A Novel Transfer Learning-Based Approach for Screening Pre-existing Heart Diseases Using Synchronized ECG Signals and Heart Sounds
Ramith Hettiarachchi, Udith Haputhanthri, Kithmini Herath, Hasindu Kariyawasam, Shehan Munasinghe, Kithmin Wickramasinghe, Duminda Samarasinghe, Anjula De Silva, Chamira U. S. Edussooriya
AA Novel Transfer Learning-Based Approach forScreening Pre-existing Heart Diseases UsingSynchronized ECG Signals and Heart Sounds
Ramith Hettiarachchi , Udith Haputhanthri , Kithmini Herath , Hasindu Kariyawasam , Shehan Munasinghe ,Kithmin Wickramasinghe , Duminda Samarasinghe , Anjula De Silva and Chamira U. S. Edussooriya Department of Electronic and Telecommunication Engineering, University of Moratuwa, Sri Lanka Lady Ridgeway Hospital for Children, Sri Lanka
Abstract —Diagnosing pre-existing heart diseases early in lifeis important as it helps prevent complications such as pul-monary hypertension, heart rhythm problems, blood clots, heartfailure and sudden cardiac arrest. To identify such diseases,phonocardiogram (PCG) and electrocardiogram (ECG) wave-forms convey important information. Therefore, effectively usingthese two modalities of data has the potential to improve thedisease screening process. We evaluate this hypothesis on asubset of the PhysioNet Challenge 2016 Dataset which containssimultaneously acquired PCG and ECG recordings. Our novelDual-convolutional neural network based approach uses transferlearning to tackle the problem of having limited amounts ofsimultaneous PCG and ECG data that is publicly available, whilehaving the potential to adapt to larger datasets. In addition, weintroduce two main evaluation frameworks named record-wise and sample-wise evaluation which leads to a rich performanceevaluation for the transfer learning approach. Comparisons withmethods which used single or dual modality data show that ourmethod can lead to better performance. Furthermore, our resultsshow that individually collected ECG or PCG waveforms are ableto provide transferable features which could effectively help tomake use of a limited number of synchronized PCG and ECGwaveforms and still achieve significant classification performance.
Keywords —Automated Screening, Phonocardiogram, Electro-cardiogram, Convolutional Neural Networks, Transfer Learning.
I. I
NTRODUCTION
Pre-existing heart diseases are a range of abnormalitieswhich affect the heart, that are either existing from birth(congenital heart diseases) or acquired later in life. Thestethoscope is considered as the tool-of-choice ubiquitouslyused for initial medical examination of such complications byauscultation. However, accurate diagnosis using a stethoscoperequires considerable expertise [1]. In such occasions, anautomated mechanism to help diagnose or screen for specificcardiac conditions can reduce the dependence on inadequatehealthcare infrastructure and lead to significant improvementsin the quality of patient care. While such a solution in generalwould deliver considerable healthcare benefits, the recentnovel coronavirus pandemic, where persons with underlying © 2021 IEEE. Personal use of this material is permitted. Permission fromIEEE must be obtained for all other uses, in any current or future media,including reprinting/republishing this material for advertising or promotionalpurposes, creating new collective works, for resale or redistribution to serversor lists, or reuse of any copyrighted component of this work in other works. conditions including those with cardiac complications aredeemed to be at a higher risk of co-morbidity and mortality,has made this endeavor more critical [2], [3].To this end, studies have reported deep learning approachesfor detecting cardiac abnormalities using either: (a) PCG sig-nals [4]–[6] used for traditional stethoscope-based screeningor, (b) ECG signals [7] typically reserved for advanced diag-nosis. Most deep learning techniques in the literature translatetime domain PCG or ECG waveforms to higher dimensionalrepresentations depicting their time-frequency characteristicswhich are then fed to deep learning algorithms. However,these studies do not evaluate the performance improvementthat could be achieved via the integration of both PCG andECG signals. This is especially important as there can be manyoccasions where the two types of signals convey mutuallyexclusive information regarding disease status [8]. That is, acombination of ECG and PCG can provide accurate informa-tion about heart murmurs and other abnormal heart soundsthat help to do accurate diagnosis. Therefore, integrating theanalysis of PCG and ECG has the potential to significantlyimprove disease screening.A recent study by Chakir et al. [9] used traditional machinelearning techniques to detect cardiac abnormalities from si-multaneous PCG and ECG waveforms. By utilizing a subset ofrecordings from the PhysioNet [10] Challenge 2016 (PHY16)[11] (training-a) dataset, they demonstrate that classificationperformance can be improved via integration of PCG andECG beyond the sole use of PCG signals. However, usuallytraditional machine learning techniques have difficulty in per-forming well on large datasets [12], whereas to tackle thisproblem through deep learning, there should be a sufficientlylarge set of simultaneous PCG and ECG recordings. Thiscreates a significant constraint for efficacious training and sub-sequent deployment: simultaneous PCG and ECG acquisitiondatasets are minimal, as such considerable effort and timehas to be invested afresh to create new datasets for reliablytraining classifiers before its usage. Since devices capable ofsimultaneous PCG and ECG acquisition are either expensiveor not yet widely adopted, this extra consideration can translateinto years worth of new data collection before its deployment.To address this critical limitation, we propose a novel dual- a r X i v : . [ ee ss . SP ] F e b PhysioNet 2017PhysioNet 2016 SamplewiseRecordwiseEvaluationPCG-only ClassificationECG-only Classification
PCGScalogram GenerationECGECGPCG+ECGPCG
Balanced / ImbalancedSimultaneous PCG + ECG Hybrid Classification
PCG+ECG
Fig. 1:
The proposed transfer learning based pipeline : Balanced datasets created using setting-1 and setting-2 were fedinto PCG-only and ECG-only models separately. The trained weights before the flatten layers of those models were used toinitialize the feature extraction path weights of the hybrid model. The hybrid model was trained without freezing the featureextraction path weights using balanced or imbalanced datasets created based on setting-3. After obtaining the predictions fromthe hybrid model, sample-wise and record-wise evaluations were done separately.convolutional neural network (CNN) based architecture whichemploys transfer learning [13]. This approach allows us to uselarge pre-existing datasets of individually acquired PCG andECG waveforms, train two distinct CNNs to identify cardiaccomplications using each type of recordings, and transferthe learnt features onto a single integrated CNN. This canthen be deployed for screening using dual-input PCG andECG waveforms. We use the PHY16 (training-a) dataset fortraining, validating and testing our approach and demonstrateits feasibility for a reasonably eminent deployment.II. M
ETHODS
Here, we present the proposed transfer learning basedcardiovascular abnormality detection approach. Fig. 1 showsthe overall pipeline where the generation of scalograms of theraw signals is followed by the hybrid classification model.
A. Scalogram Generation
We employ scalograms to represent the time-frequencydistribution of PCG and ECG signals. To this end, we applya continuous wavelet transform (CWT) [14] to both ECG andPCG signals because it is suitable for analysing non-stationarysignals using time-frequency representations. We select theMorlet wavelet [15] as the mother wavelet for the PCG signaltransformation, since it has desirable properties of localizationin both time and frequency domains of PCG signals [16].We choose the scaling parameter to be in the range of - , based on the frequency content of the the PCG signalthat the wavelet should be sensitive to when applying thetransform, to extract the frequency content of the signal inthe time-frequency domain [15]. We select the complex Morletwavelet [15] as the mother wavelet for the wavelet transform ofthe ECG signals. The complex wavelet allows both amplitude and phase information to be obtained in wavelet space andlow values of the center frequency can be particularly usefulfor detecting short-duration signal transients. [ ? ] We set thebandwidth parameter and center frequency to . and Hz,respectively. The range of the scaling parameter is set to - (the reasoning was the same as for PCG signals). Examplescalograms generated from the wavelet coefficients of a PCGand ECG signal are shown in Fig. 1. B. Data Description
We used the following three settings for creating datasetsfrom the PHY16 and Physionet Challenge 2017 (PHY17) [18]datasets for model training. • Setting 1 - For this setting, we only included the PHY16(training- b, c, d, e, f) datasets which contained only PCGrecordings. We used a local peak detector [19] to detectthe main peak in each heart cycle and non-overlappingsamples of . s windows were selected with respect tothe detected peaks. Then, scalograms were generated foreach of these samples. • Setting 2 - PHY17 dataset was used for this setting whichcontained only ECG recordings. A QRS detector [20] wasused to detect R peaks [20] of each ECG signal. From thedetected n number of peaks, we then selected samplesof . s windows centering the (cid:4) n (cid:5) th R peak. Finally,we generated scalograms for each selected sample of aparticular record. • Setting 3 - We used simultaneously recorded PCG andECG signals available in the PHY16 (training-a) dataset( recordings) in this setting. Since the number ofrecordings were limited, the recordings were split intosamples of . s windows using the same approachemployed in Setting 1. ECG signals were also segmentedrom the same points that the PCG signals were seg-mented. This effectively produced simultaneousPCG and ECG samples. Finally, we generated scalogramsfrom the segmented samples.The scalograms in each of the above settings were divided intosets of 70%, 10%, 20% for training, validation and testingpurposes respectively. This segregation was done randomlybased on the record identities instead of sample identities.Therefore, we were able to make sure that the samplesgenerated from the same record were not shared across thetraining, validation and test sets. Moreover, we also createdimbalanced and balanced datasets for all the above settings. C. Novel CNN Architecture
The following CNN architectures were designed to classifyPCG and ECG signals as abnormal and normal .
1) PCG/ECG Classification:
We implemented separateECG-only, PCG-only classification models to classify abnor-mal/normal conditions of ECG, PCG datasets separately. Here,three repetitive downsampling blocks were implemented asfeature extractors followed by a convolutional layer alongwith a flatten layer. The downsampling block contained astack of convolution, maxpooling and rectified linear unit(ReLU) layers each. The resultant feature vector was fed intoa multilayer perceptron consisting of 3 dense layers followedby a softmax output.
2) Hybrid Model:
The intuition behind this hybrid modelis to use a superimposed feature representation of both PCGand ECG signals. Therefore, this architecture comprises ofseparate PCG and ECG-feature extraction paths. To constructthe hybrid model, outputs of the flatten layers of the twoarchitectures used in ECG-only and PCG-only classification(subsection II-C1) were concatenated and then followed by ashared multilayer perceptron (See Fig. 1).
D. Transfer Learning
Since the number of simultaneous PCG and ECG recordingsobtained from the PHY16 (training-a) data is too low (407recordings) to train the hybrid model, we have adopted atransfer learning approach [13]. Initially, individual PCG-onlyand ECG-only models were trained on Setting 1 and Setting2, respectively for both balanced and imbalanced datasets. Totransfer the learnt knowledge, the weights before the flattenlayers of both ECG-only, PCG-only models were then used inthe PCG-feature extraction and ECG feature extraction pathsof the hybrid model architecture (See Fig. 1).
E. Determining an Optimal Threshold for Classification
Obtaining predictions using the default threshold (0.5) onthe softmax output of the model will not be the best method forimbalanced datasets [21]. Therefore we computed the optimalthreshold such that it maximizes the G-mean metric givenin (1) for the validation set [22]. This effectively improvedsensitivity and specificity parameters for the validation setwhich is a key requirement in medical diagnostic tests. G - M ean = (cid:112) Sensitivity × Specif icity (1) The computed optimal threshold was then used to evaluatethe model on the test set. Results showed that this methodsignificantly removed the correlation between the imbalancednature of the dataset and model performance (See Table I).III. E
XPERIMENTS
A. Evaluation Criteria
We conducted a comprehensive analysis to measure therobustness of our method based on the factors: 1. transferlearning, 2. balanced/imbalanced dataset setting, 3. predictionsusing optimal/default thresholding and 4. sample-wise/record-wise evaluation.In sample-wise evaluation, final prediction scores were ob-tained based on the output of the hybrid CNN while for record-wise evaluation, the final prediction scores were obtained byaggregating the sample-wise predictions of samples whichcorresponded to a particular record.
B. Evaluation Metrics
To measure the robustness of our method, we evaluatedour models’ performance on the test set with respect toAccuracy (Acc), Sensitivity (Sen), Specificity (Spe) [24], G-mean and Area under the Receiver Operating CharacteristicCurve (AUC). We have focused on improving the G-meanscore since it ensures that both sensitivity and specificity arehigher (1). Furthermore, based on the results, we have foundthat the G-mean score is practically a better measurementto capture the cases where the sensitivity is higher with acomparable specificity as well (Table I).We then explored the effect of the above-mentioned factorsusing different training settings which included different modelarchitectures and hyper-parameters. The default training set-tings included: ADAM optimizer with a learning rate of 0.001,batch size of 20, categorical cross-entropy as the loss functionwith the architectures presented in section II-C. The trainingsettings of the best results are described in section IV.IV. R
ESULTS
Table I shows a comparison based on different evaluationcriteria explained in the section III. Our method [D] wasable to achieve a high sensitivity of while having areasonable specificity of 75% when using dilated convolutionand dropouts. This was on par with the results of Chakir et al. [9] which used 100 simultaneous PCG and ECG records wherethe class distribution was unknown. Even though their selectedhandcrafted features were able to produce good results on thisdata, for a larger population, those features might not be ableto capture properties of diverse abnormalities [25]. Therefore,it poses a limitation to extend their method to large datasets.On the other hand, since the feature extraction is learnt bythe CNN itself, it has the potential to capture properties ofcomplex abnormalities. Furthermore, support vector machine’salgorithm complexity hinders it from being applied to largedatasets [26].The reason for the decreased specificity in method [D] ismainly due to class imbalance (117 Normal, 290 Abnormal)present in the dataset. However, our method [A] showed thatABLE I: Comparison of previous approaches vs. our experiments using standard statistical evaluation parameters.
Author Method Database TransferLearning OptimalThreshold Input Results (%)Chakir et al. [9](2020) Support Vector Machine(SVM) PhysioNet 2016 (a)subset of 100 records - - record-wise(PCG+ECG) Sen = 92.31Spe = 92.86 Acc = 92.5AUC = 95.05 G-mean= 92.58 (cid:63) Li et al. [6](2020) CNN PhysioNet 2016(a,b,c,d,e,f) - - record-wise(PCG only) Sen = 87Spe = 86.6 Acc = 86.8 G-mean= 86.8Ren et al. [23](2018) Learnt VGG Netcom/w a SVM PhysioNet 2016(PCG signals only) Yes - Sen = 24.6Spe = 87.8 Acc = 56.2 G-mean= 46.47Our Method [A] Hybrid CNN Setting 3(Imbalanced) Yes Yes record-wise(PCG+ECG) Sen = 87.72Spe = Acc = 87.67AUC =
G-mean=
Our Method [B] Hybrid CNN Setting 3(balanced) Yes Defaultthresh=0.5 record-wise(PCG+ECG) Sen = 85.7Spe = 82.6 Acc = 84.1AUC = 87.06 G-mean= 84.15Our Method [C] Hybrid CNN Setting 3(balanced) Yes Yes sample-wise(PCG+ECG) Sen = 81.71Spe = 81.22 Acc = 81.45AUC = 85.83 G-mean= 81.47Our Method [D] Hybrid CNN Setting 3(Imbalanced) No Yes record-wise(PCG+ECG) Sen =
Spe = 75 Acc =
AUC = 91.06 G-mean= 84.29Our Method [E] Hybrid CNN Setting 3(balanced) No Defaultthresh=0.5 record-wise(PCG+ECG) Sen = 80.95Spe = 78.26 Acc = 79.55AUC = 82.5 G-mean= 79.6Our Method [F] Hybrid CNN Setting 3(balanced) No Yes sample-wise(PCG+ECG) Sen = 78.86Spe = 77.67 Acc = 78.23AUC = 86.04 G-mean= 78.26Our Method [G] PCG-only Setting 3(balanced) No Defaultthresh=0.5 sample-wisePCG Sen = 56.02Spe = 49.50 Acc = 52.69AUC = 55.26 G-mean= 52.65Our Method [H] ECG-only Setting 3(balanced) No Defaultthresh=0.5 sample-wiseECG Sen = 70.16Spe = 81.73 Acc = 76.03AUC = 82.16 G-mean= 75.72 (cid:63)
Distribution of class proportions is unknown even with an imbalanced dataset, it is possible to keep afine balance between sensitivity and specificity by leveragingtransfer learning. Performance improvements from methods[F] to [C] and [E] to [B] further strengthens this claim.Our method [A] outperforms the work done by Li et al. [6]which uses only PCG data. However, a direct comparison isnot possible due to different utilization settings of the PHY16dataset. Our model used PCG and ECG feature extraction pathweights obtained from best PCG-only (PCG-22) and ECG-only (ECG-4) models trained on setting 1 and 2, respectively.Furthermore, we outperform the work by Ren et al. [23] whofollowed a transfer learning approach on a VGG architecture.Methods [B] and [E] which used the default threshold (0.5)were the best results obtained for the balanced record-wise set-ting. Meanwhile, methods [C] and [F] which used the optimalthreshold were the best results for the balanced sample-wisesetting. Even though the optimal threshold improved resultssignificantly in the imbalanced dataset setting, in the balanceddataset setting it has not consistently given the best results.Methods [A], [B], [D] and [E] show that record-wiseevaluation gives better performance over methods [C] and[F] which uses sample-wise evaluation. This is because, theutilization of the complete time domain signals in record-wiseevaluation makes it more robust to temporal distortions thatmay occur during the acquisition of signals. Methods [G] and[H] shows the results obtained from the models
PCG-22 and
ECG-4 when they were trained on setting 3 balanced dataset.It can be clearly seen that the performances of all hybridmodels are significantly better than the performances of PCG-only, ECG-only models. The AUC scores obtained from thereceiver operating characteristic curves in Fig. 2 have been
False Positive Rate T r u e P o s i t i v e R a t e Method : AMethod : BMethod : CMethod : DMethod : EMethod : FMethod : GMethod : H
Fig. 2: Receiver operating characteristic curvessummarized in Table I, which shows that our methods achieveda maximum sensitivity of . and a maximum G-meanscore of . with an AUC of . .V. C ONCLUSION
We have demonstrated a proof-of-concept implementationof a novel classification framework for automated screening ofcardiac complications by using dual-input PCG and ECG data.By employing transfer learning techniques, we circumventedthe need to train our dual-input CNN solely on simultane-ous PCG and ECG data, opening the potential to use vastamounts of individually acquired PCG and ECG waveformscontained in pre-existing datasets. Our current performanceindices which were achieved only using 407 simultaneousPCG and ECG recordings, already show performance on-parto existing literature. We believe that this performance can beimproved via expanded training with other well known onlinedatasets which tabulate individual PCG or ECG data.I. A
CKNOWLEDGEMENTS
Authors thank the University of Moratuwa for the finan-cial support. Furthermore, authors thank Mr. Jathushan Ra-jasegaran for helpful suggestions and Mr. Vidura Dhananjayafor providing computing resources on Amazon AWS.R
EFERENCES[1] M. Mehmood, H. L. Abu Grara, J. S. Stewart, and F. A. Khasawneh,“Comparing the auscultatory accuracy of health care professionals usingthree different brands of stethoscopes on a simulator,”
Medical Devices:Evidence and Research , 2014.[2] T. Guo, Y. Fan, M. Chen, X. Wu, L. Zhang, T. He, H. Wang, J. Wan,X. Wang, and Z. Lu, “Cardiovascular Implications of Fatal Outcomesof Patients with Coronavirus Disease 2019 (COVID-19),”
JAMACardiology , vol. 5, no. 7, pp. 811–818, 2020.[3] D. Wang, B. Hu, C. Hu, F. Zhu, X. Liu, J. Zhang, B. Wang, H. Xi-ang, Z. Cheng, Y. Xiong, Y. Zhao, Y. Li, X. Wang, and Z. Peng,“Clinical Characteristics of 138 Hospitalized Patients with 2019 NovelCoronavirus-Infected Pneumonia in Wuhan, China,”
JAMA - Journalof the American Medical Association , vol. 323, no. 11, pp. 1061–1069,2020.[4] S. Aziz, M. U. Khan, M. Alhaisoni, T. Akram, and M. Altaf, “Phonocar-diogram signal processing for automatic diagnosis of congenital heartdisorders through fusion of temporal and cepstral features,”
Sensors(Switzerland) , vol. 20, no. 13, pp. 1–20, 2020.[5] T. Dissanayake, T. Fernando, S. Denman, S. Sridharan, H. Ghaem-maghami, and C. Fookes, “Understanding the Importance of HeartSound Segmentation for Heart Anomaly Detection,” pp. 1–9, 2020.[6] F. Li, H. Tang, S. Shang, K. Mathiak, and F. Cong, “Classificationof heart sounds using convolutional neural network,”
Applied Sciences(Switzerland) , vol. 10, no. 11, 2020.[7] J. Huang, B. Chen, B. Yao, and W. He, “ECG Arrhythmia ClassificationUsing STFT-Based Spectrogram and Convolutional Neural Network,”
IEEE Access , vol. 7, pp. 92871–92880, 2019.[8] F. Sato, “Electrocardiogram of Congenital Heart Disease in Relationto their Operative Findings : Electrocardiogram and Phonocardiogramin Relation to Operative Findings : Symposium on Limitation of Diag-nostic Value of Electrocardiography and Phonocardiography,”
JapaneseCirculation Journal , vol. 30, no. 12, pp. 1537–1541, 1966.[9] F. Chakir, A. Jilbab, C. Nacir, and A. Hammouch, “Recognitionof cardiac abnormalities from synchronized ECG and PCG signals,”
Physical and Engineering Sciences in Medicine , vol. 43, no. 2, pp. 673–677, 2020.[10] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. Ch.Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C. K. Peng, and H. E.Stanley, “PhysioBank, PhysioToolkit, and PhysioNet,”
Circulation ,2000. [11] G. D. Clifford, C. Liu, B. Moody, D. Springer, I. Silva, Q. Li, and R. G.Mark, “Classification of normal/abnormal heart sound recordings: ThePhysioNet/Computing in Cardiology Challenge 2016,” in
Computing inCardiology , 2016.[12] P. Curtis, K. Bigelow, B. A. T. S. Cook, A. E. Flanders, and P. Matthew,“A roadmap for the foundational research on artificial intelligence inmedical imaging,”
Radiology , vol. 291, no. 3, pp. 781–791, 2019.[13] L. Pratt, J. Mostow, and C. Kamm, “Direct Transfer of LearnedInformation Among Neural Networks,”
AAAI-91 Proceedings , pp. 584–589, 1991.[14] J. L. Semmlow and B. Griffel,
Biosignal and Medical Image Processing ,CRC Press, 2014.[15] F. Wasilewski, “Continuous wavelet transform (cwt) — pywavelets doc-umentation,” pywavelets.readthedocs.io. https://pywavelets.readthedocs.io/en/latest/ref/cwt.html (accessed Jul. 4, 2020).[16] Z. M. Zin, S. H. Salleh, S. Daliman, and M. D. Sulaiman, “Analysisof heart sounds based on continuous wavelet transform,”
StudentConference on Research and Development: Networking the Future Mindin Convergence Technology, SCOReD 2003 - Proceedings , pp. 19–22,2003.[17] M. K. Stiles, D. Clifton, N. R. Grubb, J. N. Watson, and P. S. Addison,“Wavelet-based analysis of heart-rate-dependent ecg features,” October2004, vol. 9, pp. 316–322.[18] G. D. Clifford, C. Liu, B. Moody, L. H. Lehman, I. Silva, Q. Li,A. E. Johnson, and R. G. Mark, “AF classification from a short singlelead ECG recording: The PhysioNet/computing in cardiology challenge2017,” in
Computing in Cardiology , 2017.[19] “Processing — wfdb 3.1.1 documentation,” wfdb.readthedocs.io. https://wfdb.readthedocs.io/en/latest/processing.html (accessed Jul. 4, 2020).[20] W. A. H. Engelse and C. Zeelenberg, “Single Scan Algorithm for QRS-Detection And Feature Extraction,” in
Computers in Cardiology , 1979.[21] Z. H. Zhou and X. Y. Liu, “Training cost-sensitive neural networks withmethods addressing the class imbalance problem,”
IEEE Transactionson Knowledge and Data Engineering , vol. 18, no. 1, pp. 63–77, 2006.[22] M. Kubat and S. Matwin, “Addressing the curse of imbalanced trainingsets: one-sided selection,” in
International Conference on MachineLearning , 1997.[23] Z. Ren, N. Cummins, V. Pandit, J. Han, K. Qian, and B. Schuller,“Learning image-based representations for heart sound classification,”
ACM International Conference Proceeding Series , pp. 143–147, 2018.[24] R. Trevethan, “Sensitivity, Specificity, and Predictive Values: Founda-tions, Pliabilities, and Pitfalls in Research and Practice,”
Frontiers inPublic Health , 2017.[25] D. Xie, L. Zhang, and L. Bai, “Deep Learning in Visual Computingand Signal Processing,”
Applied Computational Intelligence and SoftComputing , vol. 2017, 2017.[26] J. Cervantes, F. Garcia-Lamont, L. Rodr´ıguez-Mazahua, and A. Lopez,“A comprehensive survey on support vector machine classification:Applications, challenges and trends,”