Fetal distress auxiliary diagnosis system based on continuous wavelet transform and shallow convolutional neural network

Abstract

The invention discloses a fetal distress auxiliary diagnosis system based on continuous wavelet transform and a shallow convolutional neural network, and the system comprises a data obtaining and preprocessing module which is responsible for obtaining an original fetal heart rate signal and carrying out the data preprocessing of the original fetal heart rate signal; the time-frequency characteristic graph construction module is responsible for converting the signals processed by the data acquisition and preprocessing module from one-dimensional data into a two-dimensional time-frequency characteristic graph by using continuous wavelet transform; and the classification and recognition module is responsible for inputting the time-frequency characteristic pattern into a shallow convolutional neural network model and outputting a fetal distress state classification result. According to the method, a one-dimensional time sequence signal is converted into a two-dimensional time-frequency image, and joint distribution information of the signal in a time domain and a frequency domain is subjected to visual coding, so that a shallow convolutional neural network model can fully utilize the advantages of the shallow convolutional neural network model in image processing, deep pathological features hidden in an FHR signal are mined, and the deep pathological features hidden in the FHR signal are extracted. And the defect that the one-dimensional convolution is insensitive to the time sequence is overcome.

Description

A fetal distress auxiliary diagnostic system based on continuous wavelet transform and shallow convolutional neural network Technical Field

[0001] This invention relates to the fields of smart healthcare and biomedical signal processing technology, and in particular to a fetal distress auxiliary diagnostic system based on fetal heart rate signals. Specifically, it is a fetal distress auxiliary diagnostic system based on continuous wavelet transform and shallow convolutional neural network. Background Technology

[0002] Fetal heart rate monitoring is a core tool for assessing fetal health in utero. Clinically, doctors analyze fetal heart rate curves and uterine contraction pressures to determine the risk of fetal distress. However, the fetal heart rate (FHR) signal is jointly regulated by the fetal central nervous system, autonomic nervous system, and external environment, exhibiting complex morphology and high variability. Relying solely on visual interpretation by doctors is highly subjective and prone to misdiagnosis and missed diagnosis. Statistics show that the rate of unnecessary cesarean sections due to misinterpretation of cardiotocography (CTG) is increasing year by year, placing an unnecessary burden on both pregnant women and fetuses.

[0003] Currently, the interpretation of CTG in clinical practice heavily relies on obstetricians' visual analysis of the images. However, this traditional method has inherent limitations: it is highly subjective, the FHR signal morphology is complex and highly variable, and different doctors, or even the same doctor at different times, may interpret the same image differently, leading to inconsistent diagnostic results; it is also highly experience-dependent, with the accuracy of interpretation heavily reliant on the doctor's clinical experience, and young or inexperienced doctors are prone to misinterpretation.

[0004] To overcome the limitations of manual interpretation, computer-aided diagnostic systems have emerged. Early systems were mostly based on feature engineering, extracting one-dimensional features such as baseline rate, variability, acceleration, and deceleration from FHR and uterine contraction pressure signals according to clinical guidelines like FIGO, and then using traditional machine learning algorithms (such as support vector machines and decision trees) for classification. While this approach partially reduced subjectivity, it had significant drawbacks: incomplete feature extraction, with manually designed features failing to fully capture the complex nonlinear and time-varying characteristics of FHR signals, leading to the loss of crucial pathological information; poor model generalization ability, with strong correlations between extracted features and classifiers, resulting in unstable performance across different datasets and generally low specificity; and cumbersome and time-consuming processes, with feature extraction and selection heavily reliant on expert knowledge and exhibiting low computational efficiency.

[0005] In recent years, deep learning technology, especially convolutional neural networks (CNNs), has demonstrated powerful advantages in image recognition and medical signal analysis. CNNs can automatically learn hierarchical features from raw data end-to-end, avoiding the drawbacks of manual feature design. However, directly inputting a one-dimensional fetal distress signal into a CNN makes it difficult to effectively capture its frequency domain information and temporal dynamic characteristics. Therefore, how to combine the joint time-frequency features of the FHR signal with the image processing advantages of CNNs to construct an efficient, accurate, and lightweight model for assisting in the diagnosis of fetal distress has become a pressing technical problem in this field. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for assisting in the diagnosis of fetal distress based on continuous wavelet transform and convolutional neural networks. Its core objective is to achieve automated and high-precision identification of fetal distress by converting a one-dimensional FHR signal into a two-dimensional image rich in time-frequency information, and then using an optimized shallow CNN model for feature learning and classification, thereby providing reliable decision support for clinicians.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The present invention provides a fetal distress auxiliary diagnostic system based on continuous wavelet transform and shallow convolutional neural network, comprising: a data acquisition and preprocessing module, responsible for acquiring the raw fetal heart rate signal and performing data preprocessing on it; a time-frequency feature map construction module, responsible for using continuous wavelet transform to convert the signal processed by the data acquisition and preprocessing module from one-dimensional data into a two-dimensional time-frequency feature map; and a classification and recognition module, responsible for inputting the time-frequency feature map into the shallow convolutional neural network model and outputting the fetal distress state classification result.

[0008] Preferably, in the data acquisition and preprocessing module, the preprocessing includes filling in missing values, eliminating interference signals, and extracting continuous valid segments.

[0009] Preferably, the missing values ​​are filled using linear interpolation.

[0010] Preferably, the interference elimination signal is a signal that uses a medium-range filter to eliminate spike pulses caused by interference in the signal.

[0011] Preferably, the extraction of the continuous valid segment is the extraction of the last 10 minutes of the signal.

[0012] Preferably, the specific implementation process of the time-frequency feature map construction module is as follows: for each segment of fetal heart rate signal after data acquisition and preprocessing by the data acquisition and preprocessing module, a continuous wavelet transform is performed using the mother wavelet function at n scale factors. The Convolutional Transform (CWT) calculation generates 2n time-frequency feature maps, where n ≥ 2. Preferably, in the time-frequency feature map construction module, the mother wavelet function is selected as Morlet wavelet and Gaussian 8th order wavelet. Preferably, in the time-frequency feature map construction module, the n scale factors are 20, 25, and 30 to capture features of different frequency components. Preferably, the shallow convolutional neural network model includes an input layer, a convolutional layer, an activation layer, a normalization layer, a pooling layer, a fully connected layer, a Dropout layer, and a classification layer. The input layer receives the time-frequency feature map, and the convolutional layer uses a convolutional kernel to extract features from the time-frequency feature map of the input layer. Subsequently, the extracted features are nonlinearly mapped, normalized, and downsampled by the activation layer, normalization layer, and fully connected layer. Then, the extracted feature map is flattened and overfitted by the fully connected layer and Dropout layer. Finally, the classification layer outputs the classification result of the fetal state probability distribution.

[0013] The beneficial effects of this invention are at least as follows: strong feature expression capability: This invention uses CWT to convert one-dimensional time-series signals into two-dimensional time-frequency images, and visualizes and encodes the joint distribution information of the signal in the time domain and frequency domain, so that the shallow convolutional neural network model CNN can make full use of its advantages in image processing, and dig out the deep pathological features hidden in the FHR signal, overcoming the disadvantage of one-dimensional convolution being insensitive to time.

[0014] High diagnostic accuracy: The end-to-end CNN model (i.e., shallow convolutional neural network model) used in this invention can automatically learn the complex mapping relationship between time-frequency images and diagnostic results, avoiding information loss and bias caused by manual feature design and selection in traditional methods, thereby significantly improving the accuracy, sensitivity and specificity of classification.

[0015] The model combines efficiency and practicality: This invention abandons complex and deep network structures, designing a shallow CNN with only 8 layers. While ensuring high performance, this model has fewer parameters, lower computational complexity, and faster training and inference speeds. It does not require expensive high-performance hardware, making it more suitable for real-time or near-real-time diagnostic deployment in clinical settings. Attached Figure Description

[0016] 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.

[0017] Figure 1 is a schematic diagram of the fetal distress auxiliary diagnosis system module provided in an embodiment of the present invention.

[0018] Figure 2 is a schematic diagram of the data acquisition and preprocessing module in an embodiment of the present invention.

[0019] Figure 3 is a network structure diagram of a shallow convolutional neural network in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way.

[0021] This embodiment provides a fetal distress auxiliary diagnosis system based on continuous wavelet transform and shallow convolutional neural network. Referring to Figure 1, it includes a data acquisition and preprocessing module, which is responsible for acquiring the raw FHR signal from the publicly available fetal heart rate database CTU-UHB and performing data preprocessing on it.

[0022] In one implementation, the preprocessing includes filling in missing values, eliminating interference signals, and truncating continuous valid segments. For example, missing values ​​in the signal are filled using linear interpolation; a mid-range filter is applied to eliminate spikes in the signal caused by interference; and the last 10 minutes (corresponding to approximately 2400 data points) of the signal are truncated as continuous valid segments (i.e., stable segments) as the sample data to be analyzed.

[0023] The time-frequency feature map construction module is responsible for converting the signal processed by the data acquisition and preprocessing modules from one-dimensional data into a two-dimensional time-frequency feature map using continuous wavelet transform. In one implementation, the specific process of the time-frequency feature map construction module is as follows: for each segment of fetal heart rate signal processed by the data acquisition and preprocessing modules, CWT calculation is performed using the mother wavelet function at n scale factors, generating a total of 2n time-frequency feature maps; n≥2. The time-frequency feature map is an RGB three-channel image (i.e., 64×64×3 pixels) with a uniform size of 64×64 pixels, converted from the wavelet coefficient matrix to pixel values.

[0024] For example, the mother wavelet function is selected from Morlet wavelet and Gaussian 8th order wavelet; in this embodiment, three scaling factors are used, set to 20, 25, and 30 respectively, to capture the features of different frequency components; the horizontal axis of the time-frequency feature map represents time, and the vertical axis represents frequency. Therefore, each FHR signal can be converted into 6 time-frequency feature maps, as shown in Figure 2.

[0025] The classification and recognition module is responsible for inputting any time-frequency feature map obtained by the time-frequency feature map construction module into the shallow convolutional neural network model, and outputting the classification result of fetal distress state, thereby realizing the auxiliary diagnosis of fetal distress.

[0026] One implementation, as shown in Figure 3, is a shallow convolutional neural network model with 8 layers, including an input layer, a convolutional layer, an activation layer, a normalization layer, a pooling layer, a fully connected layer, a Dropout layer, and a classification layer.

[0027] The input layer receives a time-frequency feature map with a size of 64×64×3; the convolutional layer uses a convolutional kernel to extract features from the time-frequency feature map of the input layer; the activation layer uses the ReLU function to perform non-linear mapping on the above features; the normalization layer normalizes the activated features; the pooling layer selects max pooling to downsample the features; the fully connected layer flattens and synthesizes the extracted feature map; the Dropout layer sets the dropout probability to 0.5 to prevent overfitting; the classification layer uses the Softmax function to output the probability distribution of the fetal state as "normal" or "pathological".

[0028] During model training, the Adam optimizer was used with an initial learning rate of 0.001, a batch size of 32, and a training epoch of 100. The loss function was classification cross-entropy. The shallow convolutional neural network model was trained using the FHR sample set to optimize network weights and hyperparameters. The FHR sample set used in the training process consisted of all time-frequency feature maps obtained by the time-frequency feature map construction module.

[0029] The performance of the above methods is evaluated below. Table 1 shows some representative research in the field of fetal distress assisted diagnosis in recent years. The training model of this invention achieved a classification accuracy of 98.34% on the dataset. Existing technologies 1, 2, and 3 are traditional machine learning classification algorithms based on feature engineering, and their classification performance is weaker than that of deep learning-based classification algorithms. Compared with existing technology 4, the shallow convolutional neural network based on two-dimensional time-frequency feature maps constructed in this invention achieves higher classification accuracy than one-dimensional convolutional neural networks. Compared with existing technology 5, the time-frequency feature map based on continuous wavelet transform constructed in this invention has advantages over the time-frequency feature map based on discrete wavelet transform, and can better reflect the feature information of FHR signals.

[0030] Table 1. Comparison of diagnostic performance of different methods

[0031] The sources of each existing technology are as follows: Prior Art 1: A study based on gray level co-occurrence matrix and neural network community for determination of hypoxic fetuses. Prior Art 2: Investigating pH based evaluation of fetal heart rate (FHR) recordings. Prior Art 3: Prognostic model based on image-based time-frequencyfeatures and genetic algorithm for fetal hypoxia assessment. Prior Art 4: Modeling segmented cardiotocography time-series signals using one-dimensional convolutional neural networks for the early Detection of abnormal birth outcomes Existing technology 5: An attention-based CNN-BiLSTM hybrid neural network enhanced with features of discrete wavelet transformation for fetal acidosis classification.

Claims

1. A fetal distress auxiliary diagnostic system based on continuous wavelet transform and shallow convolutional neural network, characterized in that, include: The data acquisition and preprocessing module is responsible for acquiring the raw fetal heart rate signal and performing data preprocessing on it. The time-frequency feature map construction module is responsible for converting the signal after data acquisition and preprocessing into a two-dimensional time-frequency feature map using continuous wavelet transform; the classification and recognition module is responsible for inputting the time-frequency feature map into a shallow convolutional neural network model and outputting the classification result of fetal distress status.

2. The system according to claim 1, characterized in that, In the data acquisition and preprocessing module, preprocessing includes filling in missing values, eliminating interference signals, and extracting continuous valid segments.

3. The system according to claim 2, characterized in that, The missing values ​​in the signal are filled using linear interpolation.

4. The system according to claim 2, characterized in that, The interference cancellation signal is obtained by using a medium-range filter to eliminate spike pulses caused by interference in the signal.

5. The system according to claim 2, characterized in that, The term "truncation of continuous valid segments" refers to extracting the last 10 minutes of continuous valid segments from the signal.

6. The system according to claim 1, characterized in that, The specific implementation process of the time-frequency feature map construction module is as follows: for each segment of fetal heart rate signal after data acquisition and preprocessing, CWT calculation is performed using the mother wavelet function under n scale factors, generating a total of 2n time-frequency feature maps, where n≥2.

7. The system according to claim 6, characterized in that, In the time-frequency feature map construction module, the mother wavelet function is selected from Morlet wavelet and Gaussian 8th order wavelet.

8. The system according to claim 6, characterized in that, In the time-frequency feature map construction module, the n scale factors are 20, 25, and 30 to capture the features of different frequency components.

9. The system according to claim 6, characterized in that, The shallow convolutional neural network model includes an input layer, a convolutional layer, an activation layer, a normalization layer, a pooling layer, a fully connected layer, a Dropout layer, and a classification layer. The input layer receives a time-frequency feature map, and the convolutional layer uses convolutional kernels to extract features from the time-frequency feature map of the input layer. Subsequently, the activation layer, normalization layer, and fully connected layer perform nonlinear mapping, normalization, and downsampling on the extracted features. Then, the fully connected layer and Dropout layer flatten and overfit the extracted feature map. Finally, the classification layer outputs the classification result of the fetal state probability distribution.

Images