Classification between normal and adventitious lung sounds using
Classification between normal and adventitious lung sounds using deep neural network Lin Li 1, Wenhao Xu 1, Qingyang Hong 1, Feng Tong 2, Jinzhun Wu 3 1 School of Information Science and Technology, Xiamen University, China 2 Key Lab of Underwater Acoustic Communication and Marine Information Technology of MOE, Xiamen University, China 3 The First Affiliated Hospital of Xiamen University, China 1 Acoustic Signal and Speech Processing Lab: http: //speech. xmu. edu. cn 2020/9/17
Outline 1、 Background 2、 Proposed Methods 3、 The Lung Sound Database 4、 Results and Analysis 5、 Conclusion 2 2020/9/17
Background (1) 3 2020/9/17
Background (1) The European Respiratory Society 4 & Computerized Respiratory Sound Analysis (CORSA) Wardship Aim Doctors practice 2020/9/17
Background (2) Ø GMM [Jin 2011] Ø HMM [Matsutake 2013] Ø ANN [Abbas 2010] Ø Wavelet Transform [Pesu 1998] Ø SVM [Abbasi 2013] [Jin 2011] F Jin, S Krishnan, F Sattar. Adventitious sounds identification and extraction using temporal-spectral dominance-based features [J]. Biomedical Engineering, IEEE Transactions on 58(11): 3078 -3087, 2011. [Matsutake 2013] S Matsutake, M Yamashita, S Matsunaga. Discrimination between healthy subjects and patients using lung sounds from multiple auscultation points [C]. Acoustics, Speech and Signal Processing (ICASSP), 1296 -1300, 2013. [Abbas 2010] A Abbas, A Fahim. An automated computerized auscultation and diagnostic system for pulmonary diseases [J]. Journal of medical systems, 34(6): 1149 -1155, 2010. [Pesu 1998] Pesu, L. , et al. "Classification of respiratory sounds based on wavelet packet decomposition and learning vector quantization. " Technology and Health Care 6. 1, 65 -74, 1998. [Abbasi 2013] Abbasi, Samira, et al. "Classification of normal and abnormal lung sounds using neural network and support vector machines. " Proceedings of the 21 st Iranian Conference on Electrical Engineering, May. 2013. 5 2020/9/17
The Proposed Method Automatic Lung Sound Recognition Figure 1: The flowchart of the proposed DNN-HMM system 6 2020/9/17
ANC [Harrison 1986 ] Clean lung sounds Collected lung sounds Enhanced lung sounds Noise Figure 2: The structure of ANC 7 [Harrison 1986] W. Harrison, J. S. Lim, and E. Singer, “A new application of adaptive noise cancellation, ” IEEE Transactions on Acoustics Speech & Signal Processing, vol. 34, no. 1, pp. 21– 27, 1986. 2020/9/17
GMM-HMM Figure 3: The structure of GMM-HMM 8 2020/9/17
DNN-HMM 9 atoms 2 hidden layers (500 atoms/each layer) 216 atoms 9 frames MFCC connection 9 Figure 4: The structure of DNN-HMM system 2020/9/17
The Lung Sound Database Table 1: The number of the lung sounds train and test sets Train Figure 5: The lung sounds auscultation points u u u 10 u Test Normal Adventitious 140 93 77 43 Sampled at 25 seconds. Collected from the hospital. Labeled by professional doctors. Enhanced by ANC, MFCC was extracted. 2020/9/17
Results –ANC (1) 11 Figure 6: The spectrum of lung sound samples 2020/9/17
Results –ANC (2) Figure 7: The performance of GMM-HMM compared with and without ANC 12 Ø In GMM-HMM baseline system, ANC achieved 8. 7% improvement. 2020/9/17
Results (3) DNN structure Ø Cepstral-mean variance normalization (CVN) was used. 4% Ø Two hidden layers were included. Ø Concatenations of 9 frames of MFCCs as input data. Figure 8: The DNN-HMM classification performance compared with GMM-HMM 13 Ø 216 units for input layer. Ø 500 units for each hidden layer. 2020/9/17
Results (4) Table 2: The classification performances on DNN architecture with different input frames and hidden atoms No. of Input Frames No. of Hidden Atoms Accuracy Rate% 3 2* 72 =144 87. 50 5 2*120=240 85. 29 7 2*168=336 94. 12 9 2*216=432 91. 91 11 2*264=528 90. 44 Ø The DNN-HMM structure was comparable with GMM-HMM when the input frames were more than 7 frames. 14 2020/9/17
Results (5) Table 3: The classification performances on DNN architecture with 500 hidden atoms 15 No. of Input Frames No. of Hidden Atoms Accuracy Rate% 3 500 81. 62 5 500 83. 82 7 500 86. 03 9 500 95. 59 11 500 95. 59 Ø The performance was degraded rapidly when the number of input frames was less than 9. Ø The accuracy rate was optimized to 95. 59% when the number of 2020/9/17 input frames was 9 or 11.
Results (6) Table 4: Comparison between Table 2 and Table 3 16 No. of Input Frames Atom Changing Accuracy Rate Changing % 3 144→ 500 87. 50→ 81. 62 5 240→ 500 85. 29→ 83. 82 7 336→ 500 94. 12→ 86. 03 9 432→ 500 91. 91→ 95. 59 11 528→ 500 90. 44→ 95. 59 Ø When the hidden atom number changed, the accuracy rates following changed obviously. Ø Unsuitable configuration of input and hidden layer would reduce the performance of the proposed system. 2020/9/17
Results (7) Figure 9: The accuracy using different percentage of pre-trained lung sound samples 17 Ø The classification achieved the best one when the percentage was 10%. 2020/9/17
Conclusion Ø A new strategy to discriminate between normal and adventitious lung sounds for automatic classification in practical environment. Ø ANC based enhancement at the signal pre-processing stage. Ø GMM-HMM (baseline) V. S. DNN-HMM(proposed) Ø The proposed DNN-HMM method outperformed the GMM-HMM baseline system and achieved the accuracy of 95. 59%. 18 2020/9/17
Prospect Ø Multi classification in adventitious lung sounds. Ø Optimized the DNN-HMM system parameters. Ø Increase the lung sound database. 19 2020/9/17
Thank you! 20 2020/9/17
- Slides: 20