IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 44, NO. 12, DECEMBER 1997 1269 Separation of Discontinuous Adventitious Sounds from Vesicular Sounds Using a Wavelet-Based Filter Leontios J. Hadjileontiadis,* Student Member, IEEE and Stavros M. Panas, Member, IEEE Abstract— The separation of pathological discontinuous ad- ventitious sounds (DAS) from vesicular sounds (VS) is of great importance to the analysis of lung sounds, since DAS are re- lated to certain pulmonary pathologies. An automated way of revealing the diagnostic character of DAS by isolating them from VS, based on their nonstationarity, is presented in this paper. The proposed algorithm combines multiresolution analysis with hard thresholding in order to compose a wavelet transform- based stationary–nonstationary filter (WTST–NST). Applying the WTST–NST filter to fine/coarse crackles and squawks, selected from three lung sound databases, the coherent structure of DAS is revealed and they are separated from VS. When compared to other separation tools, the WTST–NST filter performed more accurately, objectively, and with lower computational cost. Due to its simple implementation it can easily be used in clinical medicine. Index Terms— Discontinuous adventitious sounds, multireso- lution analysis, vesicular sounds, wavelet transform, wavelet transform-based stationary–nonstationary filter. I. INTRODUCTION P ULMONARY diagnosis is often based on the analysis of acoustic pulmonary signals, since the generated acoustic energy, produced by air flow during inspiration and expiration, is highly correlated with pulmonary dysfunction. Pulmonary dysfunction is caused by anatomical or physiological changes in the pulmonary system and is characterized by changes in the acoustic properties of the various parts or organs involved [1]. Thus, when narrowing of a portion of the tracheobronchial tree occurs, turbulent flow may cause the generation of specific acoustic noise, i.e., adventitious sounds. Adventitious sounds are divided into two major classes: continuous and discontinuous sounds [2], and are only heard in pathological cases, indicating an underlying physiological malfunction. The first class contains wheezes and rhonchi, characterized by a relatively long duration (250 ms), and a sharp peak in the power spectral density function in the range of 400 Hz (wheezes) or in the range of 200 Hz or less (rhonchi). The second class contains crackles and squawks, characterized mainly by their time domain features such as a relatively short duration (20 ms), the initial deflection width (IDW), and the two cycle duration (2CD). Manuscript received July 25, 1996; revised June 16, 1997. Asterisk indicates corresponding author. *L. J. Hadjileontiadis is with the Department of EE and CE, Aristo- tle University of Thessaloniki, Thessaloniki 54006 Greece (e-mail: leon- tios@ccf.auth.gr). S. M. Panas is with the Department of EE and CE, Aristotle University of Thessaloniki, Thessaloniki 54006 Greece. Publisher Item Identifier S 0018-9294(97)07603-9. Crackles are discrete, nonmusical sounds, which, when they appear, behave as a nonstationary explosive noise superim- posed on breath sounds. Their only useful categorization is between fine and coarse crackles, with IDW 0.90 ms; 2CD 6.0 ms, and IDW 1.25 ms; 2CD 9.50 ms [1], respectively. Fine crackles (or Velcro sounds) are exclu- sively inspiratory events which tend to occur in mid-to-late inspiration and repeat in similar patterns over subsequent breaths. They have been credibly established to result from the explosive reopening of small airways that had closed during the previous expiration. They are connected either to congestive heart failure or to pulmonary fibrotic diseases such as asbestosis and idiopathic interstitial fibrosis. Coarse crackles are found in early inspiration and occasionally in expiration as well. They are of a “popping” quality (not Velcro like) and tend to be less reproducible from breath to breath. Furthermore, they apparently arise from fluid in small airways; can change pattern or clear after coughing, implying a transient character in their production mechanism; and are related with chronic bronchitis. Squawks are a combination of wheezes and crackles; although they appear as short inspiratory wheezes, they are heard in association with fine crackles (in fact, they may be initiated with a crackle). They are related to allergic alveolitis and interstitial fibrosis and are caused by the explosive opening and fluttering of the unstable airway which causes the short wheeze [3]. From the aforementioned description of discontinuous ad- ventitious lung sounds (DAS) it is evident that their sepa- ration from vesicular sounds (VS) could reveal significant information, since the structure of the DAS isolates their diagnostic character. In order to achieve automated separa- tion, the nonstationarity of DAS must be taken into account. Consequently, the use of highpass filtering fails to separate the nonstationary sounds, destroying the waveforms. Furthermore, a level slicer cannot overcome the small amplitude of fine crackles. Application of time-expanded waveform analysis in crackle time domain analysis [2], [4], results in separation; it is, however, time consuming, with large inter-observer variability. Nonlinear processing [5], obtains more accurate results, but requires empirical definition of the set of parameters of its stationary–nonstationary (ST–NST) filter. The wavelet transform (WT) provides a new perspective in analysis of lung sounds, since it can decompose them into multiscale details, describing their power at each scale and position [6]. Applying a threshold-based criterion at each scale, a filtering scheme which weights WT coeffi- cients according to signal structure can be composed. A 0018–9294/97$10.00  1997 IEEE