CN110956197A - Method and system for establishing pulse wave noise signal identification model based on convolutional neural network - Google Patents
Method and system for establishing pulse wave noise signal identification model based on convolutional neural network Download PDFInfo
- Publication number
- CN110956197A CN110956197A CN201911030823.8A CN201911030823A CN110956197A CN 110956197 A CN110956197 A CN 110956197A CN 201911030823 A CN201911030823 A CN 201911030823A CN 110956197 A CN110956197 A CN 110956197A
- Authority
- CN
- China
- Prior art keywords
- noise signal
- signal data
- neural network
- convolutional neural
- network model
- 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
Links
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 105
- 238000000034 method Methods 0.000 title claims abstract description 58
- 238000011176 pooling Methods 0.000 claims abstract description 41
- 230000004913 activation Effects 0.000 claims abstract description 37
- 210000002569 neuron Anatomy 0.000 claims abstract description 26
- 238000012549 training Methods 0.000 claims abstract description 26
- 238000007781 pre-processing Methods 0.000 claims abstract description 13
- 238000010276 construction Methods 0.000 claims description 11
- 238000012545 processing Methods 0.000 claims description 10
- 108010076504 Protein Sorting Signals Proteins 0.000 claims description 7
- 239000006185 dispersion Substances 0.000 claims description 7
- 239000000284 extract Substances 0.000 claims description 5
- 239000010410 layer Substances 0.000 description 52
- 230000006870 function Effects 0.000 description 43
- 238000012360 testing method Methods 0.000 description 14
- 238000013528 artificial neural network Methods 0.000 description 13
- 230000008569 process Effects 0.000 description 10
- 238000005457 optimization Methods 0.000 description 7
- 238000003745 diagnosis Methods 0.000 description 6
- 238000005070 sampling Methods 0.000 description 6
- 238000009826 distribution Methods 0.000 description 4
- 238000003759 clinical diagnosis Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000010606 normalization Methods 0.000 description 3
- 210000002321 radial artery Anatomy 0.000 description 3
- 238000013529 biological neural network Methods 0.000 description 2
- 239000003814 drug Substances 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000001788 irregular Effects 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 239000002356 single layer Substances 0.000 description 2
- 239000012855 volatile organic compound Substances 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000003287 bathing Methods 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000007477 logistic regression Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 230000002040 relaxant effect Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 230000035939 shock Effects 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/48—Other medical applications
- A61B5/4854—Diagnosis based on concepts of traditional oriental medicine
-
- 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
- A61B5/7264—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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
- G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
- G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
- G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
- G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Theoretical Computer Science (AREA)
- Medical Informatics (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Biomedical Technology (AREA)
- Evolutionary Computation (AREA)
- Veterinary Medicine (AREA)
- Public Health (AREA)
- Animal Behavior & Ethology (AREA)
- Surgery (AREA)
- Heart & Thoracic Surgery (AREA)
- Pathology (AREA)
- Biophysics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Signal Processing (AREA)
- Physiology (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Psychiatry (AREA)
- Biodiversity & Conservation Biology (AREA)
- Alternative & Traditional Medicine (AREA)
- Fuzzy Systems (AREA)
- Mathematical Physics (AREA)
- Multimedia (AREA)
- Cardiology (AREA)
- Measuring Pulse, Heart Rate, Blood Pressure Or Blood Flow (AREA)
Abstract
The invention discloses a method and a system for establishing a pulse wave noise signal identification model based on a convolutional neural network, wherein the method comprises the following steps: preprocessing pulse wave noise signal data; extracting features of the processed noise signal data through a plurality of convolutional layers of a convolutional network model; extracting key information of the processed noise signal data through a maximum pooling layer; calculating the weight of each neuron of the convolutional neural network model through a dense layer activation function; calculating the processed noise signal data through a convolutional neural network model, and outputting an identification result of the noise signal data; adjusting the weight of each neuron of the convolutional neural network model through the error of the recognition result of the noise signal data, and performing multiple iterative cycle training on the convolutional neural network model; and verifying the convolutional neural network model by using the unidentified noise signal data in the processed noise signal data to obtain the verified convolutional neural network model.
Description
Technical Field
The invention relates to the technical field of pulse wave signal processing, in particular to a method and a system for establishing a pulse wave noise signal identification model based on a convolutional neural network.
Background
The digital pulse diagnosis acquires pulse wave signals through a pressure sensor attached to the radial artery. The technology gradually leads the pulse diagnosis to be quantitative and standardized, provides technical support for more objectively and accurately carrying out clinical diagnosis, auxiliary diagnosis and curative effect judgment on the pulse diagnosis of the traditional Chinese medicine, and further promotes the modernization process of the traditional Chinese medicine.
The nature of the problem faced in identifying digitized pulse noise signals is to perform a binary classification of successive one-dimensional time series. Traditional classifiers such as (logistic regression, linear SVM) are mostly used for linear segmentation; for a nonlinear separable sample, a plurality of kernel functions or characteristic mappings can be added to form a curve or a curved surface to separate the sample, and the problems of overfitting and characteristic extraction are easy to occur because the distribution of the sample characteristics is irregular.
Therefore, a technology is needed to realize a technology for establishing a pulse wave noise signal identification model based on a convolutional neural network.
Disclosure of Invention
The technical scheme of the invention provides a method and a system for establishing a pulse wave noise signal identification model based on a convolutional neural network, which aim to solve the problem of how to establish the pulse wave noise signal identification model based on the convolutional neural network.
In order to solve the above problem, the present invention provides a method for establishing a pulse wave noise signal identification model based on a convolutional neural network, the method comprising:
preprocessing pulse wave noise signal data, reducing the discrete degree of the pulse wave signal data, and acquiring the processed noise signal data;
inputting the processed noise signal data into a convolutional neural network model, and extracting the characteristics of the processed noise signal data through a plurality of convolutional layers of the convolutional neural network model;
extracting key information of the processed noise signal data through a maximum pooling layer based on characteristics of the noise signal data;
calculating the weight of each neuron of the convolutional neural network model through a dense layer activation function;
calculating the processed noise signal data through the convolutional neural network model, and outputting an identification result of the noise signal data;
adjusting the weight of each neuron of the convolutional neural network model according to the error of the recognition result of the noise signal data, and performing multiple iterative cycle training on the convolutional neural network model;
and verifying the convolutional neural network model by using the unrecognized noise signal data in the processed noise signal data to obtain the verified convolutional neural network model.
Preferably, the preprocessing the pulse wave noise signal data includes:
the pulse wave noise signal data is normalized by the following formula:
y＝(x-a)/c
wherein x represents a discretized digitized pulse noise signal;
a ═ avg (x) represents the mean, median, or average calculated from the whole sample of the individual pulse noise signal sequences, or the center of the pulse noise signal predicted by other methods;
c is a constant value, and is a dispersion mean value of x obtained through experimental data statistics or a standard deviation after calibration according to pulse wave noise signal data.
Preferably, the extracting, by the max-pooling layer, key information of the processed noise signal data includes:
and extracting key information of the processed noise signal data through a maximum value downsampling method through a maximum pooling layer.
Preferably, the pooling radius is proportional to the convolution kernel size of the corresponding convolution layer.
Preferably, the weights of the neurons of the convolutional neural network model are calculated by a dense layer activation function, wherein the activation function uses a ReLU activation function.
According to another aspect of the present invention, there is provided a system for establishing a pulse wave noise signal identification model based on a convolutional neural network, the system comprising:
the device comprises an initial unit, a data processing unit and a data processing unit, wherein the initial unit is used for preprocessing pulse wave noise signal data, reducing the discrete degree of the pulse wave signal data and acquiring the processed noise signal data;
a first construction unit, configured to input the processed noise signal data to a convolutional neural network model, and extract features of the processed noise signal data through a plurality of convolutional layers of the convolutional neural network model;
a second construction unit, configured to extract, based on features of the noise signal data, key information of the processed noise signal data through a maximum pooling layer;
the third construction unit is used for calculating the weight of each neuron of the convolutional neural network model through a dense layer activation function;
the output unit is used for calculating the processed noise signal data through the convolutional neural network model and outputting the identification result of the noise signal data;
the training unit is used for adjusting the weight of each neuron of the convolutional neural network model according to the error of the recognition result of the noise signal data and carrying out multiple iterative cycle training on the convolutional neural network model;
and the obtaining unit is used for verifying the convolutional neural network model by using the unidentified noise signal data in the processed noise signal data to obtain the verified convolutional neural network model.
Preferably, the initialization unit is configured to pre-process the pulse noise signal data, and is further configured to:
the pulse wave noise signal data is normalized by the following formula:
y＝(x-a)/c
wherein x represents a discretized digitized pulse noise signal;
a ═ avg (x) represents the mean, median, or average calculated from the whole sample of the individual pulse noise signal sequences, or the center of the pulse noise signal predicted by other methods;
c is a constant value, and is a dispersion mean value of x obtained through experimental data statistics or a standard deviation after calibration according to pulse wave noise signal data.
Preferably, the second construction unit is configured to extract key information of the processed noise signal data through a maximum pooling layer, and includes:
and extracting key information of the processed noise signal data through a maximum value downsampling method through a maximum pooling layer.
Preferably, the pooling radius is proportional to the convolution kernel size of the corresponding convolution layer.
Preferably, the weights of the neurons of the convolutional neural network model are calculated by a dense layer activation function, wherein the activation function uses a ReLU activation function.
The technical scheme of the invention provides a method and a system for establishing a pulse wave noise signal identification model based on a convolutional neural network, wherein the method comprises the following steps: preprocessing pulse wave noise signal data, reducing the discrete degree of the pulse wave signal data, and acquiring the processed noise signal data; inputting the processed noise signal data into a convolutional neural network model, and extracting the characteristics of the processed noise signal data through a plurality of convolutional layers of the convolutional neural network model; extracting key information of the processed noise signal data through a maximum pooling layer based on the characteristics of the noise signal data; calculating the weight of each neuron of the convolutional neural network model through a dense layer activation function; calculating the processed noise signal data through a convolutional neural network model, and outputting an identification result of the noise signal data; adjusting the weight of each neuron of the convolutional neural network model through the error of the recognition result of the noise signal data, and performing multiple iterative cycle training on the convolutional neural network model; and verifying the convolutional neural network model by using the unidentified noise signal data in the processed noise signal data to obtain the verified convolutional neural network model. The convolutional neural network CNN is characterized by parallel distributed processing capability, high fault tolerance, intellectualization, self-learning capability and the like, combines processing and storage of information, and has a unique knowledge representation mode and intelligent self-adaptive learning capability. The convolutional neural network is actually a complex network formed by connecting a large number of simple elements with each other, has high nonlinearity, and can perform complex logic operation and nonlinear relation implementation. A neural network is an operational model, which is formed by connecting a large number of nodes (or neurons). Each node represents a particular output function, called an activation function. Every connection between two nodes represents a weighted value, called weight, for the signal passing through the connection, in such a way that the neural network simulates the human memory. The output of the network depends on the structure of the network, the way the network is connected, the weights and the activation functions. The network itself is usually an approximation to some algorithm or function in nature, and may also be an expression of a logic strategy. The construction concept of the neural network is inspired by the operation of the biological neural network. The CNN is realized by combining the knowledge of the biological neural network with a mathematical statistical model and by means of a mathematical statistical tool. On the other hand, in the artificial perception field of artificial intelligence, the neural network can have the human-like decision ability and simple judgment ability by a mathematical statistics method. According to the technical scheme, the CNN is applied to identifying the noise signal of the digital pulse wave, the noise signal of the digital pulse wave is automatically identified through an efficient algorithm, and further guarantee is provided for digital diagnosis and treatment of pulse conditions, so that the CNN becomes a key problem for collecting effective pulse condition data. The technical scheme of the invention has high stability for pulse wave noise signal identification, and obtains very high precision rate and recall rate on different types of noise signals.
Drawings
A more complete understanding of exemplary embodiments of the present invention may be had by reference to the following drawings in which:
fig. 1 is a flow chart of a method for establishing a pulse wave noise signal identification model based on a convolutional neural network according to a preferred embodiment of the present invention;
FIG. 2 is a diagram illustrating a convolution operation according to a preferred embodiment of the present invention; and
fig. 3 is a system diagram for establishing a pulse wave noise signal recognition model based on a convolutional neural network according to a preferred embodiment of the present invention.
Detailed Description
The exemplary embodiments of the present invention will now be described with reference to the accompanying drawings, however, the present invention may be embodied in many different forms and is not limited to the embodiments described herein, which are provided for complete and complete disclosure of the present invention and to fully convey the scope of the present invention to those skilled in the art. The terminology used in the exemplary embodiments illustrated in the accompanying drawings is not intended to be limiting of the invention. In the drawings, the same units/elements are denoted by the same reference numerals.
Unless otherwise defined, terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Further, it will be understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense.
Fig. 1 is a flowchart of a method for establishing a pulse wave noise signal identification model based on a convolutional neural network according to a preferred embodiment of the present invention. The data of the application are based on the radial artery digital signals acquired by the pressure sensor, and the digital pulse wave noise signals caused by various reasons are accurately and efficiently identified through the CNN algorithm, so that guarantee is provided for accurate clinical diagnosis. As shown in fig. 1, a method for establishing a pulse wave noise signal identification model based on a convolutional neural network includes:
the application trains and adjusts the optimization to pulse wave signal classification model, includes:
1.1 making training samples
Manually classifying and labeling 2000 groups of acquired pulse wave signal data, then taking 1500 groups of data as a training set, taking 500 groups of data as a verification set, and making all data into a VOC (volatile organic compound) data set by using a caffe tool;
configuration of network structure and training model
The Caffe profile generally consists of two parts: prototxt and net. They actually correspond to two very key entities in the Caffe system architecture, namely Solver Solver and network structure Net. Prototxt is a model parameter configuration file, and the parameters are continuously adjusted according to an experiment result when a model is trained each time, and the flow of the solver is roughly divided into:
1. designing an object to be optimized, a training network for learning and a testing network for evaluation;
2. tracking new parameters through iterative optimization of forward and backward;
3. periodically evaluating the test network;
4. displaying the states of the model and the solution during the optimization process
We perform detailed configuration description of parameters in solution.
net:"examples/mobilenets/mobilenet_train.prototxt"
The method comprises the steps of setting a deep network model, wherein each model is a net, configuring the net in a special configuration file, and each net is composed of a plurality of layers.
2)test_iter:10
This is to be understood in conjunction with the batch _ size in the test layer. The total number of test samples in the data set is 500, and the data efficiency of one-time execution is low, so that the test data is divided into several batches to be executed, and the number of each batch is the batch _ size. Assuming we set the batch _ size to 50, 10 iterations are required to complete 500 data. Therefore, the test _ iter is set to 10, and the whole data is executed once, which is called an epoch.
3)test_interval:100
The test is performed at test intervals, i.e., 100 times per training.
4)base_lr:0.01
lr_policy:"inv"
gamma:0.0001
power:0.75
These four parameters can be understood together for the setting of the learning rate. As long as the gradient descent method is used for solving the optimization, a learning rate, also called a step length, exists. The base _ lr is used for setting a basic learning rate, and in an iteration process, the basic learning rate can be adjusted, which is an adjustment strategy, and is set by lr _ policy.
5)momentum:0.9
Weight of last gradient update
6)type:SGD
Optimization algorithm selection
7)weight_decay:0.0005
Weighting decay term, a parameter that prevents overfitting
8)display:50
Every 50 times of training, displayed on the screen. If set to 0, no display is made
9)max_iter:1000
The maximum number of iterations. This number setting is too small, resulting in no convergence and very low accuracy; too large a setting can result in shock and waste of time.
10)snapshot:100
snapshot_prefix:"examples/mobilenets/mobilenet"
And (6) snapshotting. And storing the trained model and the state of the solution, wherein the snapshot is used for storing after setting the number of times of training, and the default is 0 and is not stored. The snapshot _ prefix sets the save path.
11)solver_mode:CPU
An operating mode is set. The default is GPU, if there is no GPU, the CPU needs to be modified, otherwise, errors can occur.
In the application, the learning rate can be initially considered to be 0.1 or 0.01, a fixed step size mode is adopted, and the interval of tens of thousands of turns is reduced by one order of magnitude; weight-decay initialization may be a little smaller, such as 0.00005; the Batchsize is set between 50 and 200, if the size is too small, SGD is too random, if the size is too large, video memory is considered, and if the size is too large, local minimum is easy to fall into. The main concern in the training process is test loss and train loss, if both the test loss and the train loss are reduced, the training process is in an ideal state; if the train loss is reduced and the test loss is unchanged or increased, an overfitting phenomenon occurs, and weight _ decay is tried to be increased; the train loss is unchanged or increased, the test loss is decreased, the phenomenon does not belong to the normal phenomenon, and the problem of the configuration is likely to occur; both train loss and test loss are unchanged, with two possibilities: 1) approaching a minimum point, considering reducing the learning rate 2) to not converge, and if the data has no problem, relaxing the over-fitting constraint; both the train loss and the test loss rise, configuration errors or data problems occur.
Prototxt, which includes the basic structure and type of the network and the parameter configuration of each network layer. The application is not limited to the type of network used.
And preprocessing the pulse wave signal data based on the mean value and the fixed standard deviation.
Preferably, in step 101: and preprocessing the pulse wave noise signal data, reducing the discrete degree of the pulse wave signal data, and acquiring the processed noise signal data.
Preferably, the preprocessing the pulse noise signal data includes:
the pulse wave noise signal data is normalized by the following formula:
y＝(x-a)/c
wherein x represents a discretized digitized pulse noise signal;
a ═ avg (x) represents the mean, median, or average calculated from the whole sample of the individual pulse noise signal sequences, or the center of the pulse noise signal predicted by other methods;
c is a constant value, and is a dispersion mean value of x obtained through experimental data statistics or a standard deviation after calibration according to pulse wave noise signal data.
Since the original data is directly input to the input layer of the CNN, a stable model cannot be trained due to the irregular distribution of the data, and the accuracy is affected. The application carries out standardized processing to the digital pulse wave signal, and the formula is as follows:
y＝(x-a)/c
wherein x represents a discretized digitized pulse signal;
a ═ avg (x) denotes the mean of the individual signal sequences, where a is intended for data decentralization, and an alternative method could be median or mean calculated from the whole sample, or other predicted mean;
c is a constant and is the mean deviation of x obtained by experimental data statistics. Here, the constant c is set for signal normalization, and different values may be adopted according to different data. An alternative approach may be to calibrate the standard deviation against the data. The suggested range for c is 10,40 grams.
After normalization, the value range distribution of the data becomes relatively stable from the original uncertainty. Since c is a fixed value and not a standard deviation calculated from the samples. This prevents the noise of small amplitude from becoming close in scale to the normal signal amplitude after normalization, thereby improving the noise discrimination effect.
Preferably, at step 102: inputting the processed noise signal data into a convolutional neural network model, and extracting the characteristics of the processed noise signal data through a plurality of convolutional layers of the convolutional neural network model.
The frame structure of the CNN neural network is relatively complex. Some layers of artificial neural networks need to be referenced. And a network is built on the basis of the layers, so that parameter adjustment can be effectively realized. The neural network layer of the present application is described as follows:
TABLE 1 CNN neural network layer description
The top layer of the network uses convolutional layers in combination with pooling layers, and Droupout layers, to reduce the number of parameters for training and prevent overfitting. The activation function used in the present application is not limited to the ReLU function and may be a sigmod function or a bi-tangent function, but the use of this function is still proposed due to the superior characteristics of the ReLU function.
The training process of the convolutional neural network refers to a back propagation algorithm, and the forward propagation and the backward feedback are divided into two stages:
first stage, forward propagation stage
a) Taking a sample (X, Yp) from the sample set, inputting X into the network;
b) the corresponding actual output Op is calculated.
Second stage, backward feedback stage
a) Calculating the difference between the actual output Op and the corresponding ideal output Yp;
b) the adjustment weight matrix is propagated back in a way that minimizes the error.
1. Feature extraction based on different scale convolution kernels
Convolution (Convolution): the operations performed by the convolutional layer can be considered to be inspired by the concept of local receptive field, and the pooling layer is mainly to reduce the data dimension. Each neuron looks as a filter, which computes local data by sliding a window (receptavefield). In addition, a particularly important feature of convolutional layers is the parameter sharing mechanism, i.e., the weight of each neuron connection data window is fixed.
Classical CNN, such as LeNet, input images first enter a convolutional layer C1, consisting of 6 convolution kernels of 5x5, convolved out into a 28x28 image, and then downsampled to 14x14 (S2). Next, a convolutional layer C3 is further formed, which is composed of 16 convolutional kernels of 5x5, and then downsampled to 5x5 (S4). Defining a discrete two-dimensional convolution formula for CNN convolution operations for image processing:
in the formula, the ordering of the three matrices starts from 0, and the convolution process and the generated result of the matrix A and the data are as follows:
the present application uses a plurality of convolution kernels, each convolution kernel representing a data pattern, and if a data block is convolved with the convolution kernel to a large value, the data block is considered to be very close to the convolution kernel. Here, two-dimensional image convolution, which increases the amount of computation and increases the space, may be substituted, and it is recommended to use one-dimensional linear convolution Conv 1D. The proposed range of values for the parameters of the convolution kernel is 3,5,7,9, 11.
Preferably, in step 103: key information of the processed noise signal data is extracted through a max-pooling layer based on characteristics of the noise signal data. Preferably, extracting key information of the processed noise signal data through the maximum pooling layer includes: and extracting key information of the processed noise signal data through a maximum pooling layer by a maximum value down-sampling method. Preferably, the pooling radius is proportional to the convolution kernel size of the corresponding convolution layer.
The application extracts key information from the largest Pooling layer with variable scale, Pooling (Pooling): CNN simulates feature differentiation through convolution, reduces the magnitude of network parameters through weight sharing and pooling of the convolution, and finally completes tasks such as classification through a traditional neural network. The technique used for pooling is downsampling. In practical applications, the Pooling is divided into maximum value down-sampling (Max-Pooling) and average value down-sampling (Mean-Pooling) according to the down-sampling method. This patent uses a maximum downsampling method. The use of a variable scale pooling radius as a network parameter in combination with Dropout prevents the network from overfitting.
Pooling methods as used herein include, but are not limited to: Max-Pooling, Mean-Pooling, Filter-Pooling. The selection suggestion of the pooling radius is in direct proportion to the size of the convolution kernel of the corresponding convolution layer, and the suggestion range is 2-16.
Preferably, at step 104: and calculating the weight of each neuron of the convolutional neural network model through the dense layer activation function. Preferably, the weights of the neurons of the convolutional neural network model are calculated by a dense layer activation function, wherein the activation function uses a ReLU activation function.
The ReLu function is adopted as a dense layer activation function in the application. The ReLU function has the following properties:
non-linearity: when the activation function is non-linear, the two-layered neural network can approximate substantially all of the function. If the activation function is an identity activation function, then not satisfied; if the MLP uses an identical activation function, the entire network is equivalent to a single-layer neural network.
Micro-property: this property is necessary when the optimization method is based on gradients.
Monotonicity: when the activation function is monotonic, a single layer network can be guaranteed to be a convex function.
Stability: if the initialization of the parameters is random small, the training of the neural network will be very efficient; if this property is not satisfied, an initial value needs to be set.
Output value range: when the output value of the activation function is limited, the optimization method based on the gradient is more stable, and the representation of the characteristic is influenced by the limited weight more obviously.
The SGD obtained using ReLU will converge much faster than sigmoid/tanh. This is because it is linear, and non-bathing. In contrast to sigmoid/tanh, ReLU requires only one threshold to obtain the activation value without computing a large complex heap of operations.
The normalized range of the signal is around 0 for the digitized pulse signal. And the distribution difference of the noise signal and the normal signal around the mean value is obvious. The function substituted here may be a sigmod function or a bi-tangent function.
Preferably, at step 105: and calculating the processed noise signal data through a convolutional neural network model, and outputting the recognition result of the noise signal data.
Preferably, at step 106: and adjusting the weight of each neuron of the convolutional neural network model through the error of the recognition result of the noise signal data, and performing multiple iterative cycle training on the convolutional neural network model.
The application relates to neuron weight training based on momentum stochastic gradient. Among the random gradient descent and the batch gradient descent, the update formula of the parameters is as follows:
W＝W-αdW
b＝b-αdb
where α is the learning rate and dW, db are the partial derivatives of the cost function with respect to w and b the difference between the stochastic gradient descent and the batch gradient descent is only that the input data is mini-batch and all, respectively.
Instead of subtracting α dW and α db directly from the momentum gradient, a vdW and vdb are calculated, where an exponentially weighted average is introduced to link the previous dW and db, each time the gradient is no longer independent, β is a self-settable hyperparameter, typically defaulted to 0.9 (and possibly other values), β represents that the current vdW and vdb are related to the previous 1/(1- β) vdW and vdb, 0.9 is the result of averaging the previous 10 day's vdW and vdb for the current vdW and vdb, the current gradient is no longer just the current data gradient, but the previous gradient with a weight, the weight "momentum" at the previous time is used to influence the current weight adjustment direction and magnitude, the updated learning rate alternative of this patent can be three methods:
step decrease, i.e. multiplying the learning rate by a small decay factor after a certain number of iterations. Typical practices include multiplying the learning rate by 0.5 after 5 iterations (epoch), or multiplying by 0.1 after 20 iterations.
Exponential decay (Exponential decay), the mathematical expression of which can be expressed as: α - α 0e-kt α - α 0e-kt, where α 0 and k are hyper-parameters to be set and t is the number of iterations.
Reciprocal attenuation (1/t decade), the mathematical expression of which can be expressed as α - α 0/(1+ kt) α - α 0/(1+ kt), where α 0 and k are hyper-parameters to be set and t is the number of iterations.
Preferably, in step 107: and verifying the convolutional neural network model by using the unidentified noise signal data in the processed noise signal data to obtain the verified convolutional neural network model.
The data of the application is based on the radial artery digital signals acquired by the pressure sensor, and the digital pulse wave noise signals caused by various reasons are accurately and efficiently identified through the CNN algorithm, so that the guarantee is provided for accurate clinical diagnosis. The method and the device perform convolution on the digitized pulse wave signals through different scale kernels, and extract the multidimensional characteristics including the whole information and the local information. The convolution mode used in the patent is different from the general CNN in convolution kernel, convolution dimension and convolution step size. The mode that the step length of the convolution kernel is changed in proportion to the size can not only ensure that the noise signal cannot be highlighted due to overlapping convolution, but also ensure that the information of the normal signal cannot be mixed with the noise information. Features are extracted through a plurality of convolution layers and a plurality of variable-scale channels, and pooling sampling is performed on the convolution extracted features so as to avoid overfitting. The pulse wave signal data preprocessing method based on the mean value and the fixed standard deviation is adopted. Since the original digitized pulse signal is not uniform in scale, especially the data range and the range of the feature of the noise signal are too dispersed. The raw data is not conducive to the training of CNN neural networks. The method and the device perform pulse wave signal data preprocessing on the original digital pulse wave signals based on the mean value and the fixed standard deviation, so that the network can keep higher accuracy on noise under various conditions.
Experiments prove that the embodiment of the application obtains good performance in the aspects of identification accuracy and stability. The algorithm processing flow provided by the application is widely applicable to recognition of various high-frequency and low-frequency noise signals such as myoelectricity and jitter, the model has the excellent characteristics of convenience in parameter training, small calculated amount, convenience in deployment and the like, and the problem of low recognition accuracy of the current mainstream digital pulse diagnosis instrument on the noise signals is solved. The embodiment of the application has high identification stability: the method has the advantages that very high precision rate and recall rate are obtained on different types of noise signals; the algorithm parameter structure is simple: and automatically generating a parameter system based on the data and the label.
Fig. 3 is a system diagram for establishing a pulse wave noise signal recognition model based on a convolutional neural network according to a preferred embodiment of the present invention. As shown in fig. 3, a system for establishing a pulse wave noise signal identification model based on a convolutional neural network includes:
the initialization unit 301 is configured to pre-process the pulse wave noise signal data, reduce the dispersion degree of the pulse wave noise signal data, and obtain the processed noise signal data.
Preferably, the initialization unit 301 is configured to pre-process the pulse noise signal data, and further configured to:
the pulse wave noise signal data is normalized by the following formula:
y＝(x-a)/c
wherein x represents a discretized digitized pulse noise signal;
a ═ avg (x) represents the mean, median, or average calculated from the whole sample of the individual pulse noise signal sequences, or the center of the pulse noise signal predicted by other methods;
c is a constant value, and is a dispersion mean value of x obtained through experimental data statistics or a standard deviation after calibration according to pulse wave noise signal data.
A first constructing unit 302, configured to input the processed noise signal data to a convolutional neural network model, and extract features of the processed noise signal data through a plurality of convolutional layers of the convolutional neural network model.
A second construction unit 303, configured to extract key information of the processed noise signal data through the maximum pooling layer based on features of the noise signal data. Preferably, the second construction unit 303 is configured to extract key information of the processed noise signal data through a maximum pooling layer, and includes:
and extracting key information of the processed noise signal data through a maximum pooling layer by a maximum value down-sampling method. Preferably, the pooling radius is proportional to the convolution kernel size of the corresponding convolution layer.
And a third constructing unit 304, configured to calculate a weight of each neuron of the convolutional neural network model through a dense layer activation function. Preferably, the weights of the neurons of the convolutional neural network model are calculated by a dense layer activation function, wherein the activation function uses a ReLU activation function.
An output unit 305, configured to calculate the processed noise signal data through a convolutional neural network model, and output an identification result of the noise signal data.
And the training unit 306 is configured to adjust the weight of each neuron of the convolutional neural network model according to the error of the recognition result of the noise signal data, and perform multiple iterative loop training on the convolutional neural network model.
An obtaining unit 307, configured to verify the convolutional neural network model by using the unidentified noise signal data in the processed noise signal data, and obtain a verified convolutional neural network model.
The system 300 for establishing a pulse wave noise signal identification model based on a convolutional neural network according to a preferred embodiment of the present invention corresponds to the method 100 for establishing a pulse wave noise signal identification model based on a convolutional neural network according to another preferred embodiment of the present invention, and will not be described herein again.
The invention has been described with reference to a few embodiments. However, other embodiments of the invention than the one disclosed above are equally possible within the scope of the invention, as would be apparent to a person skilled in the art from the appended patent claims.
Generally, all terms used in the claims are to be interpreted according to their ordinary meaning in the technical field, unless explicitly defined otherwise herein. All references to "a/an/the [ device, component, etc ]" are to be interpreted openly as referring to at least one instance of said device, component, etc., unless explicitly stated otherwise. The steps of any method disclosed herein do not have to be performed in the exact order disclosed, unless explicitly stated.
Claims (10)
1. A method of building a pulse wave noise signal identification model based on a convolutional neural network, the method comprising:
preprocessing pulse wave noise signal data, reducing the discrete degree of the pulse wave signal data, and acquiring the processed noise signal data;
inputting the processed noise signal data into a convolutional neural network model, and extracting the characteristics of the processed noise signal data through a plurality of convolutional layers of the convolutional neural network model;
extracting key information of the processed noise signal data through a maximum pooling layer based on characteristics of the noise signal data;
calculating the weight of each neuron of the convolutional neural network model through a dense layer activation function;
calculating the processed noise signal data through the convolutional neural network model, and outputting an identification result of the noise signal data;
adjusting the weight of each neuron of the convolutional neural network model according to the error of the recognition result of the noise signal data, and performing multiple iterative cycle training on the convolutional neural network model;
and verifying the convolutional neural network model by using the unrecognized noise signal data in the processed noise signal data to obtain the verified convolutional neural network model.
2. The method of claim 1, the pre-processing pulse noise signal data, comprising:
the pulse wave noise signal data is normalized by the following formula:
y＝(x-a)/c
wherein x represents a discretized digitized pulse noise signal;
a ═ avg (x) represents the mean, median, or average calculated from the whole sample of the individual pulse noise signal sequences, or the center of the pulse noise signal predicted by other methods;
c is a constant value, and is a dispersion mean value of x obtained through experimental data statistics or a standard deviation after calibration according to pulse wave noise signal data.
3. The method of claim 1, the extracting key information of the processed noise signal data by max-pooling layer, comprising:
and extracting key information of the processed noise signal data through a maximum value downsampling method through a maximum pooling layer.
4. The method of claim 3, pooling radius being proportional to a convolution kernel size of the corresponding convolutional layer.
5. The method of claim 1, wherein the weights for the neurons of the convolutional neural network model are calculated by a dense layer activation function, wherein the activation function uses a ReLU activation function.
6. A system for building a pulse wave noise signal identification model based on a convolutional neural network, the system comprising:
the device comprises an initial unit, a data processing unit and a data processing unit, wherein the initial unit is used for preprocessing pulse wave noise signal data, reducing the discrete degree of the pulse wave signal data and acquiring the processed noise signal data;
a first construction unit, configured to input the processed noise signal data to a convolutional neural network model, and extract features of the processed noise signal data through a plurality of convolutional layers of the convolutional neural network model;
a second construction unit, configured to extract, based on features of the noise signal data, key information of the processed noise signal data through a maximum pooling layer;
the third construction unit is used for calculating the weight of each neuron of the convolutional neural network model through a dense layer activation function;
the output unit is used for calculating the processed noise signal data through the convolutional neural network model and outputting the identification result of the noise signal data;
the training unit is used for adjusting the weight of each neuron of the convolutional neural network model according to the error of the recognition result of the noise signal data and carrying out multiple iterative cycle training on the convolutional neural network model;
and the obtaining unit is used for verifying the convolutional neural network model by using the unidentified noise signal data in the processed noise signal data to obtain the verified convolutional neural network model.
7. The system of claim 6, the initialization unit to preprocess pulse noise signal data, further to:
the pulse wave noise signal data is normalized by the following formula:
y＝(x-a)/c
wherein x represents a discretized digitized pulse noise signal;
a ═ avg (x) represents the mean, median, or average calculated from the whole sample of the individual pulse noise signal sequences, or the center of the pulse noise signal predicted by other methods;
c is a constant value, and is a dispersion mean value of x obtained through experimental data statistics or a standard deviation after calibration according to pulse wave noise signal data.
8. The system of claim 6, the second construction unit for extracting key information of the processed noise signal data through a maximum pooling layer, comprising:
and extracting key information of the processed noise signal data through a maximum value downsampling method through a maximum pooling layer.
9. The system of claim 8, a pooling radius proportional to a convolution kernel size of the corresponding convolutional layer.
10. The method of claim 1, wherein the weights for the neurons of the convolutional neural network model are calculated by a dense layer activation function, wherein the activation function uses a ReLU activation function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911030823.8A CN110956197A (en) | 2019-10-28 | 2019-10-28 | Method and system for establishing pulse wave noise signal identification model based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911030823.8A CN110956197A (en) | 2019-10-28 | 2019-10-28 | Method and system for establishing pulse wave noise signal identification model based on convolutional neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110956197A true CN110956197A (en) | 2020-04-03 |
Family
ID=69975758
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911030823.8A Pending CN110956197A (en) | 2019-10-28 | 2019-10-28 | Method and system for establishing pulse wave noise signal identification model based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110956197A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111582236A (en) * | 2020-05-27 | 2020-08-25 | 哈尔滨工程大学 | LPI radar signal classification method based on dense convolutional neural network |
| CN111721401A (en) * | 2020-06-17 | 2020-09-29 | 广州广电计量检测股份有限公司 | Low-frequency noise analysis system and method |
| CN112472048A (en) * | 2021-01-15 | 2021-03-12 | 郑州大学 | Neural network structure for pulse condition recognition of cardiovascular disease patient |
| CN113762349A (en) * | 2021-08-11 | 2021-12-07 | 同济大学 | Lightweight aliasing dense network classification method and system for marine organisms |
| CN114469017A (en) * | 2022-01-27 | 2022-05-13 | 上海国民集团健康科技有限公司 | Pulse condition data detection method and system |
| CN117224092A (en) * | 2023-11-16 | 2023-12-15 | 常熟理工学院 | Photoelectric volume pulse wave interference band real-time detection method and system based on decision tree |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102824166A (en) * | 2012-07-05 | 2012-12-19 | 华东师范大学 | Device for sorting treatment of pulse wave |
| CN104706337A (en) * | 2015-02-11 | 2015-06-17 | 华东师范大学 | Automatic pulse wave crest and trough detection method |
| CN106137167A (en) * | 2016-07-21 | 2016-11-23 | 浙江师范大学 | A kind of motion artifacts detection method based on photoplethysmographic signal |
| WO2017152990A1 (en) * | 2016-03-11 | 2017-09-14 | Telecom Italia S.P.A. | Convolutional neural networks, particularly for image analysis |
| CN107609488A (en) * | 2017-08-21 | 2018-01-19 | 哈尔滨工程大学 | A kind of ship noise method for identifying and classifying based on depth convolutional network |
| CN108234370A (en) * | 2017-12-22 | 2018-06-29 | 西安电子科技大学 | Modulation mode of communication signal recognition methods based on convolutional neural networks |
| CN109157200A (en) * | 2018-10-31 | 2019-01-08 | 安徽华米信息科技有限公司 | noise detecting method and device |
| CN109389058A (en) * | 2018-09-25 | 2019-02-26 | 中国人民解放军海军航空大学 | Sea clutter and noise signal classification method and system |
| CN110037668A (en) * | 2019-04-10 | 2019-07-23 | 常熟理工学院 | The system that pulse signal time-space domain binding model judges age, health status and malignant arrhythmia identification |
| CN110164472A (en) * | 2019-04-19 | 2019-08-23 | 天津大学 | Noise classification method based on convolutional neural networks |
-
2019
- 2019-10-28 CN CN201911030823.8A patent/CN110956197A/en active Pending
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102824166A (en) * | 2012-07-05 | 2012-12-19 | 华东师范大学 | Device for sorting treatment of pulse wave |
| CN104706337A (en) * | 2015-02-11 | 2015-06-17 | 华东师范大学 | Automatic pulse wave crest and trough detection method |
| WO2017152990A1 (en) * | 2016-03-11 | 2017-09-14 | Telecom Italia S.P.A. | Convolutional neural networks, particularly for image analysis |
| CN106137167A (en) * | 2016-07-21 | 2016-11-23 | 浙江师范大学 | A kind of motion artifacts detection method based on photoplethysmographic signal |
| CN107609488A (en) * | 2017-08-21 | 2018-01-19 | 哈尔滨工程大学 | A kind of ship noise method for identifying and classifying based on depth convolutional network |
| CN108234370A (en) * | 2017-12-22 | 2018-06-29 | 西安电子科技大学 | Modulation mode of communication signal recognition methods based on convolutional neural networks |
| CN109389058A (en) * | 2018-09-25 | 2019-02-26 | 中国人民解放军海军航空大学 | Sea clutter and noise signal classification method and system |
| CN109157200A (en) * | 2018-10-31 | 2019-01-08 | 安徽华米信息科技有限公司 | noise detecting method and device |
| CN110037668A (en) * | 2019-04-10 | 2019-07-23 | 常熟理工学院 | The system that pulse signal time-space domain binding model judges age, health status and malignant arrhythmia identification |
| CN110164472A (en) * | 2019-04-19 | 2019-08-23 | 天津大学 | Noise classification method based on convolutional neural networks |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111582236A (en) * | 2020-05-27 | 2020-08-25 | 哈尔滨工程大学 | LPI radar signal classification method based on dense convolutional neural network |
| CN111582236B (en) * | 2020-05-27 | 2022-08-02 | 哈尔滨工程大学 | LPI radar signal classification method based on dense convolutional neural network |
| CN111721401A (en) * | 2020-06-17 | 2020-09-29 | 广州广电计量检测股份有限公司 | Low-frequency noise analysis system and method |
| CN111721401B (en) * | 2020-06-17 | 2022-03-08 | 广州广电计量检测股份有限公司 | Low-frequency noise analysis system and method |
| CN112472048A (en) * | 2021-01-15 | 2021-03-12 | 郑州大学 | Neural network structure for pulse condition recognition of cardiovascular disease patient |
| CN112472048B (en) * | 2021-01-15 | 2023-10-24 | 郑州大学 | Method for realizing neural network for identifying pulse condition of cardiovascular disease patient |
| CN113762349A (en) * | 2021-08-11 | 2021-12-07 | 同济大学 | Lightweight aliasing dense network classification method and system for marine organisms |
| CN113762349B (en) * | 2021-08-11 | 2024-03-29 | 同济大学 | Marine organism-oriented lightweight aliasing dense network classification method and system |
| CN114469017A (en) * | 2022-01-27 | 2022-05-13 | 上海国民集团健康科技有限公司 | Pulse condition data detection method and system |
| CN117224092A (en) * | 2023-11-16 | 2023-12-15 | 常熟理工学院 | Photoelectric volume pulse wave interference band real-time detection method and system based on decision tree |
| CN117224092B (en) * | 2023-11-16 | 2024-02-09 | 常熟理工学院 | Photoelectric volume pulse wave interference band real-time detection method and system based on decision tree |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110956197A (en) | Method and system for establishing pulse wave noise signal identification model based on convolutional neural network | |
| CN105117611B (en) | Based on the determining method and system of the TCM tongue diagnosis model of convolutional Neural metanetwork | |
| Castellano et al. | Variable selection using neural-network models | |
| Kong et al. | Hexpo: A vanishing-proof activation function | |
| EP1534122B1 (en) | Medical decision support systems utilizing gene expression and clinical information and method for use | |
| CN112529146B (en) | Neural network model training method and device | |
| Cheng et al. | Classification of long sequential data using circular dilated convolutional neural networks | |
| CN109907733A (en) | A kind of ECG signal analysis method towards abnormal heart rhythms classification | |
| Arulkumar et al. | A novel usage of artificial intelligence and internet of things in remote‐based healthcare applications | |
| Fakih et al. | An efficient prediction of diabetes using artificial neural networks | |
| Patil et al. | Designing a model to detect diabetes using machine learning | |
| Alqudah et al. | Reduced number of parameters for predicting post-stroke activities of daily living using machine learning algorithms on initiating rehabilitation | |
| Zebardast et al. | A new generalized regression artificial neural networks approach for diagnosing heart disease | |
| Subbarao et al. | Detection of Retinal Degeneration via High-Resolution Fundus Images using Deep Neural Networks | |
| Furtado et al. | Human Visual System vs Convolution Neural Networks in food recognition task: An empirical comparison | |
| Zhang et al. | Pulsar candidate recognition with deep learning | |
| CN111160412A (en) | Method for classifying medical waste based on neural network | |
| CN114021630B (en) | Ordinal regression problem solving method for class unbalanced data set | |
| CN116310477A (en) | Image classification method based on S_ReLU pulse neural network | |
| Palaniappan et al. | Using genetic algorithm to identify the discriminatory subset of multi-channel spectral bands for visual response | |
| Mudunuru | Comparison of activation functions in multilayer neural networks for stage classification in breast cancer | |
| Sun et al. | Coupling matrix extraction for microwave filter design using neural networks | |
| Rout | A Novel Grid Ann for Prediction of Heart Disease | |
| Mohanty et al. | An optimized KELM approach for the diagnosis of COVID-19 from 2D-SSA reconstructed CXR Images | |
| Islam et al. | Retinal Vessels Detection Using Convolutional Neural Networks in Fundus Images |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200403 |