CN114569116A - Three-channel image and transfer learning-based ballistocardiogram ventricular fibrillation auxiliary diagnosis system - Google Patents

Three-channel image and transfer learning-based ballistocardiogram ventricular fibrillation auxiliary diagnosis system Download PDF

Info

Publication number
CN114569116A
CN114569116A CN202210039478.XA CN202210039478A CN114569116A CN 114569116 A CN114569116 A CN 114569116A CN 202210039478 A CN202210039478 A CN 202210039478A CN 114569116 A CN114569116 A CN 114569116A
Authority
CN
China
Prior art keywords
dimensional
channel
image
bcg
matrix
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210039478.XA
Other languages
Chinese (zh)
Inventor
邬小玫
万容茹
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Yiwu Research Institute Of Fudan University
Fudan University
Original Assignee
Fudan University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Fudan University filed Critical Fudan University
Priority to CN202210039478.XA priority Critical patent/CN114569116A/en
Publication of CN114569116A publication Critical patent/CN114569116A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/103—Detecting, measuring or recording devices for testing the shape, pattern, colour, size or movement of the body or parts thereof, for diagnostic purposes
    • A61B5/11—Measuring movement of the entire body or parts thereof, e.g. head or hand tremor, mobility of a limb
    • A61B5/1102—Ballistocardiography
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/02—Detecting, measuring or recording pulse, heart rate, blood pressure or blood flow; Combined pulse/heart-rate/blood pressure determination; Evaluating a cardiovascular condition not otherwise provided for, e.g. using combinations of techniques provided for in this group with electrocardiography or electroauscultation; Heart catheters for measuring blood pressure
    • A61B5/024—Detecting, measuring or recording pulse rate or heart rate
    • A61B5/02405—Determining heart rate variability
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7203—Signal processing specially adapted for physiological signals or for diagnostic purposes for noise prevention, reduction or removal
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7253—Details of waveform analysis characterised by using transforms
    • A61B5/7257—Details of waveform analysis characterised by using transforms using Fourier transforms
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/24323—Tree-organised classifiers
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

The invention belongs to the technical field of medical instruments, and particularly relates to an auxiliary diagnosis system for ventricular fibrillation of a ballistocardiogram based on three-channel images and transfer learning. The system comprises a signal preprocessing module, an image conversion module, a feature extraction module, a normalization processing module and a classifier; the method comprises the steps of converting one-dimensional BCG signals into three single-channel two-dimensional graphs in a self-defining mode, splicing the three-channel graphs into a three-channel image, taking the three-channel image as an input tensor, migrating a plurality of CNN models obtained by pre-training of a large-scale image library to a BCG domain for learning, extracting characteristic values from each model intermediate layer, splicing the characteristic values into one-dimensional characteristic vectors, and then automatically judging ventricular fibrillation, sinus rhythm and motion artifact through a full-connection layer or a machine learning classifier and the like. The method can effectively solve the problem of difficult ventricular fibrillation identification caused by the diversity of BCG waveforms and the characteristic of small sample size, and provides a feasible scheme for the long-term health monitoring of the cardiovascular diseases of the family.

Description

Three-channel image and transfer learning-based ballistocardiogram ventricular fibrillation auxiliary diagnosis system
Technical Field
The invention belongs to the technical field of medical instruments, and particularly relates to an auxiliary diagnosis system for ventricular fibrillation of a ballistocardiogram based on three-channel images and transfer learning.
Background
Sudden Cardiac Death (SCD) is a major public health burden worldwide, and due to medical constraints, most SCDs occur outside the hospital. Ventricular fibrillation is the leading cause of SCD, is uncertain and highly critical, and once it occurs, ventricular output is reduced sharply, the patient's circulation is interrupted, and sudden death occurs within minutes if it is not intervened in time. Electrical defibrillation is the only effective method for treating ventricular fibrillation, the success probability of the electrical defibrillation is inversely proportional to the attack duration of the ventricular fibrillation, and the resuscitation success rate is reduced by nearly 10% every minute of delay, so that a method for monitoring the cardiac function in real time in a non-contact manner and accurately identifying the ventricular fibrillation is urgently needed.
At present, the clinical diagnosis of ventricular fibrillation is mainly based on electrocardiogram. However, the electrocardiographic examination requires the electrodes to be attached to the body surface of the patient, and the patient feels inconvenience and discomfort after long-term use, so that the electrocardiographic examination is not suitable for long-term heart monitoring. The above problems can be solved by a Ballistocardiogram (BCG), an interference-free, non-contact cardiac monitoring technique. BCG catches the heart and penetrates the removal of blood time health barycenter through the sensor of integration in daily environment such as mattress, cushion, back, and then provides the holistic performance information of circulation system for the user. At present, the traditional machine learning method is mostly adopted for the heart disease detection algorithm research of the BCG, various characteristics of signals are generally required to be extracted, and the BCG is easily influenced by detection equipment, sitting and lying postures, external vibration interference and the like, so that the waveform diversity is increased, the individual difference is large, and the universality of the machine learning method based on the characteristic engineering is insufficient. Meanwhile, the current BCG lacks a relatively perfect database, and deep learning models such as a Convolutional Neural Network (CNN) have large-batch requirements on input data sets so as to be convenient for full training and debugging of the network, so that the method is not suitable for small sample researches such as the BCG. In conclusion, a computer-aided diagnosis method for ventricular fibrillation suitable for daily BCG heart monitoring is needed to meet the long-term non-binding health management requirement of cardiovascular disease families.
Disclosure of Invention
Aiming at the ventricular fibrillation detection algorithm requirements and the technical defects of the conventional BCG research, the invention provides a Ballistocardiogram (BCG) ventricular fibrillation auxiliary diagnosis system based on three-channel images and transfer learning, so as to solve the problem of difficult ventricular fibrillation identification caused by the waveform diversity and the small sample quantity characteristic of the BCG.
The method converts one-dimensional BCG signals into three-channel images in a self-defined single-channel image combination mode, takes the three-channel images as an input tensor, migrates a plurality of CNN models obtained by pre-training of a large image library to a BCG domain for learning, extracts characteristic values from the middle layers of the models and splices the characteristic values into one-dimensional characteristic vectors, and then realizes the discrimination of ventricular fibrillation, sinus rhythm and motion artifact through a full-connection layer or a machine learning classifier and the like.
The BCG ventricular fibrillation auxiliary diagnosis system based on the three-channel image and the transfer learning comprises a signal preprocessing module, an image conversion module, a feature extraction module, a normalization processing module and a classifier; wherein:
and the signal preprocessing module is used for preprocessing the acquired BCG signals. Specifically, mother wavelet Daubechie 6 is selected to decompose an original BCG signal, high and low frequency noise components are removed, and a heartbeat related frequency band part (0.5-15Hz) is selected to reconstruct the signal.
The image conversion module is used for converting the one-dimensional BCG signal obtained by the signal preprocessing module into a three-channel two-dimensional image so as to construct a transfer learning input tensor. Specifically, three single-channel images are acquired, including but not limited to a time-frequency map, a Gram Angular Field (GAF), and a custom autocorrelation sequence map (ACM). Wherein:
the time-frequency diagram acquisition mode is as follows: acquiring a time-frequency coefficient matrix through short-time Fourier transform (STFT) or Continuous Wavelet Transform (CWT) and the like, and then standardizing the size of the matrix through matrix compression transform to obtain a required single-channel time-frequency gray-scale image;
the GAF graph acquisition mode is as follows: converting the one-dimensional BCG signal under the Cartesian coordinate system into a polar coordinate system for representation, and generating a GAF matrix by using a trigonometric function, thereby obtaining a required single-channel gray scale image;
the ACM graph acquisition mode is as follows: reducing the dimension of the one-dimensional BCG signal from a long sequence to a short sequence, calculating an autocorrelation coefficient sequence of a short sequence sample and carrying out absolute value calculation, and normalizing the autocorrelation sequence value to a corresponding range and carrying out integer calculation according to the tensor size; then, a horizontal axis represents a time dimension, a vertical axis represents an amplitude dimension, and the processed autocorrelation sequence values are mapped into a two-dimensional gray scale map;
and splicing the three single-channel gray-scale images along the channel dimension to obtain the three-dimensional input tensor required by the transfer learning.
The feature extraction module takes the three-dimensional tensor obtained by the image conversion module as input and applies transfer learning to extract effective features; the method specifically comprises the steps that three classic CNN network models including VGG16, IncepotionV 3 and ResNet50 are pre-trained by a computer vision standard data set ImageNet, and the models are transferred to a three-dimensional input tensor obtained by BCG conversion; and extracting the characteristic graphs from the model intermediate layers respectively, and obtaining the one-dimensional characteristic vector through transformation and combination.
The normalization processing module is used for analyzing and cleaning the feature vectors extracted by the feature extraction module through variance and mutual information parameters, and performing normalization processing on the features to be used as input of a subsequent classifier.
The classifier adopts a full connection layer or adopts machine learning classifiers such as logistic regression, support vector machines or random forests. Dividing three types of signals of ventricular fibrillation, sinus rhythm and motion artifact into a training set and a test set through layered sampling; acquiring the optimal hyper-parameters of the model by a cross validation and grid search method on a training set; on the basis, the model is trained again by applying the whole training set, and the network structure and parameters of the classifier model are determined; and evaluating the generalization performance of the final model through the test set.
In the invention, the three-channel image is constructed by the following specific processes:
(1) specifying the image size as (N, N), where N is a custom number greater than 139;
(2) converting the preprocessed one-dimensional BCG signal into an ACM (adaptive computer graphics) diagram, wherein the specific mode is as follows:
reducing the dimension of a long sequence data segment into a short sequence by a classic segmented aggregation approximation (PAA) algorithm according to a formula (1); the long sequence data segment is represented by x, the length is L, the short sequence is represented by xpaaDenotes a length of N;
② calculating short sequence sample xpaaAfter absolute value normalization to [0, N-1 ]]Within the range, and integer to obtain the sequence xacf；
Establishing an all-zero matrix A with the size of (N, N) aiming at each sample point x in the autocorrelation sequenceacf[n]Updating the matrix A according to the formula (2) when N is 0 … N-1, and taking each data of the matrix A as the gray value of a pixel corresponding to the single-channel gray map ACM, wherein the matrix A comprises three levels of 0, 128 and 255;
Figure RE-GDA0003631323480000031
wherein
Figure RE-GDA0003631323480000032
Figure RE-GDA0003631323480000033
Wherein I ═ xacf[n]，n＝0…N-1，(2)
(3) Converting the preprocessed one-dimensional BCG signal into a GAF image, wherein the specific mode is as follows:
reducing dimension of long sequence data segment into short sequence by PAA algorithm, taking x aspaaDenotes a length of N;
② normalizing xpaaCompressing the sequence value to [ -1,1 [ -1 [ ]]In the range of xnormRepresents;
thirdly, according to the relation shown in the formula (3), the time sequence value x under the rectangular coordinate system is processednormConverting the polar coordinate system into a polar coordinate system (r, theta) representation, and solving a polar angle theta;
using m and n to respectively represent different time points, as shown in formula (4), applying angle sum trigonometric function transformation to obtain a needed GAF matrix G, and mapping matrix values to the range of [0,255] to obtain a single-channel gray scale map;
Figure RE-GDA0003631323480000034
g [ m ] ═ cos (θ [ m ] + θ [ N ]), where m, N is 0 … N-1, (4)
(4) Converting the preprocessed one-dimensional BCG signal into a time-frequency diagram, wherein the specific mode is as follows: obtaining a time-frequency coefficient matrix through STFT or CWT, adjusting the size of the matrix to be (N, N), and mapping the matrix value to the range of [0,255] to obtain a single-channel gray-scale map;
(5) combining the three single-channel graphs into a (N, N,3) or (3, N, N) three-dimensional image according to a mode of adding one layer of ACM graph and one layer of GAF graph and one layer of time-frequency graph or other self-defining modes, and normalizing the gray value of each channel image to be in a [0,1] range to be used as an input tensor of a feature extraction module of subsequent transfer learning.
In the invention, the characteristic extraction module extracts effective characteristics by applying transfer learning; the specific process is as follows:
(1) three classical CNN network models were constructed: VGG16, IncepotionV 3 and ResNet50, pre-training and tuning of the model is carried out through a computer vision standard data set ImageNet;
(2) freezing the model parameters, and taking the three-dimensional tensor obtained by normalizing the user-defined three-channel image as network input for transfer learning; and extracting the feature map from each model intermediate layer, and obtaining a one-dimensional feature vector through global average pooling and splicing.
In the present invention, the feature cleaning is developed in a filtering concept, first at 10-8Removing low variance features as a threshold; then 10% of the low quality features are removed based on the mutual information value. Normalizing each feature retained to [0, 1%]In range, as a subsequent classifier input.
In the invention, the classifier takes the extracted features as input to judge the ventricular fibrillation, the sinus rhythm and the motion artifact of the signal. The classifier can be constructed by a full connection layer or a machine learning classifier, wherein the full connection layer selects a Softmax activation function as an output layer, model optimization is carried out by adopting an Adam algorithm, and the advantages and disadvantages are evaluated by a cross entropy loss function; the machine learning classifier includes logistic regression, support vector machine or random forest.
The invention has the advantages that the BCG ventricular fibrillation auxiliary diagnosis system based on the three-channel image and the transfer learning is provided, the BCG acquired noninductively can be applied to assist in identifying ventricular fibrillation attacks, the BCG is mapped from one dimension to three-dimensional tensor by skillfully converting, and a classic model obtained by the pre-training of the ImageNet data set is transferred to the BCG actual measurement data of a small batch of samples by utilizing the richness of an open source image library and the maturity of computer vision related deep learning research to perform the discrimination of ventricular fibrillation and non-ventricular fibrillation. The algorithm provided by the invention needs fewer retraining parameters, has stronger model robustness, better diagnosis accuracy and good practicability, and provides a feasible idea for the long-term health monitoring of cardiovascular diseases in families.
Drawings
FIG. 1 is a system block diagram of an embodiment of the invention.
Fig. 2 is a single-channel gray-scale image of the measured ventricular fibrillation BCG signal and its conversion according to the embodiment of the present invention. In the figure, (a), (b), (c) and (d) are respectively gray scale graphs of one-dimensional ventricular fibrillation BCG signals, ACM, GAF and STFT.
Fig. 3 is a single channel gray scale image of the measured sinus rhythm BCG signal and its transformation according to an embodiment of the present invention. In the figure, (a), (b), (c) and (d) are one-dimensional sinus rhythm BCG signal and ACM, GAF and STFT gray scale images, respectively.
Fig. 4 is a single-channel gray-scale image of the measured motion artifact BCG signal and its conversion according to the embodiment of the present invention. In the figure, (a), (b), (c) and (d) are gray scale graphs of one-dimensional motion artifact BCG signal, ACM, GAF and STFT respectively.
Fig. 5 is a structural diagram of a feature extraction module of transfer learning according to an embodiment of the present invention.
Detailed Description
Specific embodiments of the present invention will be further described with reference to the accompanying drawings. The following examples are intended to illustrate the invention but are not intended to limit the scope of the invention as claimed.
In this embodiment, a BCG ventricular fibrillation auxiliary diagnosis system based on a custom three-channel image and transfer learning, as shown in fig. 1, includes the following steps:
step 1: the signal preprocessing module preprocesses the signal to obtain a pure BCG signal.
Step 1.1: artificially extracting a motion artifact signal in the BCG signal;
step 1.2: wavelet transforms are applied for noise filtering. Decomposing the original BCG signal into seven layers by using a mother wavelet Daubechie 6 to obtain seven detail components (D1-D7), and recombining the detail components D3-D6 containing heartbeat-related frequency bands to obtain a filtered BCG signal;
step 1.3: the signal is divided into data segments of 7 seconds in length.
Step 2: the image conversion module converts the image, and the image size is (256 ).
Step 2.1: converting the preprocessed one-dimensional BCG signal into an ACM (adaptive finite-order-of-arrival) diagram, wherein the specific mode is as follows:
reducing dimension of a long sequence data segment (represented by x and with the length of L) into a short sequence by a PAA algorithm according to a formula (1), wherein x is used forpaaExpressed, length N is 256;
② calculating short sequence sample xpaaIs normalized to [0,255] after being absolute-valued]Within the range, and integer to obtain the sequence xacf；
Creating a full zero matrix A with the size of (256 ) aiming at each sample point x in the autocorrelation sequenceacf[n]N is 0 … N-1, according to the formula (2)The new matrix a. Taking each data of the matrix A as the gray value of the corresponding pixel of the single-channel gray-scale map ACM, as shown in a subgraph (b) in FIGS. 2 to 4;
Figure RE-GDA0003631323480000051
wherein
Figure RE-GDA0003631323480000052
Figure RE-GDA0003631323480000053
Wherein I ═ xacf[n]，n＝0…255，(2)
Step 2.2: converting the preprocessed one-dimensional BCG signal into a GAF image, wherein the specific mode is as follows:
reducing dimension of long sequence data segment into short sequence by PAA algorithm, taking x aspaaIndicated, length 256;
② normalizing xpaaCompressing the time series values to [ -1,1 [)]In the range of xnormRepresents;
thirdly, according to the relation shown in the formula (3), the sequence value x under the rectangular coordinate system is processednormConverting the coordinate system into a polar coordinate system (r, theta) expression, and solving a polar angle theta;
using m and n to respectively represent different time points, as shown in formula (4), applying angle sum trigonometric function transformation to obtain a GAF matrix G required by the embodiment, and mapping matrix values to a range of [0,255] to obtain a single-channel gray scale map, as shown in subgraph (c) in fig. 2 to fig. 4;
Figure RE-GDA0003631323480000054
g [ m ] ═ cos (θ [ m ] + θ [ n ]), where m, n is 0 … 255, (4)
Step 2.3: the preprocessed one-dimensional BCG signal is converted into a single channel time-frequency (STFT) graph. And short-time Fourier transform is adopted, a hann window is selected for the transform, the window length is set to be 2 seconds, and the offset length is set to be 0.04 seconds. Adjusting the shape of the time-frequency coefficient matrix to (256 ), and mapping the matrix value to the range of [0,255] to obtain the required single-channel gray-scale map, as shown in subgraph (d) in fig. 2 to fig. 4;
step 2.4: combining three single-channel images into a (256, 3) three-dimensional image according to a mode of adding one layer of ACM image and one layer of GAF image and one layer of STFT image, and normalizing the gray value of each channel image to be in a range of [0,1] to be used as an input tensor of subsequent transfer learning.
And 3, step 3: and constructing a feature vector by a feature extraction module through transfer learning.
Step 3.1: three classical CNN network models were constructed: VGG16, IncepotionV 3 and ResNet50, pre-training and tuning of the model is carried out through a computer vision standard data set ImageNet; the structure of the feature extraction module for migration learning is shown in fig. 5, in which FC, BN, ReLU, Conv2D, MaxPooling2D, globalagepooling 2D, and ZeroPadding2D respectively represent a full connection layer, batch normalization, a ReLU activation layer, a two-dimensional convolution layer, a two-dimensional maximum pooling layer, a two-dimensional global average pooling layer, and a two-dimensional zero padding layer; vgg _ block, initiation _ block and res _ block respectively represent functional blocks of three CNN models, the internal structures are different, and since the above are all classical models, a more detailed description is not provided herein.
Step 3.2: and (4) freezing the model parameters, and performing transfer learning by taking the three-dimensional tensor obtained in the step (2.4) as network input. And extracting feature maps from the middle layers of the models, and obtaining a one-dimensional feature vector with the length of 1056 through global average pooling and splicing.
And 4, step 4: the features are cleaned and normalized by a normalization processing module.
And (4) cleaning the characteristics by adopting a filtering type thought. Firstly, the number of the particles is 10-8Removing low variance features as a threshold; then 10% of the low quality features are removed based on the mutual information value. Normalizing each feature retained to [0, 1%]Within the range;
and 5: classifying and judging by a classifier;
step 5.1: constructing a classifier by a full connection layer, and selecting a Softmax activation function as an output layer; optimizing the model by adopting an Adam algorithm, and evaluating the advantages and disadvantages by using a cross entropy loss function;
step 5.2: and (4) forming the input of the classifier by the feature vector set obtained in the step (4), wherein the output is the result of judging the sample to be ventricular fibrillation, sinus rhythm or motion artifact.
The above is a preferred embodiment of the present invention, but the scope of the present invention is not limited to this example, and any other examples made by those skilled in the art within the technical scope of the present invention through simple replacement and replacement should be covered within the scope of the present invention. The protection scope of the present invention is subject to the protection scope defined by the claims.

Claims (5)

1. A ballistocardiogram ventricular fibrillation auxiliary diagnosis system based on three-channel images and transfer learning is characterized by comprising a signal preprocessing module, an image conversion module, a feature extraction module, a normalization processing module and a classifier; wherein:
the signal preprocessing module is used for preprocessing the acquired BCG signals; specifically, mother wavelet Daubechie 6 is selected to decompose original BCG signals, high and low frequency noise components are removed, and a frequency band part of 0.5-15Hz related to heartbeat is selected to reconstruct the signals;
the image conversion module is used for converting the one-dimensional BCG signal obtained by the signal preprocessing module into a three-channel two-dimensional image so as to construct a transfer learning input tensor; the method specifically comprises the steps of firstly, acquiring three single-channel images, wherein the three single-channel images comprise a time-frequency image, a Gram Angle Field (GAF) and a self-defined autocorrelation sequence mapping (ACM); wherein:
the time-frequency diagram acquisition mode is as follows: acquiring a time-frequency coefficient matrix through short-time Fourier transform (STFT) or Continuous Wavelet Transform (CWT); then, the size of the matrix is standardized through matrix compression transformation, and a required single-channel time-frequency gray-scale image is obtained;
the GAF graph acquisition mode is as follows: converting the one-dimensional BCG signal under the Cartesian coordinate system into a polar coordinate system for representation, and generating a GAF matrix by using a trigonometric function, thereby obtaining a required single-channel gray scale image;
the ACM graph acquisition mode is as follows: reducing the dimension of the one-dimensional BCG signal from a long sequence to a short sequence, calculating an autocorrelation coefficient sequence of a short sequence sample and carrying out absolute value calculation, and normalizing the autocorrelation sequence value to a corresponding range and carrying out integer calculation according to the tensor size; then, a horizontal axis represents a time dimension, a vertical axis represents an amplitude dimension, and the processed autocorrelation sequence values are mapped into a two-dimensional gray scale map;
splicing the three single-channel gray level graphs along the channel dimension to obtain a three-dimensional input tensor required by the transfer learning;
the feature extraction module takes the three-dimensional tensor obtained by the image conversion module as input and applies transfer learning to extract effective features; the method specifically comprises the steps that three classic CNN network models including VGG16, IncepotionV 3 and ResNet50 are pre-trained by a computer vision standard data set ImageNet, and the models are transferred to a three-dimensional input tensor obtained by BCG conversion; extracting feature maps from the model intermediate layers respectively, and obtaining one-dimensional feature vectors through transformation and combination;
the normalization processing module is used for analyzing and cleaning the feature vectors extracted by the feature extraction module through variance and mutual information parameters, and performing normalization processing on the features to be used as the input of a subsequent classifier;
the classifier adopts a full connection layer or adopts a machine learning classifier such as logistic regression, a support vector machine or random forest; dividing three types of signals of ventricular fibrillation, sinus rhythm and motion artifact into a training set and a test set through layered sampling; acquiring the optimal hyper-parameters of the model by a cross validation and grid search method of a training set; on the basis, the model is trained again by applying the whole training set, the network structure and the parameters of the classifier model are determined, and the generalization performance of the final model is evaluated through the test set.
2. The ballistocardiogram ventricular fibrillation auxiliary diagnostic system according to claim 1, wherein the three-channel image is constructed by the following specific procedures:
(1) specifying the image size as (N, N), where N is a custom number greater than 139;
(2) converting the preprocessed one-dimensional BCG signal into an ACM (adaptive finite-order-of-arrival) diagram, wherein the specific mode is as follows:
reducing the dimension of a long sequence data segment into a short sequence by a classic segmented aggregation approximation (PAA) algorithm according to a formula (1); the long sequence data segment is represented by x, the length is L, the short sequence is represented by xpaaDenotes a length of N;
② calculating short sequence sample xpaaAfter absolute value normalization to [0, N-1 ]]Within the range, and integer to obtain the sequence xacf；
Establishing an all-zero matrix A with the size of (N, N) aiming at each sample point x in the autocorrelation sequenceacf[n]Updating the matrix A according to a formula (2), wherein the data of the matrix A is used as the gray value of the pixel corresponding to the single-channel gray-scale map ACM and comprises three levels of 0, 128 and 255;
Figure RE-FDA0003631323470000021
Figure RE-FDA0003631323470000022
(3) converting the preprocessed one-dimensional BCG signal into a GAF image, wherein the specific mode is as follows:
reducing dimension of long sequence data segment into short sequence by PAA algorithm, taking x aspaaDenotes, length is N;
② normalizing xpaaCompressing the sequence value to [ -1,1 [ -1 [ ]]In the range of xnormRepresents;
thirdly, according to the relation shown in the formula (3), the time sequence value x in the rectangular coordinate system is calculatednormConverting the coordinate system into a polar coordinate system (r, theta) expression, and solving a polar angle theta;
respectively representing different time points by m and n, as shown in formula (4), obtaining a needed GAF matrix G by applying angle sum trigonometric function transformation, and mapping matrix values to a range of [0,255] to obtain a single-channel gray scale image;
Figure RE-FDA0003631323470000023
g [ m ] ═ cos (θ [ m ] + θ [ N ]), where m, N is 0 … N-1, (4)
(4) Converting the preprocessed one-dimensional BCG signal into a time-frequency diagram, wherein the specific mode is as follows: obtaining a time-frequency coefficient matrix through STFT or CWT, adjusting the size of the matrix to be (N, N), and mapping the matrix value to the range of [0,255] to obtain a single-channel gray-scale map;
(5) combining three single-channel graphs into a (N, N,3) or (3, N, N) three-dimensional image according to a mode of adding one layer of ACM graph and one layer of GAF graph and one layer of time-frequency graph or other self-defining modes, and normalizing the gray value of each channel image to be in a range of [0,1] to be used as an input tensor of a subsequent transfer learning feature extractor.
3. The ballistocardiogram ventricular fibrillation auxiliary diagnosis system according to claim 2, wherein the feature extraction module applies transfer learning to extract effective features, and the specific process is as follows:
(1) three classical CNN network models were constructed: VGG16, IncepotionV 3 and ResNet50, pre-training and tuning of the model is carried out through a computer vision standard data set ImageNet;
(2) freezing the model parameters, and taking the three-dimensional tensor obtained by normalizing the user-defined three-channel image as network input for transfer learning; and extracting the feature map from each model intermediate layer, and obtaining a one-dimensional feature vector through global average pooling and splicing.
4. The ballistocardiogram ventricular fibrillation-assisted diagnosis system of claim 3, wherein the features are washed first at 10-8Removing low variance features as a threshold; then, removing 10% of low-quality features according to the mutual information value; normalizing each feature retained to [0, 1%]In range, as a subsequent classifier input.
5. The ballistocardiogram ventricular fibrillation aided diagnosis system according to claim 4, wherein the classifier is constructed by a full connection layer, wherein the full connection layer adopts a Softmax activation function as an output layer, model optimization is performed by adopting an Adam algorithm, and the advantages and disadvantages are evaluated by using a cross entropy loss function.
CN202210039478.XA 2022-01-13 2022-01-13 Three-channel image and transfer learning-based ballistocardiogram ventricular fibrillation auxiliary diagnosis system Pending CN114569116A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210039478.XA CN114569116A (en) 2022-01-13 2022-01-13 Three-channel image and transfer learning-based ballistocardiogram ventricular fibrillation auxiliary diagnosis system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210039478.XA CN114569116A (en) 2022-01-13 2022-01-13 Three-channel image and transfer learning-based ballistocardiogram ventricular fibrillation auxiliary diagnosis system

Publications (1)

Publication Number Publication Date
CN114569116A true CN114569116A (en) 2022-06-03

Family

ID=81771875

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210039478.XA Pending CN114569116A (en) 2022-01-13 2022-01-13 Three-channel image and transfer learning-based ballistocardiogram ventricular fibrillation auxiliary diagnosis system

Country Status (1)

Country Link
CN (1) CN114569116A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117274117A (en) * 2023-11-23 2023-12-22 合肥工业大学 Frequency domain pseudo-color enhanced magnetocardiogram signal characteristic image generation method and storage medium

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117274117A (en) * 2023-11-23 2023-12-22 合肥工业大学 Frequency domain pseudo-color enhanced magnetocardiogram signal characteristic image generation method and storage medium
CN117274117B (en) * 2023-11-23 2024-02-02 合肥工业大学 Frequency domain pseudo-color enhanced magnetocardiogram signal characteristic image generation method and storage medium

Similar Documents

Publication Publication Date Title
CN110876626B (en) Depression detection system based on optimal lead selection of multi-lead electroencephalogram
Li et al. Deep convolutional neural network based ECG classification system using information fusion and one-hot encoding techniques
Luo et al. Patient-specific deep architectural model for ECG classification
CN110840402B (en) Atrial fibrillation signal identification method and system based on machine learning
CN111990989A (en) Electrocardiosignal identification method based on generation countermeasure and convolution cyclic network
CN107736894A (en) A kind of electrocardiosignal Emotion identification method based on deep learning
EP4042445A1 (en) Systems and methods for reduced lead electrocardiogram diagnosis using deep neural networks and rule-based systems
CN112070785A (en) Medical image analysis method based on computer vision
CN111914925B (en) Patient behavior multi-modal perception and analysis system based on deep learning
CN115530788A (en) Arrhythmia classification method based on self-attention mechanism
CN110575178A (en) Diagnosis and monitoring integrated medical system for judging motion state and judging method thereof
CN114569116A (en) Three-channel image and transfer learning-based ballistocardiogram ventricular fibrillation auxiliary diagnosis system
CN113116300A (en) Physiological signal classification method based on model fusion
Allam et al. Patient-specific ECG beat classification using EMD and deep learning-based technique
CN116807491A (en) Myocardial infarction global lead reconstruction method applied to real-time electrocardiograph monitoring equipment
CN115736920A (en) Depression state identification method and system based on bimodal fusion
CN113509186B (en) ECG classification system and method based on deep convolutional neural network
CN116172573A (en) Arrhythmia image classification method based on improved acceptance-ResNet-v 2
Khanmohammadi et al. Fetal gender identification using machine and deep learning algorithms on phonocardiogram signals
Vimala et al. Classification of cardiac vascular disease from ECG signals for enhancing modern health care scenario
Hurr et al. Feature extraction and recognition of human physiological signals based on the convolutional neural network
Sanamdikar et al. Classification of ECG Signal for Cardiac Arrhythmia Detection Using GAN Method
CN115251889B (en) Method for describing characteristics of dynamic connection network of functional magnetic resonance image
Hakim et al. Tuberculosis detection analysis using texture features on CXRs images
Vysiya et al. Automatic detection of cardiac arrhythmias in ECG signal for IoT application

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
TA01 Transfer of patent application right
TA01 Transfer of patent application right

Effective date of registration: 20221216

Address after: 200433 No. 220, Handan Road, Shanghai, Yangpu District

Applicant after: FUDAN University

Applicant after: Yiwu Research Institute of Fudan University

Address before: 200433 No. 220, Handan Road, Shanghai, Yangpu District

Applicant before: FUDAN University