Journal of Donghua University ( Eng. Ed. ) Vol. 40, No.

2 (2023) 185

DOI: 10. 19884 / j. 1672-5220. 202112002

Fine-Grained Sleep Apnea Detection Method from Multichannel Ballistocardiogram

Using Convolution Neural Network

HUANG Yongfeng( 黄永锋) 1∗ , HUANG Qihong( 黄琦洪) 1 , SUN Chenxi( 孙晨汐) 1 , YANG Shuchen( 杨树
臣) 2 , ZHANG Zhiming( 张智明) 2
1 School of Computer Science and Technology, Donghua University, Shanghai 201620, China
2 Shanghai Yueyang Medtech Co. , Ltd. , Shanghai 201203, China

Abstract: Sleep apnea is a common health condition that can increasing social pressure and various psychological
affect numerous aspects of life and may cause a lot of health diseases. Some researchers believe that anxiety and
problems especially in the middle-aged and elderly depression are gradually becoming the main or even all
population. Polysomnography ( PSG ) , as the gold inducement of sleep apnea [2] . Due to the universality of
standard, is an expensive and inconvenient way to diagnose
above-mentioned problems, sleep apnea has become a
sleep apnea. However, ballistocardiogram can be collected
nonnegligible disease in modern society.
by devices embedded in the surrounding environment,
enabling inperceptible sleep apnea detection. Moreover, to Since sleep apnea requires monitoring of sleep
obtain the fine-grained apnea fragments, a multistage sleep conditions throughout the night, polysomnography
apnea detection model has been proposed. This model firstly ( PSG) has been proposed in clinical medicine for the
uses an improved convolution neural network ( CNN) model diagnosis of apnea. PSG, performed in sleep laboratory,
to coarsely identify apnea events and then a U-Net based is considered the gold standard for sleep apnea diagnosis
model is applied to finely segment apnea fragments. In the by recording various physiological signals, including
experiment, sleep data of 11 patients with apnea for about nasal inspiratory airflow, electroencephalogram ( EEG )
70 h have been collected, including BCG data derived from and electrocardiogram ( ECG ) , allowing researchers to
18 piezoelectric polyvinylidene fluoride ( PVDF ) sensors achieve accurate results. However, it is uncomfortable
embedded in the mattress and PSG data collected
and may cause psychological burden of patients.
synchronously. The results show the accuracy of the
Furthermore, for a severe apnea patient, monitoring at
classification model as good as 95. 7% with 0. 818 dice
coefficient of the segmentation model, which indicates that home and inspection in the hospital should be given equal
the proposed model can almost match the performance of attention. Ballistocardiogram ( BCG ) , as a universal
PSG in detecting apnea. medical technique, records the change of pressure caused
Key words: sleep apnea; ballistocardiogram; convolution by cardiovascular and respiration in a cost-effective and
neural network ( CNN) ; deep learning noninvasive way [3-4] . Since the signal can be measured by
CLC number: TP399 Document code: A sensors embedded in the ambient environments without
Article ID: 1672-5220(2023)02-0185-08 the need for medical staff, it can be a potential method
for home sleep apnea monitoring.
Open Science Identity However, most of the past studies [5-9] on sleep apnea
( OSID) have focused only on classification tasks using traditional
methods such as template matching, heart rate detection
Introduction paired with machine learning and deep learning methods.
Specifically, the classification results can only identify
Sleep apnea, as an extremely common sleep disorder whether or what kind of sleep apnea occurrs in a period.
in modern society, may lead to shortness of breath and In fact, more detailed information such as the frequency
affect our sleep quality. Studies have shown that repeated of sleep apnea and the approximate period of sleep apnea
episodes of apnea will lead to nocturnal hypoxia, which is of equally significance. Hence, past studies fail to
in turn can cause many serious complications such as provide meaningful information for clinical diagnosis in
hypertension, diabetes, cardiovascular and sleep apnea, for example, apnea-hypopnea index
cerebrovascular diseases, and even sudden death at ( AHI) . So, we propose a fine-grained multistage sleep
night [1] . Moreover, about 10% of population has apnea detection model. This model first identifies apnea
suffered a lot from sleep apnea, mainly owing to the events using convolutional neural network ( CNN) , and

Received date: 2021-12-01

∗Correspondence should be addressed to HUANG Yongfeng, email: 741239226@ qq. com
Citation: HUANG Y F, HUANG Q H, SUN C X, et al. Fine-grained sleep apnea detection method from multichannel ballistocardiogram using
convolution neural network[ J] . Journal of Donghua University( English Edition) , 2023, 40(2) : 185-192.
 186 HUANG Yongfeng, HUANG Qihong, SUN Chenxi, et al.

then uses a U-Net based model to segment the time period roughly equivalent to the head and chest of the human
during which the apnea occurs. body.
Furthermore, a high accuracy with single channel In this study, we collected BCG data from 11 apnea
BCG signal was achieved in Ref. [ 10 ] . However, a patients. Data acquisition time was approximately 7 h of
single sensor system may cause measurement problems sleep throughout the night for each subject. The BCG
such as low coverage and non-continuity since BCG sampling frequency is 50 Hz. To improve the
can only be detected when the human body produces generalization ability of the model, we also added a
pressure on the mechanical sensor. To tackle this normal human BCG data collected in the same way.
problem, we propose that, due to the noninvasive For comparison, the PSG data of the above 11
characteristic, multiple-sensor systems can be used to patients were also recorded and the segments where apnea
collect BCG signal at the same time and make full use occurred were labeled by a professional sleep physician.
of BCG signal through channel fusion to improve the Professional sleep specialists determined if apnea occurred
accuracy of detection. based on the nasal airflow and thoracoabdominal
The main contributions are the following two points. breathing conditions collected by the PSG. Central apnea
Firstly, we propose a more fine-grained multistage sleep is characterized by the absence of nasal airflow and
apnea detection model that can reach the sample point thoracoabdominal breathing, while obstructive apnea is
level of resolution, which is at a higher resolution characterized by undetectable nasal airflow and normal
compared with state-of-the-art methods. Secondly, thoracoabdominal breathing. Sleep physicians use the
different multichannel data fusion methods are compared, above methods to determine whether apnea occurs and
and squeeze-and-excitation net ( SENet) shows the best how long it has lasted.
results under the contrast experiment of different Then the Butterworth digital filter is applied to
parameters. remove high-frequency noise and the cut-off frequency of
3 dB filter is 10 Hz and the order is 7. The Z-score
1 Methods normalization method is used to normalize the 18 channel
signals to accelerate the convergence speed of the model.
Figure 1 shows the global architecture of the Moreover, we use the sliding window method to organize
proposed U-Breath model. The model will settle the the data. Specifically, Fig. 2 shows the distribution of the
multichannel BCG signal fusion problem firstly, and then duration of apnea patients in all samples. It can be seen
determine whether sleep apnea events happens or not. If that the duration of most of the samples is concentrated in
it happens, the fine-grained segmentation is further the time range from 10 to 60, and the duration of the
performed and then the marked apnea segment will be longest apnea sample is around 120. Therefore, the
output. length of each window is 90, and to expand the sample
data set, we slide the window in 10-second steps. Within
the 90 second window, the data that contain apnea
fragments longer than 10 are regarded as positive
samples, while the rest are negative samples.

Fig. 2 Frequency histogram of all apnea samples

Fig. 1 Global model architecture

1. 2 Channel fusion methods
After data preprocessing, 18 channel BCG signals
1. 1 Dataset and preprocessing are sent to the channel fusion layer for gating or fusion.
The BCG data acquisition equipment is composed We use LR, average and SENet [10] to process
of 18 PVDF array sensors embedded in the mattress. multichannel BCG signals, and the target is to output
Eighteen sensors are placed horizontally in two rows, single channel BCG signal as the input data of subsequent
and the position of the sensors in the mattress is network. The architecture of SENet is illustrated in
 Fine-Grained Sleep Apnea Detection Method from Multichannel Ballistocardiogram Using
Convolution Neural Network 187

Fig. 3, in which the sign 18 × 4 500 means that the size of parameter. SENet consists of three parts: squeeze,
the input signal consists of 18 channels and 4 500 one- excitation and scale. Its main purpose is to make the
dimensional sample points; the signs 18 × 1, c@ 1 × 1 and model pay more attention to the channel features with the
1 × 4 500 represent the output signal size after the current largest amount of information and suppress those
layer processing of the model, where c is a variable unimportant channel features.

Fig. 3 SENet model architecture

1. 3 Classification methods apnea events. As shown in Fig. 4, the classification

Classification network architecture, inspired by network is composed of six one-dimensional ( 1D )
the work of Banluesombatkul et al.[ 11] , is illustrated in convolution layers and one full connection layer.
Fig. 4. In Fig. 4, the signs 50 @ 20 × 1, 30 @ 24 × 1, Specifically, each convolution layer contains
24@ 10 × 1 and 12 @ 10 × 1 represent the channels and convolution, activation, batch normalization and
size of filter. The sign 192 × 1 is the full connection pooling operations. In addition to the first convolution
block size. The sign 1 × 4 500 is the input signal size. layer, the dropout block is added to solve overfitting
The goal of classification network is to determine problem. And the full connection layer is made up of
whether BCG data over a window of time contain sleep full connection and activation blocks.

Fig. 4 Classification model architecture

1. 4 Segmentation methods symbols in Fig. 5 are as follows. The numbers 1, 16,

As shown in Fig. 5, we use a U-Net [12-13] based 32, 64, 128 and 256 to the right of the down-arrow and
network to achieve fine-grained segmentation task. U-Net to the left of the up-arrow represent the size of the filter.
has demonstrated impressive performance on medical The numbers 2, 4, 6 and 8 to the left of the down-arrow
image segmentation as well as time series segmentation. and to the right of the up-arrow represent the size of
Within the U-Net network, it has two main parts, namely maximum pooling kernel. And 128 + 256, 128 + 64, 64 +
an encoder and a decoder, also known as a down- 32 and 32 + 16 represent the output dimension after
sampling part and a up-sampling part. There are four concatenation. Besides, 1 × 4 500 represents the
convolution layers in each part. The meanings of the dimension of input data and output data.
 188 HUANG Yongfeng, HUANG Qihong, SUN Chenxi, et al.

Fig. 5 Segmentation model architecture

input. Alivar et al. [14] found that the channel with better
2 Experiments signal quality should be normal distribution. Therefore,
we choose the data that best fit the normal distribution as
2. 1 Classification the input of the classification network in our experiments.
To verify the effectiveness of the classification Figure 6 shows the data distribution of the 18 sensors for
network, we manually select one channel signal as the one subject we randomly selected.

Fig. 6 Data distribution of BCG signals of a patient throughout the night

 Fine-Grained Sleep Apnea Detection Method from Multichannel Ballistocardiogram Using
Convolution Neural Network 189

For validating the performance of the proposed The total number of samples is 23 573, including
model, we trained some different models for comparison. 10 541 positive samples and 13 032 negative samples.
The first one is the U-Breath model proposed in this We extracted the training set, the validation set and the
paper. Next, we reproduced the model in recent test set from the BCG data of 11 subjects in the
study [15] , where they proposed a model for apnea event proportion of 60%, 20% and 20%, respectively, and
detection based on an SVM implementation. Then, a combined them into the final training set, the final
bidirectional long short-term memory ( Bi-LSTM ) validation set and the final test set. The number of
based [16] model and a TCN based [17] model in the samples of classification network is shown in Table 1.
study [18] were also used for comparison. Note that all the
above-mentioned models use the same data input format. Table 1 Sample number of each dataset in classification network
In the U-Breath model experiment, we referred to its Type Train Validation Test Total
original study [13] to set our experimental parameters. Positive 6 325 2 108 2 108 10 541
Specifically, the dropout layer probability is 0. 2. That Negative 7 820 2 606 2 606 13 032
is, 20% of the neuron output value is set to 0
Total 14 145 4 714 4 714 23 573
randomly [19] . Then we use a mini-batch method for
training [20] , and gradient descent method with an Adam
After training 200 epochs, the U-Breath model
optimizer is performed on only 256 samples at a time.
converge. Figure 7( a) shows the train loss and validation
The learning rate is 0. 001, and the total number of
loss of the U-Breath model. Meanwhile, the training of
training epochs is 200. And we use cross entropy as the
the comparison model was also completed according to
loss function, which can measure the difference between
the respective parameter settings of the comparison
two data distributions. The calculation formula of cross
model. Table 2 shows the evaluation metrics of all
entropy loss l c is shown as
classification models, where the experimental results of
l c = - [ ylg y^ + (1 - y) lg(1 - y^ ) ] , (1) the SVM model are averaged after 5 k-fold cross
validation. Figure 8 ( a ) shows the receiver operating
where y represents the real value and y^ represents the characteristic ( ROC ) curve of all models and the
predicted value. For the other comparison models, we respective area under curve A ROC , both of which indicates
referred to the parameter settings in their respective that the U-Breath model outperforms other models and
articles to train the models. achieves elegant results.

Fig. 7 U-Breath model train loss and validation loss: ( a) classification result; ( b) segmentation result

Fig. 8 Evaluation metrics for different models: (a) ROC curves for different classification models; (b) precision-recall (PR) curve for U-Breath model
 190 HUANG Yongfeng, HUANG Qihong, SUN Chenxi, et al.

Table 2 Evaluation metrics for different models classification network, we keep the original parameters of
Model Accuracy Precision Recall F-score the classification network unchanged and only feed the
single channel signal into the classification network after
SVM 0. 636 0. 721 0. 656 0. 687 the channel selection or fusion process for validation. For
SENet, in the experiments, we also verify the
Bi-LSTM 0. 670 0. 657 0. 635 0. 646
performance of the network with different squeeze layers
TCN 0. 653 0. 581 0. 797 0. 687 c ( SE-c ) . According to Ref. [ 10 ] , we verify the
network performance when c takes five parameters of 1,
U-Breath 0. 957 0. 958 0. 958 0. 958
2, 4, 8 and 16.
Tables 3 and 4 show the experimental results of the
2. 2 Segmentation U-Breath model after adding different channel fusion
In the segmentation model, we refer to U-Net to set layers. Here, for the convenience of comparison, the
our experimental parameters and the mini-batch method is experimental results of the manual channel selection
also used to gradient down 64 samples each time. Then approach are also added in Tables 3 and 4. Similarly, we
we use the Adam optimizer to update the gradients. The also combine the channel fusion layer and segmentation
learning rate is 0. 001, and the total number of training network for the experiment, and the experimental results
epochs is 200. In addition, dice coefficient, a widely are shown in Tables 5 and 6.
used loss function in image segmentation, can evaluate
the similarity between the prediction set and the real set. Table 3 Evaluation metrics for the combination of
The calculation formula of the dice coefficient and the classification networks and different channel fusion
dice doss are given by methods

2 Mp ∩ Mt Method Train loss Accuracy Precision Recall F-score

cd = , (2) Manual 0. 326 0. 957 0. 958 0. 958 0. 958
Mp + Mt
LR 0. 423 0. 857 0. 836 0. 844 0. 840
ld = 1 - cd , (3) Average 0. 326 0. 961 0. 957 0. 953 0. 955
where M p is the prediction matrix, M t is the real matrix, c d
represents the dice coefficient, and l d represents the dice Table 4 Evaluation metrics for the combination of
classification networks and SENet under different
loss. If the similarity between two sets is higher, the dice
squeeze level
coefficient is closer to 1, while the dice loss is closer to
0. Level Train loss Accuracy Precision Recall F-score
In the previous classifier, there are 10 541 positive SE-1 0. 327 0. 958 0. 951 0. 955 0. 953
samples, that is, samples with sleep apnea. These SE-2 0. 323 0. 968 0. 957 0. 970 0. 963
samples are sent to the segmentation network, and the SE-4 0. 320 0. 981 0. 973 0. 982 0. 978
sample data are divided into the training set, the SE-8 0. 316 0. 987 0. 985 0. 986 0. 986
validation set and the test set according to the same
SE-16 0. 318 0. 975 0. 968 0. 978 0. 973
division method as the classification network is used.
After training 200 epochs, the model converges, and
Table 5 Evaluation metrics for the combination of
Fig. 7 ( b) shows that there is no over fitting or under
segmentation networks and different channel
fitting happened in the verification set and the test set.
fusion methods
The average dice coefficient of the test set is 0. 818,
which shows that the network achieves good results in the Method Train loss Validation dice
task of BCG signal apnea segmentation. Figure 8 ( b ) Manual 0. 174 0. 819
shows the PR curve of the segmentation model, which is LR 0. 295 0. 702
used for performance evaluation in scenarios with Average 0. 114 0. 854
unbalanced samples. Finally, the U-Breath segmentation
model proposed in this paper achieves an average
precision score of 0. 86. Table 6 Evaluation metrics for the combination of
2. 3 Channel fusion segmentation networks and SENet under different
In the channel fusion experiments, we evaluate the squeeze level
performance of various channel fusion methods by Level Train loss Validation dice
combining the channel fusion methods with the SE-1 0. 105 0. 874
classification and segmentation networks described above. SE-2 0. 091 0. 882
In this experiment, three channel fusion methods, namely
SE-4 0. 096 0. 887
LR, average and SENet, are compared with the manual
selection methods. SE-8 0. 094 0. 887
In the combination experiments with the SE-16 0. 099 0. 879
 Fine-Grained Sleep Apnea Detection Method from Multichannel Ballistocardiogram Using
Convolution Neural Network 191

value for the calculation of AHI and other clinical

3 Discussion indicators.

From the experimental results and evaluation metrics

of the classification network, there are obvious
differences between the various methods. From a macro
point of view, the U-Breath model proposed in this paper
is better than other models on the whole, which indicates
that the U-Breath model shows excellent classification
ability that can accurately judge whether apnea happens or
not over a window of time. However, the above
approach using manual channel selection requires manual
comparison of the data distribution of each channel,
which is almost difficult to achieve for home monitoring
scenarios. Therefore, although this manual intervention
method of channel selection is still accurate, it does not
Fig. 9 Comparison of real sleep apnea segments labeled by PSG and
overcome the shortcomings of traditional PSG and ECG
predicted apnea segments labeled by U-Breath model
methods and lacks practical value.
As shown in Tables 3 and 4, after adding the
automatic channel fusion layer, for the U-Breath network 4 Conclusions
proposed in this paper, the channel fusion layer using
both the average method and SENet has improved the This paper proposes a new model based on
evaluation metrics compared to the manual channel convolution neural network, which can well achieve the
selection method. This may be due to the fact that the detection of sleep apnea events and the extraction of the
features of multichannel data are more fully utilized specific time and duration of the occurrence of sleep
compared to those using only one channel data. Somatic apnea. Meanwhile, based on the problem that BCG
motion is an important marker for the occurrence of sleep signals are susceptible to interference, this paper uses
apnea, while manual channel selection is difficult to multichannel BCG data and compares several channel
utilize somatic motion features because it tends to use fusion methods and finds that the SENet-based method
data satisfying a normal distribution; whereas for SENet, can make full use of data from multiple channels and thus
it can adjust the amount of contribution of each channel, improve the performance of the whole model.
so it can utilize somatic motion features and thus improve However, we also notice that it may take longer
the accuracy of sleep apnea detection. SENet which uses time to train the model and detect sleep apnea by using
eight intermediate layers performs the best, and it has the the multistage model, and the reasons may include the
best value for all evaluation metrics. redundant calculation caused by the lack of correlation
Figure 9 shows the comparison between the real between stages. Therefore, we will focus on how to
sleep apnea segments labeled by PSG and the predicted integrate the multistage model into the one-stage model or
sleep apnea segments labeled by the proposed U-Breath how to share the structure and parameters between stages
model. To be specific, the blue line represents the BCG to lighten our model in the future work.
input signal, the green shadow part is the predicted
value, the pink shadow part is the real value given by References
PSG and the overlapping part of the two colors is the
accurate prediction part. It can be seen that the difference [1] SHAHAR E, WHITNEY C W, REDLINEE S,
between the real value and the predicted value is small, et al. Sleep-disordered breathing and
which indicates that the segmentation network can cardiovascular disease [ J] . American Journal of
roughly locate the time segment of sleep apnea, but there Respiratory and Critical Care Medicine, 2001,
will be some deviation in the beginning and ending parts. 163(1) : 19-25.
Compared with the problem of image segmentation, it is [2] MACEY P M, WOO M A, KUMAR R, et al.
difficult to give accurate edge of segmentation in time Relationship between obstructive sleep apnea
series, so it is also difficult to improve the segmentation severity and sleep, depression and anxiety
results. By visualizing test samples, the segmentation symptoms in newly-diagnosed patients[ J] . PLoS
network has the ability to locate the general occurrence One, 2010, 5(4) : e10211.
area of sleep apnea, and the number of sleep apnea [3] PINHEIRO E, POSTOLACHE O, GIRÃO P.
occurrence and the general duration of sleep apnea can be Theory and developments in an unobtrusive
obtained from the predicted results. These indicators can cardiovascular system representation:
provide more fine-grained information for clinical ballistocardiography [ J ] . Open Biomedical
diagnosis of sleep apnea and have a certain reference Engineering Journal, 2010, 4(1) : 201-216.
 192 HUANG Yongfeng, HUANG Qihong, SUN Chenxi, et al.

[4] SADEK I, BISWAS J, ABDULRAZAK B. segmentation [ C ] / / Medical Image Computing

Ballistocardiogram signal processing: a review and Computer-Assisted Intervention-MICCAI
[ J ] . Health Information Science and Systems, 2015, Munich, Germany. Springer International
2019, 7(1) : 10. Publishing, 2015: 234-241.
[5] URTNASAN E, PARK J U, LEE S Y, et al. [13] LIN W, FAN H, HU C X, et al. Improved
Optimal classifier for detection of obstructive medical image segmentation model based on 3D
sleep apnea using a heartbeat signal [ J ] . U-Net [ J ] . Journal of Donghua University
International Journal of Fuzzy Logic & Intelligent ( English Edition) , 2022, 39(4) : 311-316.
Systems, 2017, 17(2) : 76-81. [14] ALIVAR A, CARLSON C, SULIMAN A, et al.
[6] URTNASAN E, PARK J U, JOO E Y, et al. Motion artifact detection and reduction in bed-
Automated detection of obstructive sleep apnea based Ballistocardiogram [ J ] . IEEE Access,
events from a single-lead electrocardiogram using 2019, 7: 13693-13703.
a convolutional neural network [ J ] . Journal of [15] AZIMI H, BOUCHARD M, GOUBRAN R A,
Medical Systems, 2018, 42(6) : 104. et al. Apnea event detection methodology using
[7] MAHMUD T, KHAN I A, MAHMUD T I, pressure sensors [ C ] / / 2019 IEEE International
et al. Sleep apnea event detection from sub-frame Symposium on Medical Measurements and
based feature variation in EEG signal using deep Applications ( MeMeA) , Turkey. IEEE, 2019:
convolutional neural network [ C ] / / 2020 42nd 1-6.
Annual International Conference of the IEEE [16] YANG Q R, XU K Z, ZHENG X H, et al. Tool
Engineering in Medicine and Biology Society health condition recognition method for high-
( EMBC ) , Montreal, Canada. IEEE, 2020: speed milling of titanium alloy based on principal
5580-5583. component analysis ( PCA) and long short-term
[8] CHANG H Y, YEH C Y, LEE C T, et al. A memory ( LSTM ) [ J ] . Journal of Donghua
sleep apnea detection system based on a one- University ( English Edition ) , 2019, 36 ( 4 ) :
dimensional deep convolution neural network 364-368.
model using single-lead electrocardiogram [ J ] . [17] BAI S, KOLTER J Z, KOLTUN V. An
Sensors, 2020, 20: 4157. empirical evaluation of generic convolutional and
[9] SCHÄTZ M, PROCHÁZKA A, KUCHYŇKA J, recurrent networks for sequence modeling [ J /
et al. Sleep apnea detection with OL] . ( 2018-04-19) [ 2021-12-24] http: / / arxiv.
Polysomnography and depth sensors [ J ] . org / abs / 1803. 01271v2.
Sensors, 2020, 20(5) : 1360. [18] AZIMI H, XI P, BOUCHARD M, et al.
[10] HU J, LI S, GANG S, et al. Squeeze-and- Machine learning-based automatic detection of
excitation networks [ J ] . IEEE Transactions on central sleep apnea events from a pressure
Pattern Analysis and Machine Intelligence, 2020, sensitive mat [ J ] . IEEE Access, 2020, 8:
42(8) : 2011-2023. 173428-173439.
[11] BANLUESOMBATKUL N, RAKTHANMANON [19] SRIVASTAVA N, HINTON G, KRIZHEVSKY
T, WILAIPRASITPORN T. Single channel ECG A, et al. Dropout: a simple way to prevent
for obstructive sleep apnea severity detection neural networks from overfitting [ J] . Journal of
using a deep learning approach [ C ] / / TENCON Machine Learning Research, 2014, 15: 1929-
2018 - 2018 IEEE Region 10 Conference, Jeju, 1958.
Korea. IEEE, 2018: 2011-2016. [20] HINTON G E. A practical guide to training
[12] RONNEBERGER O, FISCHER P, BROX T. U- restricted Boltzmann machines [ J] . Momentum,
Net: convolutional networks for biomedical image 2010, 9(1) : 926-947.
Copyright of Journal of Donghua University is the property of Journal of Donghua University
Editorial Board and its content may not be copied or emailed to multiple sites or posted to a
listserv without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.