Solunum seslerinin analizi ve sınıflandırılması

Abstract

Akciğer hastalıklarının tanısında en cok kullanılan yöntemler, dinleme, ışınçekim ve işlev testi yöntemleri dir. Bu yöntemler arasında ucuzluğu, kolaylığı ve herhangi bir cerrahi müdahele gerektirmemesi açısından en ilgi çekeni dinleme yöntemidir. Ancak soluk seslerinin dinlenmesi kolay bir yöntem olmasına karşın bir çok doktor tarafından tanı için yeterli sayılmamaktadır. Bunun nedenlerinin başında hastalıklı akciğerler için tanımlanan seslerin nesnel değil öznel olması gelmektedir, örneğin, astımda duyulan müziksel sesler ıslık, akciğer yangısında duyulan ani patlama sesleri çıtırtı diye tanımlanmaktadır [1]. öznel tanımlamaların ortaya çıkardığı problemleri ortadan kaldırmak için en akılcı yaklaşım, hem normal hem de hastalıklı seslerin parametrik gösterimlerini bulmak, daha sonra bu parametreleri sesleri sınıflandır ma ve hastalığı tanıma amacıyla kullanmaktır. Bu çalışmanın amacı sık görülen akciğer ve solunum yolları hastalıklarında duyulan solunum seslerinin bilgisayar yardımıyla analizi ve sınıflandırılmasıdır. Bu amaçla göğüs kafesinden duyulan solunum sesleri bilgisayara kaydedilmiş, daha sonra kaydedilen verilere özbağlanımlı (AR) modelleme uygulanmıştır. Sesleri sınıflandırmak için Itakura uzaklık ölçüsü kullanılmıştır. Yapılan deneyler önerilen sınıflandırıcının, yüzde 85'lik bir başarı oranı ile zatürreeli sesleri sağlıklı seslerden ayırabildiğini göstermiştir.

Auscultation, function testing and roentgenography are the three widely used methods in the identification and diagnosis of lung disease and respiratory disorder. Among these, auscultation is the most attractive method in that it is cheap, simple and non-invasive. The basis of the method is to stimulate the chest by an acoustical input and then to auscultate the sounds on the chest wall and trachea. The sound source itself could be generated in the chest through breathing at and above a certain flow rate, through percussion, or with speech utterrances. The pattern variations occuring in these sounds provide to the physician important information about the pathological situations, i.e. lung lesions, obstruction and collapse of the bronchial paths, the status of the alveoli etc. However auscultation is regarded by many physicians of low diagnostic value. The reason of this approach to auscultation is the subjective definitions of adventitous lung sounds. For example, the musical' sounds heard in asthma are called wheezes, or suddenly bursting sounds heard in pulmonary fibrosis are called crackles. An experiment conducted among the qualified physicians [2] has revealed significant variability between the observers, such that the diagnostic opinions were falling midway between chance and full agreement on almost all of the symptoms. The difficulty comes from the uncertainty in relating and transferring subjective qualifications and sound assesments, and the confusing terminology [3]. In order to avoid the problems of subjective definitions that change from one physician to another, a reasonable approach is to use a parametric representa tion of both normal and abnormal sounds and then, use these parameters to compare and classify lung sounds for diagnostic purposes. This method, though seems easy +o implement in theory, is difficult to achieve in practical applications, as these waveforms and their spectral content may depend upon various conditions, such as flow rate, regional volume distribution, surrounding tissue, position of the patient, the stage of the illness. VI Respiratory sounds are divided roughly into two classes, breath sounds and adventitous sounds. Breath sounds are low-pitched, non-musical rustlings that are synchronous with the movement of air in and out of the lungs, and they arise as the flow of air changes from laminar to turbulent above a certain flow rate. Normal breath sounds themselves can be classified into four types called vesicular, bronchial, bronhco-vesicular and tracheal [4], Their amplitudes, however, have been observed to display significant variations among sites of measurement in any one patient, as well as across patients. The above mentioned breath sounds are considered normal if heard over certain areas of the throax, and abnormal otherwise. There are, however, several types of breath sounds which, when present, always indicate abnormality. For example, the cogwheel breath sound which contains several short pauses during the inspiratory phase is thought to be caused by nonuniform inflation of normal alveoli. Asthmatic breath sound is characterized by its prolonged expiratory phase and by its short inspiratory phase. It is the result of obstruction of airways and is heard primarily in patients with asthma, but also in patients with bronchitis and emphysema. The adventitous sounds, on the other hand, are much more varied, though often a more direct association exists between these sounds and their physiological causes. These sounds have been described by various adjectives such as wet, crepitant, crackling, stridor, tinkling, narghile, fricative, amphoric, cavernous, etc. For example consecutive fine crackles at the beginning and terminal stages of pneumonia, with the trickled exuded matter causing alveolar walls to stick together in expirium, while they snap out in inspirium, causing crepitation. Wet rales or subcrepitant sounds, on the other hand, occur both in inspirium and expirium. Typical occurences of wet rales are with acute, chronic bronhcitis, lung abscissa, diffuse fibrosis, bronchiectasis, diffuse interstitial lung illnesses. Fricative sounds occur if serous membranes have lost their lubricant fluid or if there exists fibrous exuda. Stridors are whistling type of sounds that occur in laryngeal oedema or acute laryngitis, chronic mediastinis, aortal aneurisma. Wheezing sounds seen in bronchial asthma and chronic, acute bronchitis are due to the constriction of bronchial lumens which narrows the air pathways due to muscle spasms or mucosa swelling. The goal of this work is the early detection of pneumonia and eventually other lung diseases with the VI i help of signal processing and computerization methods. For this purpose, a data acquisition and measurement system, shown in Figure 1, was used [8]. Respiratory sounds were preprocessed by a bandpass filter in the 80 -2000 Hz band. The lower cutoff frequency was set to eliminate the heart sounds and muscle friction noise. A 4th order Butterworth highpass filter and an 8th order Bessel lowpass filter were designed, both with almost linear phase characteristics. A preamplifier with a gain of 45 dB and flat frequency response was implemented before the analog-to-digital (A/D) converter. The signal was digitized with 12 bit accuracy (MacAdiosII Jr) at 5 kHz sampling rate and inputted into the computer (Apple Macintosh II-cx). The quality of the sound was monitored before conversion with headphones (Philips SBC 484). In order to adjust the flow rate and synchronize on the inspiration-expiration cycles, the flow signal was also recorded by a Fleisch type flowmeter (Gould Godart No. 2) [5]. computer i H flowmeter - subject N '_ "1 - >| HPF LPF AMP ADC J he adphone e Figure 1. Measurement set up for respiratory sound monitoring and recording. In order to avoid friction noise, and to isolate the ambient noise, the microphones were encapsulated within manufactured plastic casings, and then attached to the chest wall and trachea with an adhesive tape. This type of air coupling also makes the measurement less dependent upon the static force applied [6]. The microphones were Panasonic P9930 with a bandwith of 20-18000 Hz. The measurement sites were right basilar, left basilar, and tracheal. From each of the three sites recordings of 12.8 s duration were made, a period covering 4-5 respiration cycles. Flow signal was also recorded to synchronize on the inspiratory and VI 11 expiratory cycles. The recorded respiratory data was divided into frames of 256 samples and weighted with a Hamming window. Succsessive frames had 64 samples in common or overlapped by 25%. This produced a smoothing effect and prevented spurious leaps in the estimated parameters. Assuming that classes of respiratory sounds can be modelled as white-noise driven all-pole systems, an autoregressive (AR) model was built for each class of signal as: srİ,(n) = I<" «;nt«-kJ +ej"(nî ; n = 1,.,.,N k«i ' r = insp,exp (1) where s (n) was the respiratory sound signal observed at time n, a was the k'th LPC coefficient, p was the model order, N was the frame length and e (n) was the one-step prediction error. The superscript i denoted the case of the subject and f denoted the flow cycle. The major concern in selecting the model order p was the class separability, an order of p = 8 seemed adequate. As inspiration and expiration cycles display different spectral characteristics, seperate AS models were applied to each cycle. Durbin's algorithm was used for the calculation of AR coefficients. The discrimination between cycles was achieved using flowmeter data. For this purpose, a modified zero crossing method was used. The calculated AR coefficients from data sets were labeled through their concerning flow cycles. Classification was based on the k-nearest neighbor algorithm employing the p AR coefficients as the feature vectors and the Itakura (I) distance measure [7] between the classes, which is specified as: Y by s sT b r "£i ~1'1 -" ~n -1-1 -j V T T ) a s s a Lı. - m - n - n - m C ? l09 l-==5 I (2) n o 1 IX expiratory cycles. The recorded respiratory data was divided into frames of 256 samples and weighted with a Hamming window. Succsessive frames had 64 samples in common or overlapped by 25%. This produced a smoothing effect and prevented spurious leaps in the estimated parameters. Assuming that classes of respiratory sounds can be modelled as white-noise driven all-pole systems, an autoregressive (AR) model was built for each class of signal as: srİ,(n) = I<" «;nt«-kJ +ej"(nî ; n = 1,.,.,N k«i ' r = insp,exp (1) where s (n) was the respiratory sound signal observed at time n, a was the k'th LPC coefficient, p was the model order, N was the frame length and e (n) was the one-step prediction error. The superscript i denoted the case of the subject and f denoted the flow cycle. The major concern in selecting the model order p was the class separability, an order of p = 8 seemed adequate. As inspiration and expiration cycles display different spectral characteristics, seperate AS models were applied to each cycle. Durbin's algorithm was used for the calculation of AR coefficients. The discrimination between cycles was achieved using flowmeter data. For this purpose, a modified zero crossing method was used. The calculated AR coefficients from data sets were labeled through their concerning flow cycles. Classification was based on the k-nearest neighbor algorithm employing the p AR coefficients as the feature vectors and the Itakura (I) distance measure [7] between the classes, which is specified as: Y by s sT b r "£i ~1'1 -" ~n -1-1 -j V T T ) a s s a Lı. - m - n - n - m C ? l09 l-==5 I (2) n o 1 IX breath cycles, were used in the classification process. The performance of the classifier was calculated using leave-one-out method, which provides good estimates of the probability of error in case of small sample populations. The experiments showed that suggested classifier could distinguish the pneumonic breath cycles providing a percentage of correct classifications of 85 %.