Embed Size (px)
Citation preview
Unmixed Spectrum Clustering for Template Composition in Lung Sound Classification
Tomonari MASADANagasaki University
Agenda
Determine “phonemes” of lung sounds Extract elementary and constructive parts
of lung sounds
Classify unknown lung sounds Reconstruct lung sounds as a sequence
of lung sound phonemes
respiratorysounds
breathsounds
Lung Sounds (= Respiratory Sounds)
adventitioussounds
normal
abnormal
pulmonaryadventitious
sounds(rales)
miscellaneous
vesicular sounds
bronchovesicular sounds
tracheal sounds
decreased
increased
absentbronchial sounds in abnormal locations
tracheal stenosis
continuous
discontinuous(crackles)
pleural friction rub
low-pitched (rhonchi)
high-pitched (wheeze)
squawk
fine crackles
coarse crackles
But…
No consensus on
how to define “phonemes” of lung sounds(to our best knowledge)
“Phoneme” of Lung Sounds (1/2)
Extract component spectra Split a lung sound into overlapping segments
(~1sec) Split each segment into 32 overlapping wind
ows Apply FFT 32 log power spectra Unmix log power spectra by ICA or NMF
and Obtain two component spectra
fine crackle (32 log power spectra)
0
5
10
15
20
25
1 51 101 151 201
fine crackle (two component spectra)
-2
-1
0
1
2
3
4
5
1 51 101 151 201
“Phoneme” of Lung Sounds (2/2)
Construct equivalence classes of sound segments Repeat k-means 100 times Regard component spectra giving
consistent memberships as equivalent
ICA
ICA
32 log power spectra
32 log power
spectra
same cluster in all clustering results
Data Specifications
Lung sounds from a textbook CD Yonemaru, M., Sakurai, T.
Lung Sound Auscultation Training via CD for Nurses. Nanko-do, Tokyo (2001) 31 types of lung sounds ~30min length in total
Ground truth
= Disease name assigned to sound data
Details of Our Method (1/3)
Computing Power Spectra
Sound segment 0.81 sec length = 32 FFT windows 1/2 overlap
FFT window 0.093 sec length = 1024 points in 11,025Hz 3/4 overlap
Details of Our Method (2/3)
Unmixing Power Spectra
Component spectra 32 power spectra
= linear combinations of
two component spectra
fastICA NMF [Cichocki et al. 2006]
ICA, NMF = Matrix Factorization
≒ ×・・・ ・・・
ARDS (adult respiratory distress syndrome)
0
5
10
15
20
25
1 51 101 151 201
ARDS (adult respiratory distress syndrome)
-4
-3
-2
-1
0
1
2
3
4
1 51 101 151 201
Details of Our Method (3/3)
Sound Segment Equivalency
Component spectra 12,290 spectra = 6,145 sound segments
Component spectrum clustering 100 times of k-means “equivalent” = same cluster in all results
1,558 segments 305 equivalent classes
same cluster in all clustering results
ICA
ICA
32 log power spectra
32 log power
spectra
-2
-1
0
1
2
3
4
5
1 51 101 151 201
-2
-1
0
1
2
3
4
5
1 51 101 151 201
-2
-1
0
1
2
3
4
5
1 51 101 151 201
-2
-1
0
1
2
3
4
5
1 51 101 151 201
-4
-3
-2
-1
0
1
2
3
4
1 51 101 151 201
-4
-3
-2
-1
0
1
2
3
4
1 51 101 151 201
Equivalent Sound Segments== =
ICA
ICA
32 log power spectra
32 log power
spectra
Results (1/2)
9~128 9~192 9~256 9~384ICA 0.9387 / 0.0668
(186:24)0.9609 / 0.0698
(212:24)0.9605 / 0.0615
(215:26)0.9779 / 0.0523
(246:28)
NMF 0.9718 / 0.0451(198:27)
0.9880 / 0.0373(245:26)
0.9856 / 0.0375(299:28)
0.9484 / 0.0396(326:29)
mean 0.8469 / 0.0281(556:31)
0.8697 / 0.0309(601:31)
0.8588 / 0.0361(547:28)
0.8365 / 0.0425(467:27)
17~128 17~192 17~256 17~384ICA 0.9642 / 0.0702
(161:23)0.9560 / 0.0693
(165:19)0.9644 / 0.0685
(192:25)0.9696 / 0.0575
(195:28)
NMF 0.9245 / 0.0497(204:25)
0.9721 / 0.0534(204:24)
0.9929 / 0.0348(305:29)
0.9015 / 0.0408(294:27)
mean 0.8112 / 0.0316(511:31)
0.8434 / 0.0334(562:31)
0.8505 / 0.0346(535:29)
0.8452 / 0.0396(455:28)
Results (2/2)
100 times 80 times 60 times 40 times 20times
0.99290.0348
(305:29)
0.98800.0345
(320:28)
0.97690.0351
(351:29)
0.96190.0328
(411:31)
0.89580.0298
(494:30)
Unmixing method : NMF [Cichocki et al. 2006]
Frequency range : 17 ~ 256 (180~2760Hz)
Future Work
Determine “phonemes” of lung sounds Extract elementary and constructive parts
of lung sounds
Classify unknown lung sounds Reconstruct lung sounds as a sequence
of such elementary parts
