signals.

[0029] 2. Stable training and diverse samples: By combining mechanisms such as Wasserstein distance, gradient penalty, L2 regularization and mini-batch standard deviation, the training instability and pattern collapse problems of traditional GANs are effectively overcome, ensuring the diversity of generated samples.

[0030] 3. Universality and scalability: The progressive growth training framework proposed in this invention is not limited to FHR signals. Its core idea can be extended to other long-term sequence generation scenarios with class imbalance problems, such as the synthesis and data augmentation of physiological signals such as EEG and ECG. Attached Figure Description

[0031] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 The process for generating pathological signals of fetal heart rate using a progressive growth architecture.

[0033] Figure 2 A schematic diagram of the structure of a progressive generative adversarial network for training.

[0034] Figure 3 The diagram shows the structures of two types of convolutional blocks, where (a) is the first convolutional block and (b) is the second convolutional block.

[0035] Figure 4 This is a schematic diagram illustrating the evolution of the FHR signal generation process at different growth stages in a specific embodiment.

[0036] Figure 5 The figure shows a comparison of the PCA distribution of the FHR signal generated by the method of the present invention and the original FHR signal by the classical generative adversarial network. In the figure, (a) is the classical generative adversarial network DCGAN, (b) is the classical generative adversarial network WGAN, (c) is the classical generative adversarial network WGAN_gp, and (d) is the method of the present invention. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way.

[0038] Example 1: Implementation and Application of Fetal Heart Rate Pathological Data Generation Based on a Specific Progressive Growth Architecture

[0039] Step S1: Obtain the real fetal heart rate signal using the publicly available CTU-UHB dataset; perform denoising, normalization and slice processing on the real fetal heart rate signal to form a standardized training dataset, in which a few pathological samples are identified and stored separately.

[0040] Step S2: Construct a progressive generative adversarial network model and train it using a progressive growth strategy with the dataset.

[0041] See appendix Figure 2 The trained progressive generative adversarial network consists of a generator G and a discriminator D.

[0042] The generator G adopts a progressive growth structure, and its input is a random noise vector z that follows a normal distribution. The generator G includes a fully connected layer, a reshape layer, two cascaded first convolutional blocks, multiple stacked progressive growth blocks, and a convolutional output layer.

[0043] Specifically, the fully connected layer and Reshape are responsible for mapping the input noise into a low-resolution initial sequence of length m.

[0044] The two concatenated first convolutional blocks are responsible for convolutional processing of the initial sequence;

[0045] Multiple stacked progressively growing blocks are responsible for progressively learning and upsampling the output after convolution from coarse-grained structure to fine-grained details. Each progressively growing block has the same structure, including a one-dimensional upsampling layer and two cascaded first convolutional blocks. The one-dimensional upsampling layer is used to double the length of the input sequence, updating the length to... The first convolutional block consists of a one-dimensional convolutional layer, a LeakyReLU activation function, and a normalization layer; see appendix. Figure 3 (a)

[0046] The convolutional output layer uses the Tanh activation function to map the output of the final progressively growing block to the [-1,1] interval, generating the predicted FHR signal.

[0047] The discriminator D is mirror-symmetric to the generator G. Its inputs are real FHR samples and samples generated by the generator G, and its output is the discrimination result. The discriminator D includes a second convolutional block, multiple stacked progressive discriminant blocks, a mini-batch standard deviation layer, two cascaded second convolutional blocks, a flattening layer, and a fully connected layer.

[0048] Specifically, the second convolutional block is responsible for performing convolution processing on the input real fetal heart rate signal sample and the predicted fetal heart rate signal sample.

[0049] Multiple stacked progressive discriminant blocks are responsible for performing multi-level feature extraction and progressive downsampling from the output of the second convolutional block, ranging from fine-grained to coarse-grained. Each progressive discriminant block has the same structure, consisting of two cascaded second convolutional blocks and a downsampling layer. The second convolutional block includes a one-dimensional convolutional layer and a LeakyReLU activation function (see Appendix). Figure 3 (b) The downsampling layer is used to halve the length of the input sequence.

[0050] The mini-batch standard deviation layer is responsible for calculating the standard deviation of all samples in the current mini-batch on each feature dimension, forming a standard deviation vector, and concatenating it to the output feature map of the last progressive discriminant block to increase the discriminator's ability to detect mode collapse.

[0051] Two cascaded second convolutional blocks convolve the output of the mini-batch standard deviation layer; the flattening layer and the fully connected layer flatten and combine the outputs of the two second convolutional blocks in sequence to output the discrimination result.

[0052] Specifically, the loss function of the progressive generative adversarial network is optimized by introducing an L2 regularization term to obtain a loss function based on Wasserstein distance and gradient penalty. This loss function is defined as:

[0053]

[0054]

[0055]

[0056]

[0057]

[0058] in Represents the true data distribution. This indicates the distribution of the generated data output by the generator. This indicates the output of the discriminator. The generator outputs a sample of the predicted fetal heart rate signal. This is a sample of a real fetal heart rate signal. These are random sampling points on the line connecting the actual fetal heart rate signal sample points and the predicted fetal heart rate signal sample points. The gradient penalty coefficient is... The L2 regularization coefficient is... Expressing expectations, Represents the 2-norm; This indicates the discriminator's expected value for classifying all predicted fetal heart rate signal samples as true estimates. ; This represents the expected value of the discriminator in classifying all real fetal heart rate signal samples as real. This represents the expectation of the squared difference between the gradient 2-norm and 1 of the discriminator at random sampling points along the line connecting the true fetal heart rate signal sample points and the predicted fetal heart rate signal sample points; This represents the gradient 2 norm of the discriminator for random sampling points on the line connecting the real fetal heart rate signal sample points and the predicted fetal heart rate signal sample points; This represents the L2 regularized expectation of the discriminator's assessment of all real fetal heart rate signal samples as real.

[0059] This loss function effectively constrains the discriminator to satisfy the 1-Lipschitz condition, avoiding gradient anomalies. At the same time, by applying L2 regularization to the discriminant output based on real data, it further enhances the model's generalization ability and the quality of the generated signal.

[0060] See appendix Figure 1 The process of generating fetal heart rate pathological signals using a progressive growth architecture.

[0061] This embodiment uses the generation of a 9600-point pathological FHR signal as an example to illustrate the implementation of the present invention in detail.

[0062] 1. Model Building:

[0063] Generator: Input is a 100-dimensional noise vector. The fully connected layer outputs a 75×512 tensor, which is reshaped to 75×512. This is followed by two cascaded first convolutional blocks and seven stacked progressively growing blocks. Each progressively growing block has a kernel size of 3 and a stride of 1, with the number of channels decreasing sequentially in later growing blocks (512, 512, 256, 128, 64, 32, 16). Finally, a 1×1 convolution (i.e., the convolutional output layer) outputs a 9600×1 sequence. Discriminator: Input is the 9600×1 sequence generated by the generator and a 9600×1 sequence of the actual fetal heart rate signal. Its structure is strictly symmetrical with the generator, and includes a second convolutional block, 7 stacked progressive discriminant blocks, a mini-batch standard deviation layer, two cascaded second convolutional blocks, a flattening layer and a fully connected layer. The number of channels in the 7 progressive discriminant blocks increases (16, 32, 64, 128, 256, 512, 512), and the average pooling kernel size is 2.

[0064] 2. Training Details: Optimizer: Adam, learning rate 0.001, β1=0.0, β2=0.99. Hyperparameters: Gradient penalty coefficient λ=10, L2 regularization coefficient β=0.01. Batch size: 16. Training strategy: 7 growth phases (N=7). Each phase has a fade-in period and a stabilization period of 100 epochs. The discriminator is updated 5 times, and the generator is updated once.

[0065] like Figure 4 The diagram shows the progressive training process, starting with coarse-grained structural information (75 units in length), and training by doubling the length each time until finally generating fetal heart rate data with a length of 9600 units.

[0066] Step S3: Synthesize FHR samples by inputting a random noise vector z into the trained generator.

[0067] In summary, this invention successfully provides a stable and efficient scheme for generating fetal heart rate pathological samples, offering a powerful technical tool for solving the problem of imbalanced clinical data.

[0068] Figure 5 This image shows a comparison of the PCA distribution of FHR signals generated by the method of this invention and those generated by classical generative adversarial networks. It should be noted that, to cover more of the real distribution, each image consists of 500 generated data points and 90 training data points (red represents real training data, blue represents generated data). Data with similar features are clustered together after dimensionality reduction. Figure 5 (a) Figure 5 In (c), it was found that DCGAN, WAGN, and WGAN_gp generated datasets in the 40-minute data segment that were distributed in a portion of the real data, exhibiting a clear pattern collapse phenomenon. Figure 5 In (d), the distribution generated by the method of the present invention is roughly the same as the real distribution. There is a generated data distribution next to each real data, and there is no obvious pattern collapse in the generated data.

[0069] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating fetal heart rate pathological samples based on progressive generative adversarial networks, comprising the following steps: Acquire real fetal heart rate signals and preprocess them to form a dataset; Construct a progressive generative adversarial network model and train it using a progressive growth strategy on a dataset; By inputting a random noise vector into the trained generator, a sample of the predicted fetal heart rate signal is synthesized. The feature is that the trained progressive generative adversarial network includes a generator G and a discriminator D; the generator G adopts a progressive growth structure, and its input is a random noise vector z that follows a normal distribution; the generator G includes a fully connected layer, a reshape layer, two cascaded first convolutional blocks, multiple stacked progressive growth blocks, and a convolutional output layer; wherein, the fully connected layer and the reshape layer are responsible for mapping the input noise to a low-resolution fixed-length initial sequence; the two cascaded first convolutional blocks are responsible for convolutional processing of the initial sequence; the multiple stacked progressive growth blocks are responsible for successive learning and upsampling processing of the output after convolutional processing from coarse-grained structure to fine-grained details; the convolutional output layer is responsible for generating a predicted fetal heart rate signal from the output of the last progressive growth block; the input of the discriminator D is a sample of the real fetal heart rate signal in the dataset and a sample of the predicted fetal heart rate signal generated by the generator G, and the output is the discrimination result.

2. The method according to claim 1, characterized in that, Each progressive growth block has the same structure, including a one-dimensional upsampling layer and two cascaded first convolutional blocks; wherein the one-dimensional upsampling layer is used to double the length of the input sequence; the first convolutional block includes a one-dimensional convolutional layer, a LeakyReLU activation function, and a normalization layer.

3. The method according to claim 1, characterized in that, The discriminator D includes a second convolutional block, multiple stacked progressive discriminant blocks, a mini-batch standard deviation layer, two cascaded second convolutional blocks, a flattening layer, and a fully connected layer. The second convolutional block performs convolution processing on the input real fetal heart rate signal samples and the predicted fetal heart rate signal samples. The multiple stacked progressive discriminant blocks perform multi-level feature extraction and progressive downsampling on the output of the second convolutional blocks, from fine-grained to coarse-grained. The mini-batch standard deviation layer calculates the standard deviation of all samples within the current mini-batch on each feature dimension, forming a standard deviation vector, and concatenates it to the output feature map of the last progressive discriminant block. The two cascaded second convolutional blocks perform convolution processing on the output of the mini-batch standard deviation layer. The flattening layer and the fully connected layer sequentially flatten and combine the outputs of the two second convolutional blocks to output the discrimination result.

4. The method according to claim 3, characterized in that, Each progressive discriminant block has the same structure, which includes two cascaded second convolutional blocks and a downsampling layer; the second convolutional block includes a one-dimensional convolutional layer and a LeakyReLU activation function; the downsampling layer is used to halve the length of the input sequence.

5. The method according to claim 1, characterized in that, During training, the loss function of the progressive generative adversarial network is optimized by introducing an L2 regularization term to obtain a loss function based on Wasserstein distance and gradient penalty; this loss function is defined as: ； in Represents the true data distribution. This indicates the distribution of the generated data output by the generator. This indicates the output of the discriminator. The generator outputs a sample of the predicted fetal heart rate signal. This is a sample of a real fetal heart rate signal. These are random sampling points on the line connecting the actual fetal heart rate signal sample points and the predicted fetal heart rate signal sample points. The gradient penalty coefficient is... The L2 regularization coefficient is... Expressing expectations, Represents the 2-norm; This indicates the discriminator's expected value for classifying all predicted fetal heart rate signal samples as true estimates. This indicates the discriminator's expected value in classifying all real fetal heart rate signal samples as real. This represents the expectation of the squared difference between the 2-norm of the gradient and 1 for the discriminator at random sampling points along the line connecting the true and predicted fetal heart rate signal sample points. This represents the L2 regularized expectation of the discriminator's assessment of all real fetal heart rate signal samples as real.

6. A fetal heart rate pathological sample generation system for implementing the method of any one of claims 1-5, characterized in that, include: The data acquisition module is responsible for generating random noise vectors; The sample generation module is responsible for inputting random noise vectors into the trained generator to generate predicted fetal heart rate signal samples.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method of any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of any one of claims 1-5.

Images