Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Classification of Respiratory Sounds Simon Müller Institut für Stochastik und Anwendungen Universität Stuttgart CLEOS-Projekt Robert-Bosch-Krankenhaus 4. März 2012 Simon Müller Classification of Respiratory Sounds MARS Outline Introduction 1 Methods to Detect Adventitious Lung Sounds FDA Future Prospects Introduction The origin of lung sounds Locations of sound recording Vesicular breath sound Tracheal breath sound Summary Abnormal breath sound Adventitious sound 2 3 4 5 Summary Methods to Detect Adventitious Lung Sounds Separation and classification of crackles (fine and coarse) Separation and Classifcation of Wheeze (rhonchi, stridor, . . .) FDA Future Prospects MARS Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects The origin of lung sounds The origin of the lung sounds ([2], [3], [4]) The origin of lung sounds is not yet completely clear The lung itself cannot generate sounds, if there is no airflow (minimum of a flow is required) It is assumed that the breath sound is induced by turbulence of air at the level of lobar or segmental bronchi In smaller bronchi the gas velocity decreases and the flow is laminar ⇒ the point of origin is located in the larger air paths Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Locations of sound recording Locations of sound recording ([2], [3], [4]) Respiratory sounds have different characteristics depending on the location of recording and ventilation cycle (expiration or inspiration) The locations of recording can be divided into two classes: the trachea and the chest (front and back) Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Locations of sound recording Anterior Abbildung: Red: Area of the trachea, Green: The chest. Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Locations of sound recording Vesicular breath sound ([2], [3], [4]) Sound is filtered by the lung and the chest wall (low-pass filter) Sound is soft and low-pitched Pitch depends on the anatomy of the patient (e.g. age, BMI, physique) Frequency band contains also components of respiratory muscles and heart Inspiration phase is louder and has much higher frequency components than expiration phase Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Locations of sound recording Abbildung: Spectrum of an inspiration and expiration phase recorded at the trachea (the sound is taken from [6]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Locations of sound recording Tracheal breath sound ([2], [3], [4]) Sound is not or little filtered Sound is loud and high-pitched (up to 2300 Hz) Difference of the power of inspiration and expiration varies greatly among the subjects Tracheal sound has a direct connection to the flow of air ⇒ inspiration and expiration phases Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Locations of sound recording Abbildung: Spectrum of an inspiration and expiration phase recorded at the chest (the sound is taken from [6]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Locations of sound recording Summary Vesicular sound has higher diagnostic value than tracheal sound, since this part of the lung is affected by serious lung diseases Tracheal sound can be used to detect inspiration/expiration phases and eventually to estimate the flow of air ⇒ it may help to record lung sounds at both locations Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abnormal breath sound Bronchial sound ([2], [3]) Similar to tracheal sound Bronchial sound is typical for many diseases, for example Asthma Chronic bronchitis Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abnormal breath sound Crackle sound ([2], [3]) Dicontinuous lung sound Character is explosive and of transient nature Crackles can be separated into two classes: fine and coarse Coarse crackles are of less intensity and of longer duration than fine crackles Fine crackles instead are present in higher frequencies Pitch range is from 10 to 2000 Hz and duration < 20 ms Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abnormal breath sound Crackle sound ([2], [3]) Occurrence and number of the crackles are an indicator of the type of disease Crackles can be found in many diseases, for example Heart congestion failure Pneumonia Bronchiectasis Pulmonary fibrosis Chronic diffuse parenchymal lung disease They are an early sign for respiratory diseases, since fine crackles are originated in small air paths. Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abnormal breath sound Abbildung: Time/Frequency plane of a 16 year old boy with tuberculosis (the sound is taken from [7]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abnormal breath sound Wheeze Wheeze can be heard at the chest (low-pass filtered) and the trachea (unfiltered) Wheeze is a high-pitched sound (dominant frequency at 400 Hz or more) with a duration > 250 ms Is heard at expiration Wheeze can be found at many diseases, for example Congestive heart failure Asthma Pneumonia Chronic bronchities Emphysema Brochiectasis Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abnormal breath sound Abbildung: Time/Frequency plane of a 8 year old boy with asthma (the sound is taken from [7]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Abnormal breath sound Squawk Is a short wheeze Squawk can be found for example in Pneumonia Simon Müller Classification of Respiratory Sounds Future Prospects MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abnormal breath sound Rhonchi Similar to wheeze, but dominant frequency is about 200 Hz or less Usually rhonchi occur at airway narrows Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abnormal breath sound Stridor Loud wheeze and high pitched Usually stridor can be found for example in Upper airway obstruction Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abnormal breath sound Abbildung: Time/Frequency plane of a 15 month old girl with croup (the sound is taken from [7]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abnormal breath sound Pleural rub sound Sounds like running two pieces of leather against each other Pleural surface inflammation is a typical disease for this sound Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Summary Summary Many diseases can be classified by these adventitious sounds Next section deals with methods to extract them Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and classification of crackles (fine and coarse) Separation of the crackles from vesicular sound [9] A filter based on the wavelet packet transform is used for automatic separation of crackles from vesicular sounds Why does this technique work? Explosive peaks (crackles) have larger coefficients over many wavelet levels The coefficients of the background (vesicular sound) decrease with increasing scale [10]. Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and classification of crackles (fine and coarse) Abbildung: This picture shows an example of the WPST-NST filter ([9]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and classification of crackles (fine and coarse) Abbildung: Here we can see that the WPST-NST filter has just little affect to stationary signals. ([9]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and classification of crackles (fine and coarse) Abbildung: Close-up view of a crackle, Initial Deflection Width (IDW) and Two-Cycle Duration (2CD) ([8], [1]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and classification of crackles (fine and coarse) Classification of crackles [9] Next we need a classification of the crackles (fine or coarse) Fine crackle: average duration of IDW and 2CD is 0.7 and < 10 ms Coarse crackle: average duration of IDW and 2CD is 1.5 and > 10 ms Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and classification of crackles (fine and coarse) Optimal classification is achieved with the matched wavelet method ([8], [11]) Continuous wavelet transform is defined by 1 CWTx (τ, s) = s Z∞ x(t)Ψ( t −τ )dt s −∞ Ψ(t) is the mother wavelet and 1s Ψ( t−τ s ) the wavelet basis function, s is called the scale, and τ is a translation in the time axis The next two pictures show the energy distribution of fine and coarse crackles versus the scale s Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and classification of crackles (fine and coarse) Abbildung: Energy distribution of fine crackles ([11]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and classification of crackles (fine and coarse) Abbildung: Energy distribution of coarse crackles ([11]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and classification of crackles (fine and coarse) Abbildung: Schema of the crackle separation and classification algorithm ([8]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and classification of crackles (fine and coarse) Summary of crackle detection and classification [9] Xiaoguang and Bahoura achieved a classification rate of 93.9% with this wavelet based method ([8]) It is a state of the art algorithm to detect and classify crackles The information can be used to classify lung diseases (e.g. via number and character of the crackles, occurence (inspiration or/and expiration)) Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and Classifcation of Wheeze (rhonchi, stridor, . . .) Separation of the wheeze from vesicular/tracheal sound [12] A filter based on the continuous wavelet transform is used for automatic separation of wheeze from vesicular/tracheal sound Why does this technique work? Lung sound can be split into two sounds x(t) = s(t) + v (t) (wheeze and vesicular/tracheal sound) A characteristic of the wheeze s(t) is that it is sinusoidial s(t) = sin(ωs t + φ) In [12] the real Morlet mother wavelet is chosen as 2 p t cos(2πf0 t) ψ(t) = πfb exp ∆ − fb Because of this choice, one can easily get the time and scale information from the scalogram, Sc (a, b) = |CWTx (a, b)|2 . Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and Classifcation of Wheeze (rhonchi, stridor, . . .) Abbildung: (i) Lung sound with wheeze, (ii) spectrogram, (iii) scalogram ([12]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and Classifcation of Wheeze (rhonchi, stridor, . . .) Abbildung: The schema of this detection and separation algorithm ([12]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and Classifcation of Wheeze (rhonchi, stridor, . . .) Abbildung: An example of this method ([12]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and Classifcation of Wheeze (rhonchi, stridor, . . .) Summary of wheeze detection and classification There exists an algorithm to detect and seperate wheeze from normal lung sound The method also works for rhonchi and stridor Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Separation and Classifcation of Wheeze (rhonchi, stridor, . . .) Summary There are methods which are able to separate and classify adventitious sounds from lung sound With the extracted features we are able to classify different kinds of lung diseases Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects FDA - Functional Data Analysis This chapter deals with a newly developed method from speech classification Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Functional Classification in a separable Hilbert Space Let X be a random variable in a separable Hilbert space with label Y = {0, 1} Let (Xi , Yi ) be n independent copies of the pair (X , Y ) ∞ P (Xi , uk )uk with a complete orthonormal We know Xi = k=1 system (ONS) (ui )∞ i=1 of the Hilbert space Let our Hilbert space be L2 [a, b], then we can use the trigonometric basis as the complete ONS (ui )∞ i=1 Let Pd be a projection operator, which maps a functional variable to a d-dimensional subspace of the L2 [a, b] (for example the first d Fourier coefficients) Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Functional Classification in a separable Hilbert Space New functional variable Xnew with unknown label Ynew Apply Pd to Xnew Estimate Ynew via the k nearest neighbors of Xnew k k P P 0, if 1[Yi (x)=0] ≥ 1[Yi (x)=1] φk,d = i=1 i=1 1, otherwise Set Ynew = φk,d Biau et al. applied this method in speech recognition ([15]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abbildung: Waveform of the words boat and goat. This classifier achieved an error rate of 0.21 in this example ([?]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Summary Other basis functions can be used, e.g. wavelet basis or B-splines There exist some similar methods to the k-NN classifier, e.g. SVM These methods seem promising in lung sound classification Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Future prospects It seems possible to develop a low-cost automated classification software in the near future, which supports a doctor in diagnosis of lung diseases. Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Marburg Respiratory Sounds (MARS) database ([14]) Database contained in the year 2003 about 5000 sound recordings from different diseases All lung diseases were validated by three experts Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Abbildung: Database structure of MARS ([14]). Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Sovijärvi, A. R. A., Vandershoot, J., Earis, J.E. Standardization of computerized respiratory sound analysis CORSA Project, 2000 Sovijärvi, A. R. A., Malmberg, L. P., Charbonneau, Vandershoot, J., Dalmasso, F., Sacco, C., Rossi, M., Earis, J. E. Characteristics of breath sounds and adventitious respiratory sounds Eur Respir Rev 2000; 10: 77, 591-596 Moussavi, Zahra Fundmentals of Respiratory Sounds and Analysis Morgan & Claypool, ISBN 1598290975, 2006 Pasterkamp, H., Kraman, S. E., Wodicka, G. R. Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Respiratory Sounds - Advances Beyond the Stethoscope Am J Respir Care Med Vol. 156. pp. 974-987, 1997 Bats, B., Berger, M., Mühlhauser, I.. Klinische Untersuchung des Patienten. Schattauer, 2. Auflage, ISBN 3794512731, 1993. 3M Littmann Stethoscope Edition. 20 Examples of Cardiac and Pulmonary Auscultation 3M, 1996 http://lungdiseases.about.com/ Lu, Xiaoguang, Bahoura, Mohammed Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects An Automatic System For Crackles Detection And Classification. IEEE CCECE/CCGEI, 2006 Bahoura, Mohammed, Lu, Xiaoguang Separation of Crackles from Vesicular Sounds using Wavelet Packet Transform Hadjileontiadis, L. J., Panas, S.M. Separation of discontinuous adventitious sounds using a wavelet-based filter IEEE Trans Biomed Eng, vol. 44, no. 12, pp. 1269-1281, 1997 Du, M., Chan, F.H.Y., Lam, F.K. and Sun, J. Crackle Detection and Classification Based on Matched Wavelet Analysis Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects Proceedings - 19th International Conference - IEEE/EMBS Oct. 30 - Nov. 2, 1997 Chicago, IL. USA Styliani A. Taplidou, Leontios. J. Hadjileontiadis, Ilias K. Kitsas, Konstantinos I. Panoulas, Thomas Penzel, Volker Gross, and Stavros M. Panas On Applying Continuous Wavelet Transform in Wheeze Analysis Proceedings of the 26th Annual International Conference of the IEEE EMBS, San Francisco, CA, USA, September 1-5, 2004 Feng, J. and Sttar, F. A new automated approach for identification of respiratory sounds Multimedia and Expo, 2007 IEEE International Conference on Simon Müller Classification of Respiratory Sounds MARS Outline Introduction Methods to Detect Adventitious Lung Sounds FDA Future Prospects V. Gross, L: J. Hadjileontiadisz, Th. Penzel, U. Koehler, C. Vogelmeier Multimedia Database Marburg Respiratory Sounds (MARS) Proceedings of the 25th Annual Intemational Conference of the IEEE EMBS Cancun, Mexico - September 17-21, 2003 Biau, G., Bunea, F. and Wegkamp, M. H. Functional Classification in Hilbert Spaces IEEE Transactions on Information Theory, Vol. 51, No. 6, June 2005 Devroye, L., Györfi, L., Lugosi, G. A Probabilistic Theory of Pattern Recognition Springer-Verlag, ISBN 0-387-94618-7, 1996 Simon Müller Classification of Respiratory Sounds MARS