(Chest. 2001;120:1309-1321.)
© 2001
American College of Chest Physicians
Acoustic Imaging of the Human Chest*
Martin Kompis, MD, PhD;
Hans Pasterkamp, MD and
George R. Wodicka, PhD
*
From the School of Electrical and Computer Engineering, Department of Biomedical Engineering (Drs. Kompis and Wodicka), Purdue University, West Lafayette, IN; the Department of Pediatrics and Child Health (Dr. Pasterkamp), University of Manitoba, Winnipeg, Canada; and the University Clinic of ENT, Head and Neck Surgery (Dr. Kompis), Inselspital, Berne, Switzerland.
Correspondence to: Martin Kompis, MD, PhD, University Clinic of ENT, Head and Neck Surgery, Inselspital, 3010 Berne, Switzerland; e-mail: martin.kompis{at}insel.ch
 |
Abstract
|
|---|
Study objectives: A novel method for acoustic imaging
of the human respiratory system is proposed and evaluated.
Design: The proposed imaging system uses simultaneous
multisensor recordings of thoracic sounds from the chest wall, and
digital, computer-based postprocessing. Computer simulations and
recordings from a life-size gelatin model of the human thorax are used
to evaluate the system in vitro. Spatial representations
of thoracic sounds from 8-microphone and 16-microphone recordings from
five subjects (four healthy male adults and one child with lung
consolidation) are used to evaluate the system in
vivo.
Results: Results of the in
vitro studies show that sound sources can be imaged to within 2
cm, and that the proposed algorithm is reasonably robust with respect
to changes in the assumed sound speed within the imaged volume. The
images from recordings from the healthy volunteers show distinct
patterns for inspiratory breath sounds, expiratory breath sounds, and
heart sounds that are consistent with the assumed origin of the
respective sounds. Specifically, the images support the concept that
inspiratory sounds are produced predominantly in the periphery of the
lung while expiratory sounds are generated more centrally. Acoustic
images from the subject with lung consolidation differ substantially
from the images of the healthy subjects, and localize the
abnormality.
Conclusions: Acoustic imaging offers new
perspectives to explore the acoustic properties of the respiratory
system and thereby reveal structural and functional properties for
diagnostic purposes.
Key Words: acoustics • heart sounds • image processing, computer-assisted • respiratory sounds • respiratory system • thorax
|
Introduction
|
|---|
Several
imaging methods for the human lung are available and widely used.
Conventional radiography, CT, MRI, and methods provided by nuclear
medicine are well established. Other methods, such as electrical
impedance tomography, are still under development.1
Methods using acoustic signals, most notably ultrasound, have not been
successfully developed, primarily because acoustic damping of the lung
parenchyma is prohibitively high at high frequencies.2
The absence of acoustic imaging methods may be somewhat surprising,
since alterations in the structure and function of thoracic organs that
occur in disease often give rise to measurable changes in lung sound
production and transmission. Acoustic assessment by stethoscope has
been known for almost 2 centuries3
and is a widely used
clinical method to assess these changes. In the last decade, the
availability of computer technology has prompted many research efforts
in the area of respiratory sounds and provided knowledge beyond what
has been known based on classical auscultation.4
Thoracic sounds are known to contain spatial information that can
be accurately accessed using simultaneous multimicrophone recordings
but not by successive auscultation at different locations of the
thoracic surface.5
6
The use of this additional spatial
information may lead to acoustic imaging methods beyond the relatively
simple mapping of sounds on the thoracic surface, which has been
proposed both for lung7
8
and heart sounds.9
In this study, a novel acoustic imaging method for the human
respiratory system that exploits this additional spatial information is
proposed and evaluated. The attractiveness of such a technique stems
from its noninvasive nature and its relation to both the anatomy and
physiology of the respiratory system.
|
Acoustics of the Human Thorax
|
|---|
The acoustic properties of the human thorax are complex and only
partly understood. The chest consists of at least three components of
substantially different acoustic qualities: solid tissue, airways, and
lung parenchyma. Acoustic properties of the solid components of the
thorax, such as the chest wall and the heart, are relatively
well-known. Sound speeds in these tissues are approximately 1,500
m/s,10
and damping is relatively low. In the larger
airways (ie, diameter of 
1 mm) of animal models, sound
propagates at speeds (mean ± 95% confidence limit) of 268 ± 44
m/s.11
The acoustic properties of the lung
parenchyma, which fills a substantial portion of the human thorax, is a
function of the air content of the lung.2
12
Parenchymal
sound speed was estimated to be relatively low, ie, between
23 m/s2
and 60 m/s,10
12
depending on air
content. Sound speed reaches a minimum at lung densities that are
slightly higher than those at resting volume, and increases from this
minimal value of approximately 23 m/s for both higher, and lower
densities.12
Therefore, under physiologic conditions, the
sound speed is slightly higher in the upper parts of the lung and after
inspiration. At resting volume, sound speed is likely closer to 30 m/s
than to 60 m/s. As previously noted, damping of the lung parenchyma
increases with frequency.2
At low audible frequencies, for
example 400 Hz, damping is estimated to be only from 0.5 to 1.0 decibel
per centimeter.2
Aside from these differences in acoustic properties, geometry
contributes significantly to the complexity of thoracic acoustics.
High-frequency sounds are known to travel further within the
airway-branching structure, while low-frequency sounds appear to exit
predominantly the large airways via wall motion.13
Reflections, multiple delays, and interference of acoustic signals, as
well as a left-to right asymmetry of thoracic
transmission8
14
have been described.
Respiratory sounds can be roughly grouped into breath sounds,
continuous adventitious lung sounds, and discontinuous adventitious
lung sounds. From these sound categories, only discontinuous
adventitious lung sounds or "crackles" can be easily attributed a
time of arrival for direct estimation of the sound origin. Because of
the frequency-dependent damping by the lung parenchyma, most of the
signal energy of breath sounds is concentrated at low frequencies (Fig 1
). Within the frequency band of potential diagnostic value and the given
range for the sound speed in lung parenchyma, wavelengths range from
2.3 cm at 23 m/s and 1,000 Hz to 60 cm at 60 m/s and 100 Hz. The
spatial resolution of an acoustic imaging system for the human lung can
therefore not be expected to resolve differences below approximately 2
cm. Slightly higher resolution may be obtained when the signal-to-noise
ratio is high at frequencies > 1,000 Hz, eg, at higher
airflows, in smaller chests, and over areas with increased
high-frequency contents, such as the right upper lobe.
|
Imaging Algorithm
|
|---|
Requirements for an Acoustic Imaging System for the Human Thorax
Because of the complexity of the acoustic properties of the human
thorax and practical limitations, a useful imaging system for the human
lung and the underlying imaging algorithm should meet several goals:
(1) the algorithm must be robust with respect to the acoustic
properties within the thorax, most notably to changes in sound speed;
(2) the algorithm should not rely on the measurement of the time of
arrival of lung sound components; (3) the algorithm should provide
three-dimensional data sets, which can be viewed from various
perspectives and in various formats; (4) the resulting images should be
intuitively interpretable; (5) given current sensor technology, the
number of microphones should be limited in number, preferably not
exceeding 10 to 16; and (6) the algorithm should be robust with respect
to missing microphones or noisy data in individual microphones.
Several algorithms are known to process multimicrophone
signals.15
16
17
However, since the distribution of sound
speed and damping within the thorax is complex, known algorithms such
as acoustic holography15
and similar methods cannot be
directly applied to lung sound imaging within the thorax. As time of
arrival cannot be easily and unequivocally estimated for breath sounds
and continuous adventitious lung sounds, algorithms relying on the
measurement of time of arrival have only a limited number of
applications.18
In addition, the algorithm should be able
to generate a spatial representation of intrathoracic sounds, as
opposed to the mapping of sounds on the thoracic
surface.7
8
9
A Robust Acoustic Imaging Algorithm
Based on the above-mentioned goals, an acoustic imaging algorithm
was developed. The algorithm consists of two parts: the calculation of
a three-dimensional data array, and the graphic representation of this
array.
Calculation of a Three-Dimensional Data Array:
In this step, a
data array containing a reconstructed three-dimensional distribution of
the sound sources is calculated. For any given point within the thorax,
the acoustic imaging algorithm tests the hypothesis that it contains
the only relevant acoustic source. A hypothetical source signal is
calculated by a least-squares estimation procedure to explain a maximum
of the signal variance in all microphone signals as follows. Let
xi (i = 1...M) be the positions of M
microphones on the thoracic surface and si (t) the signals
recorded at these microphones, where t represents time. Assuming a
uniform sound propagation throughout the thorax (sound speed c, damping
factor per unit length d) the signal r(y,t) emitted by
this hypothetical source at the location, y can be
estimated by solving the linear system equations:
using a least-squares fit. In the above-mentioned
equations, the left-hand side represents the microphone signals taking
into account the time delay between hypothetical source and microphone,
and the right-hand side represents the source signal including
geometric and linear damping. Assuming different hypothetical source
sites y, the thoracic volume is scanned using a step size of
1 cm. From the hypothetical source signal estimated for each location,
the portion of all microphone signals that could be explained by the
given hypothetical source only, ie, a number between 0 and
1, is calculated and stored in a three-dimensional data array,
representing the thoracic volume. Note that the algorithm
does not assume any minimum (or maximum) number of microphones or any
specific microphone arrangement. Although the number of assumptions on
the acoustic properties of the thorax was also kept as low as possible,
such assumptions could not be avoided completely. Their influence is
discussed later in this article in "Influence of Model
Assumptions."
Graphic Representation of the Data Array:
In this step, a
graphic representation of the data array that contains the
reconstructed three-dimensional distribution is generated. The data
array calculated by the above-mentioned procedure can be represented in
different ways, including CT-like images and several three-dimensional
representations. In this study, a gray-level-coded spatial
representation was selected. In these diagrams, each data point is
represented using a sphere at the appropriate location. High data
values, ie, data points at which a hypothetical sound
source can explain much of the total microphone signal variance, are
depicted in dark colors, and low data values are shown in light colors.
To resolve the problem of hidden data points by overlay, only a certain
percentage of the data points with the highest values are represented
with spheres, with the remainder of them empty.
Implementation:
As the algorithm requires the spatial
coordinates of all microphones, the microphone positions must be
measured. For this purpose, a simple stereotactic device was developed
(Fig 2
). With this device, three-dimensional coordinates on the chest surface
can be estimated from the length of three strings attached to a
pointing device with an accuracy of at least 1 cm. A program to convert
the measurement results from the stereotactic device to Cartesian
coordinates, the imaging algorithm and a program for graphic
representation were implemented off-line on a personal computer. Before
the actual imaging procedure, the recorded multimicrophone signal was
split into time segments of 100-ms duration. Each segment was first
multiplied with a Tukey-window function in the time-domain and
transformed into the frequency domain using a fast Fourier transform.
In the frequency domain, a calibration procedure was performed to
correct for the (small) differences in phase and amplitude between the
individual microphones. The signals were then transformed back into the
time domain, and only the middle 50 ms of the original 100-ms window
corresponding to the flat part of the Tukey window were used and
processed by the imaging algorithm.
View larger version (22K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2. Schematic drawing of the stereotactic device to
measure the spatial coordinates of the microphones attached to the
thoracic surface with an accuracy of better than 1 cm. The size of the
vertical mounting board is 1.1 x 0.6 m.
|
|
|
Evaluation
|
|---|
Materials and Methods
Three methods were used to evaluate the acoustic imaging
algorithm: computer simulations, imaging of a gelatin model of the
human thorax, and imaging of human subjects using different numbers of
microphones.
Computer Simulations
In these experiments, microphone signals were simulated by a
computer program rather than recorded by microphones. Sixteen
microphones were assumed to be in the same positions as were used later
in multisite recordings from human subjects: 8 microphones in the front
in two rows of 4 microphones each, and 8 microphones in the back (Fig 3
). Two independent acoustic point sources, point 1 and point 2, were
simulated within the thorax, as shown schematically in Figure 3
. The
point sources emitted white noise of equal amplitude, which was
generated by a random-number generator, and the resulting microphones
signals were calculated assuming a constant sound speed of 40 m/s for
the entire thorax. Three different simulations were performed: one with
both point sources active with the same amplitude, one where only
source point 1 was active, and one with source point 2 alone. The
simulated microphone signals were then processed by the proposed
imaging algorithm in the frequency band of 100 to 1,000 Hz, assuming
intrathoracic sound speeds of 20, 30, 40, 50, and 60 m/s.
View larger version (18K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3. Computer-simulated thoracic volume of
30 x 30 x 15 cm with the positions of the 16 microphones (squares
numbered 1 through 16) and two independent acoustic point sources
("x"-marked P1 and P2) shown.
|
|
Images of a Gelatin Model of the Human Thorax
In order to obtain a more realistic, yet well-controlled model
to evaluate the imaging algorithm, a gelatin model (J. D. DePeri,
BS; unpublished study; April 1987) of the human thorax was
constructed. Figure 4
shows the model schematically. The model consisted of a solid gelatin
wall of 2.7 cm in average thickness to mimic the chest wall. Inside
this wall, 6.9 L of foamed gelatin (3.9 L of air and 3.0 L of
gelatin) with a composite density of 430 kg/m3
was used to model lung parenchyma. In the center of the model lung
parenchyma, a soft rubber tube (diameter, 17 mm; length inside the
gelatin, 11.2 cm) was used to represent the trachea. These model
parameters were chosen to match average values for adult human subjects
in terms of density of the parenchyma,12
residual lung
volume and tracheal length,19
as well as diameter of the
trachea.20
For the recordings, white noise was introduced by a compression driver
(model 22A; Peavey Electronics; Meridian, MS) attached to the upper end
of the model trachea. By repeating the exact white noise sequence nine
times and moving 14 microphones (ECM-T150; Sony; Tokyo, Japan)
in optimized couplers,21
a virtual recording with 105
microphones arranged in a rectangular grid with a spacing of 4 cm
around the surface of the gelatin model was obtained. The composite
sound speed was measured to lie between 35 m/s and 45 m/s, and acoustic
images were reconstructed assuming a uniform sound speed of 40 m/s in
four different frequency bands: 50 to 2,000 Hz, 125 to 250 Hz, 250 to
500 Hz, and 500 to 1,000 Hz.
Acoustic Images From Recordings of Human Subjects
Subjects:
Four male healthy nonsmokers (subjects 1, 2, 3,
and 4; age range, 30 to 34 years) participated in this study. Their
chest circumferences ranged from 94 to 107 cm (average, 99 cm), and
their weights ranged from 68 to 91 kg (average, 74 kg). Subjects 1 and
2 were studied using 8-microphone recordings; 16-microphone recordings
were used for subjects 3 and 4.
In addition, a child (subject 5; age, 10 years; chest
circumference, 80 cm; weight, 50 kg) with pneumonia due to
blastomycosis, resulting in a severely consolidated lower left lung,
underwent a 16-microphone recording.
The study protocol using 8 microphones was approved by the Purdue
University Committee on the Use of Human Subjects, while the protocol
with 16 microphones was approved by the Committee on the Use of Human
Subjects in Research, Faculty of Medicine, University of Manitoba. All
subjects, and the parents of subject 5, gave their informed consent to
participate in the study.
Experimental Apparatus:
Figure 5
is a schematic drawing of the experimental setting. Subjects were
seated in a soundproof room (model 108192; Industrial Acoustics; Bronx,
NY) with 8 microphones or 16 microphones (ECM-T150; Sony) in
optimized couplers21
attached to the chest. The microphone
signals were amplified and band-pass filtered between 100 Hz and 1,000
Hz by custom-built filter-amplifier units (four-pole Butterworth
filters, amplification 300). The resulting signals were digitized at a
sampling rate of 10,240 Hz using a personal computer-based
analog-to-digital converter (resolution 12 bit; model DT-2831 G; Data
Translation; Marlboro, MA). The subjects breathed through a
pneumotachograph (Fleisch No. 3) and observed their airflow signal on
an oscilloscope.
Experimental Protocol:
To establish feasibility, the initial
recordings were performed at Purdue University using eight microphones
that reflected the maximum number of signal conditioning channels
available. Subsequent recordings were performed at the University of
Manitoba to both exploit the 16-channel measurement capabilities
therein and allow studies to be performed on patients as well as
healthy subjects. For eight-microphone recordings, the microphones were
placed in the corners of two rectangles lying on the anterior and
posterior chest surfaces and measuring 18-cm horizontally by 16-cm
vertically. They were centered around the median plane, with the upper
anterior microphones overlying the second intercostal space (Fig 6
).
For 16 microphone recordings, the microphones were arranged as
follows: 8 microphones were placed in the front, arranged in a regular
pattern of 4 horizontal microphones by 2 vertical microphones.
Distances between adjacent microphones were 10 cm, both in the
horizontal and vertical direction for the recordings with adult
subjects (subject 3 and subject 4), and 8 cm for the child (subject 5).
The resulting rectangles were centered around the median plane of the
subject, with the middle two microphones of the upper row being placed
on the second intercostal space. The remaining eight microphones were
placed on the back of the thorax, using the same four-by-two
arrangement of the microphones and the same distances. Corresponding
microphones in the front and the back of the subjects were at the same
horizontal level. This microphone arrangement corresponds to that used
in the computer simulations (Fig 3)
. Microphone positions were measured
in three dimensions using the stereotactic device depicted in Figure 2
.
Respiratory sound data were acquired for at least six complete breath
cycles at target airflows of ± 30 mL/s/kg (corresponding to
approximately ± 2 L/s for adult subjects). Background noise and heart
sounds (16-microphone recordings only) were then measured while the
subjects stopped breathing at end expiration with their glottis open.
Signal Analysis and Spatial Representation:
Signals were
analyzed separately for inspiration and expiration (all recordings), as
well as for first heart sounds (16-microphone recordings only). For
inspiratory and expiratory sounds, only segments with flow rates within
± 20% of the target flow14
were analyzed. For heart
sound imaging, segments from the breath-hold section of the recordings
were used. For each spatial representation, time segments of 0.1 s
were used. To produce averaged images, data arrays from 4 to 12 time
segments of 50-ms duration were averaged before graphic representation.
|
Results
|
|---|
Imaging of Sound Sources
Figure 7
shows a spatial view of acoustic images from the computer simulations,
showing the three cases of source point 1 active only (Fig 7
,
left, a), source point 2 active only (Fig 7
,
middle, b), and both sources active (Fig 7
,
right, c). Each image represents the view of a
volume of 30 x 30 x 15 cm, overlying a (simulated) thorax. Data
points are displayed on a regular three-dimensional grid with a spacing
of 1 cm, and represented by spheres. Each data point can have a value
between 0 and 1, according to the portion of the total energy of all 16
microphone signals that could be explained by placing one single
optimal acoustic point source at that point. These values are
gray-level coded: dark points denote a large amount of explained
variance or a high probability for the presence of an acoustic source,
and bright or no points denote areas where only a small part of all
microphone signals could be explained by placing a single sound source
at this location. Each image is accompanied by a legend for the linear
gray-scale used. The higher number denotes the value for black data
points and at the same time shows the maximal value reached within the
imaged volume. The lower number denotes the cutoff value below which
data points are omitted to allow better visualization. In Figure 7
, the
cutoff value has been chosen to be half of the maximal value reached
within the imaged volume. It can be seen that both point sources are
clearly identified and correctly located by the algorithm in all three
simulations. The accuracy of the localization is better than 1 cm,
ie, more accurate than the grid width of the data
representation. Note, however, that this does not imply that resolution
between two acoustic sources is better than the 2 cm mentioned earlier.
The value of the highest data points denotes correctly that in the
first case (Fig 7
, left, a) and the second case
(Fig 7
, middle, b), only one acoustic source was
found (values close to 1.0), while in the third case (Fig 7
,
right, c), two sources of equal strength were
found (maximal portion of total explained microphone variance close to
0.5 for each of the two sources).
View larger version (13K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 7. Reconstructed acoustic images from the computer
model of the human thorax (Fig 3)
showing a spatial view of the
simulated thoracic volume of 30 x 30 x 15 cm with
(left, a) acoustic source point 1 active
only, (middle, b) acoustic point source 2
active only, and (right , c) both sources
active. The gray-scale-coded value of the imaged data points shows the
portion of the total signal variance in all 16 microphone signals that
could be explained by placing a single hypothetical point source at
that location.
|
|
Figure 8
shows the images from the gelatin model. The microphone signals were
analyzed in four different frequency bands: 50 to 2,000 Hz, 125 to 250
Hz, 250 to 500 Hz, and 500 to 1,000 Hz. To form each of the four images
in Figure 8
, 10
separate reconstructions of a time segment of 50 ms
were averaged. In each image, only 3% of all data points with the
highest values are shown. Because of the high number (105) of
microphones and the substantial spatial dimensions of the model
trachea, the maximal explained portion of the total signal energy of
all 105 microphones was very low, typically < 1%. The broad-band (50
to 2,000 Hz) and medium-band (250 to 500 Hz) images show the model
trachea in its correct length and placement to within 2 cm. For the
low-frequency band (125 to 250 Hz), the trachea is broader, presumably
because of the long wavelengths (16 to 32 cm at sound speed of 40
m/s) associated with these signals. In the high-frequency band
(500 to 1,000 Hz) the distribution becomes irregular, probably due to
the higher damping and therefore lower signal-to-noise ratio at these
frequencies, or due to small irregularities in the gelatin that become
evident only at these short wavelengths. From these images, it can be
hypothesized that low-frequency sounds (125 to 250 Hz) exit the model
trachea already relatively close to the sound source, whereas
high-frequency sounds travel further and exit toward the end of the
tube. This observation is in agreement with results from studies with
human subjects, where high-frequency sounds were found to travel
further within the airways before exiting to the parenchyma than
low-frequency sounds.13
View larger version (31K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 8. Acoustic images from the life-size gelatin model
of the human lung in four different frequency bands. Each image
represents a volume of 26 x 26 x 25 cm. The image of the model
trachea (Fig 4)
is best discernible in the frequency bands 50 to 2,000
Hz and 250 to 500 Hz.
|
|
View larger version (75K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 10. Reconstructions from the 16-microphone recordings
of two healthy male subjects for forced inspiration, expiration, and
first heart sounds during breath hold. Imaged volumes are
30 x 30 x 21 cm (subject 3) and 30 x 30 x 23 cm (subject
4).
|
|
Plausibility of Acoustic Images of Healthy Human Subjects:
Comparison With Current Notions on the Site of Production of Thoracic
Sounds
Figure 9
shows reconstructions from eight-microphone recordings of subjects 1
and 2. Images from 12 inspiratory and expiratory segments of 50 ms in
duration each were averaged to obtain the individual images. Each
image represents a volume of 30 x 30 x 20 cm (subject 1) and
30 x 30 x 16 cm (subject 2), respectively. The viewing angle is
the same as that in the thoracic outline provided in Figure 9
. Although
there are differences between the images of the two subjects,
inspiratory sounds are estimated to be produced predominantly in the
periphery of the lungs for both subjects, while expiratory sounds
appear to originate more centrally. This is in agreement with current
concepts on the site of production of respiratory sounds.5
In some of the images of Figure 9
, a central vertical structure can be
identified in the upper part of the imaged volume, which might be an
acoustic image of the trachea.
View larger version (75K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 9. Reconstructions from the eight-microphone
recordings of two healthy male subjects during forced inspiration and
expiration, respectively. Imaged volumes are 30 x 30 x 20 cm
(subject 1) and 30 x 30 x 16 cm (subject 2), respectively. The
viewing angle is shown by the thoracic outline in the upper- right
corner. Note the more peripheral sound production during inspiration
when compared to expiration in the same subject.
|
|
Figure 10
depicts the spatial representations of the three different sound
categories under investigation for the 16-microphone recordings of
subjects 3 and 4. The blocks represent a volume of 30 x 30 x 21
cm (subject 3) or 30 x 30 x 23 cm (subject 4) overlying the
thorax of the subjects. The viewing angle is from the front, right
side, from the same direction as shown for the thoracic outline in the
upper right corner. The microphone positions are not shown. However,
the represented volumes were chosen in such a manner that anterior and
posterior microphones are lying approximately in the front and rear
plane of the depicted volume. In the vertical dimension, the volumes
shown are centered around the center of gravity of all 16 microphones.
During inspiration, areas with high data values (ie,
areas in which a hypothetical sound source explains a significant
portion of the total microphone signal variance) are concentrated
mainly in the front of the thorax. A major contribution is made by a
dark spot in the upper half of the thorax, just left from the median
plane. Another, smaller maximum, which is partly hidden by the front
maxima, can be found in the back in both subjects. During expiration,
dark areas can be found more centrally in the thorax. For both
subjects, there is a significant dark region around the second
intercostal space in the front, right hemithorax.
For the first heart sounds, a major part of the microphone signal
variance can be explained by assuming sources in the front left lower
thorax, ie, at the approximate location of the heart.
Spatial representations from first heart sounds are similar for both
subjects.
Influence of Model Assumptions
Figure 11
shows five different reconstructions from the computer-simulated data
with two active sound sources (all corresponding to the image in Fig 7
,
right, c), with the exception that the sound
speed assumed by the imaging algorithm was varied from 20 to 60 m/s in
steps of 10 m/s. It can be seen that the acoustic sources are imaged
most compactly and localized most accurately if the assumed sound speed
matches the actual sound speed (40 m/s). If the sound speed assumed by
the imaging algorithm is too low, artifacts appear, mimicking several
scattered yet weak sound sources. If the sound speed is overestimated,
the images of sound sources near the microphones become more prominent
than more centrally located sound sources. Overestimation of sound
speed causes less deviation from the image based on true sound speed
than underestimation. This is also documented by the highest maximal
data value at 50 m/s and 60 m/s (0.42 and 0.35) than at 20 m/s and 30
m/s (0.28 and 0.35).
View larger version (27K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 11. Acoustic images from the computer-simulated model
of the human thorax with two acoustic point sources (Fig 3)
. The sound
speed assumed by the imaging algorithm was varied between 20 m/s and 60
m/s, while the actual sound speed of the model was kept constant at 40
m/s.
|
|
Acoustic Images From the Subject With Lung Consolidation
Figure 12
shows the acoustic images resulting from 16-microphone recordings from
the child with pneumonia due to blastomycosis, and a severely
consolidated lower left lung, and Figure 13
shows a corresponding chest radiograph. The imaged volume in Figure 12 is 30 x 30 x 21 cm, and the images were reconstructed from the
frequency band between 100 Hz and 1,000 Hz. Images have been averaged
over 17 inspiratory and 17 expiratory recorded segments of 50 ms in
duration each. Thirty percent of all data points are shown in the
images in Figure 12
, and front as well as rear views of the data are
shown. An area with high data values can be seen in the back, both
during inspiration and expiration, corresponding to the consolidated
portion of the lung.
View larger version (64K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 12. Reconstructions from the 16-microphone recordings
of subject 5 (10-year-old child with pneumonia due to fungus infection
[blastomycosis] resulting in a severely consolidated lower left
lung). The imaged volume is 30 x 30 x 21 cm. The images in the
left column show the front view, and the images in the right column
show the rear view of the subject, according to the thoracic outlines
on the top of each column. Note the active area (high data values) in
the back, both during inspiration and expiration, corresponding to the
lung consolidation.
|
|
|
Discussion
|
|---|
The results of this investigation suggest that the proposed
algorithm can image sound sources adequately, both in the computer
simulations (accuracy better than 1 cm) and in a life-size gelatin
model of the human thorax (accuracy of 2 cm). Due to the heterogeneity
of lung tissue and person-to-person differences, eg, in
lung-muscle-fat ratios the same accuracy is not expected for humans.
Nevertheless, these findings are encouraging, as they document that
useful spatial information can be extracted from recordings with as few
as 16 microphones.
Spatial representations from recordings with human subjects (Fig 9
, 10
, 12)
show similar patterns for all subjects during inspiration,
expiration, and, where evaluated, first heart sounds. Images during
inspiration and expiration confirm the hypotheses that inspiratory
sounds are produced predominantly in the periphery of the lung, while
expiratory sounds are generated more centrally.5
8
During expiration, a major sound source appears to lie close to the
second intercostal space at the upper right anterior portion of the
chest wall for four of the five subjects (subjects 2, 3, 4, and 5),
confirming a finding from earlier studies.8
22
It can be
speculated that this phenomenon is caused either by preferential
transmission of tracheal noise to the right or by particular turbulence
patterns in the right upper lobe, which typically has a more sharply
angulated origin of bronchus. No such maximum in the upper right chest
could be found in subject 1. The spatial representation algorithm
locates first heart sounds roughly at the expected location of the
heart itself (subjects 3 and 4; Fig 10
).
Images from 8-microphone and 16-microphone recordings (Fig 9
, 10) show certain differences. The influence of the number of microphones (8
or 16) and their placement is incompletely known. The optimal placement
of microphones and its influence on the images needs to be investigated
further. Because of the different placement of the microphones, the
8-microphone and 16 microphone recordings cannot be precisely compared
to each other.
Images from recordings of lung sounds in the child with pneumonia (Fig 12)
show a substantially increased acoustic activity at the location of
the abnormality. This is most probably caused by the higher sound speed
and lower attenuation in the consolidated tissue, when compared to that
in the remainder of the lung. Figure 12
illustrates that localized lung
disease can be observed in acoustic images and that acoustic images
from diseased lungs appear distinctly different from acoustic images
obtained from healthy subjects (Fig 9
, 10)
.
The computer simulations (Fig 11)
indicated that the proposed
imaging algorithm is reasonably robust with respect to the underlying
assumption of the sound speed in the imaged volume. The results suggest
that in cases where the exact sound speed is not known, it may be
better to overestimate than to underestimate to optimize the resulting
localization accuracy.
In addition to static images of the human thorax, the proposed imaging
algorithm can also be used to produce real-time imaging. For the
currently employed, nonoptimized implementation of the imaging
algorithm running on nonspecialized computer hardware, the computation
load is too high for such an application. However, using specialized
hardware (eg, multiple fast digital signal processors) and
optimized code, a real-time system could be realized today at a
fraction of the cost of standard clinical imaging systems.
Because this is the first study (to our knowledge) of the proposed
acoustic imaging method, many questions regarding its functionality and
potential utility remain. Recordings from additional healthy subjects
and patients with different lung diseases and different auscultation
findings would provide a wealth of useful information. In addition,
future efforts also include investigations into improved imaging
algorithms and recording systems with a goal to allow simultaneous
recordings with > 16 microphones.
It is unlikely that acoustic imaging of the thorax will become a method
competing with static CT or MRI. However, it might open new diagnostic
possibilities since it provides dynamic information. Acoustic imaging
can be performed during normal breathing, ie, no
breath-holding maneuvers are required. Lastly, advances in sensor
technology, such as an acoustic array replacing multiple individual
microphones, might allow for patient monitoring over extended periods
of time.
|
Summary
|
|---|
A novel method for acoustic imaging of the human respiratory
system was developed and evaluated. The system uses simultaneous
multimicrophone recordings of thoracic sounds from the chest wall and a
digital, computer-based postprocessing system. Computer simulations and
recordings from a life-size gelatin model of the human thorax indicated
that sound sources can be imaged to within 2 cm, and that the proposed
algorithm is reasonably robust with respect to changes in the assumed
sound speed within the imaged volume. Spatial representations of
thoracic sounds from recordings from four subjects depict distinct
patterns for inspiratory breath sounds, expiratory breath sounds, and
first heart sounds that are consistent with the assumed origin of the
analyzed sounds. Acoustic images from a subject with lung consolidation
were distinctly different from images of healthy subjects and localized
the abnormality.
|
Acknowledgements
|
|---|
We thank Professor D. A. Rice, Department of
Biomedical Engineering, Tulane University, for his help and advice
concerning the gelatin model of the thorax.
|
Footnotes
|
|---|
This work has been supported by the Swiss and the US National Science
Foundations, the Ciba-Geigy-Jubiläums-Foundation, and the
Children’s Hospital of Winnipeg Foundation.
Received for publication October 20, 2000.
Accepted for publication March 8, 2001.
|
References
|
|---|
-
Bronzino, JD (1995) The biomedical engineering handbook. CRC Press Boca Raton, FL.
-
Wodicka, GR, Stevens, KN, Golub, HL, et al (1989) A model of acoustic transmission in the respiratory system. IEEE Trans Biomed Eng 36,925-934[CrossRef][Medline]
-
Laënnec, RTH (1819) De l’auscultation médiate ou trité du diagnostic de maldies des poumons et du coeur, fondé principalement sur ce nouveau moyen d’exploration. Brosson et Chaudé Paris, France.
-
Pasterkamp, H, Kraman, SS, Wodicka, GR (1997) Respiratory sounds: advances beyond the stethoscope. Am J Respir Crit Care Med 156,974-987[Free Full Text]
-
Kraman, SS (1980) Determination of the site of production of respiratory sounds by subtraction phonopneumography. Am Rev Respir Dis 122,303-309[Medline]
-
Kompis, M, Wodicka, GR (1995) Coherence of inspiratory and expiratory breath sounds as a function of inter-microphone distance. Proc Annu Int Conf IEEE Eng Med Biol Soc 17,971-972
-
Benedetto, G, Dalmasso, F, Spagnolo, R (1988) Surface distribution of crackling sounds. IEEE Trans Biomed Eng 35,406-412[CrossRef][Medline]
-
Kompis, M, Pasterkamp, H, Oh, Y, et al (1997) Distribution of inspiratory and expiratory respiratory sound intensity on the surface of the human thorax. Proc Annu Int Conf IEEE Eng Med Biol Soc 19,2047-2050
-
Leong-Kon, D, Durand, LG, Durand, J, et al (1998) A system for real-time cardiac acoustic mapping. Proc Annu Int Conf IEEE Eng Med Biol Soc 20,17-20
-
Rice, DA (1985) Transmission of lung sounds. Semin Respir Med 6,166-170
-
Rice, DA, Rice, JC (1987) Central to peripheral sound propagation in excised lung. J Acoust Soc Am 82,1139-1144[CrossRef][ISI][Medline]
-
Rice, DA (1983) Sound speed in pulmonary parenchyma J Appl Physiol: Respir Environ Exer Physiol 54,304-308
-
Wodicka, GR, Aguirre, A, DeFrain, PD, et al (1992) Phase delay of pulmonary acoustic transmission from trachea to chest wall. IEEE Trans Biomed Eng 39,1053-1059[CrossRef][ISI][Medline]
-
Pasterkamp, H, Patel, S, Wodicka, GR (1997) Asymmetry of respiratory sounds and thoracic transmission. Med Biol Eng Comp 35,103-106[ISI][Medline]
-
Espeland, M, Løgberg, OJ, Rustad, R (1995) Full field tomographic reconstruction of sound fields using TV holography. J Acoust Soc Am 98,280-287
-
Rasmussen, P (1995) Spatial transformation of sound fields. Sound Vibration May,18-23
-
Kompis, M, Oberli, M, Brugger, U (2000) A novel real-time noise reduction system for the assessment of evoked otoacoustic emissions. Comput Biol Med 30,341-354[Medline]
-
Murphy R. Method and apparatus for locating the origin of intrathoracic sounds. US Patent 5 844 997, December 1, 1998
-
Habib, RH, Chalker, RB, Suki, B, et al (1994) Airway geometry and wall mechanical properties estimated from subglottal input impedance in humans. J Appl Physiol 77,441-451[Abstract/Free Full Text]
-
Horsfield, K, Cumming, G (1968) Morphology of the bronchial tree in man. J Appl Physiol 24,373-383[Free Full Text]
-
Wodicka, GR, Kraman, SS, Zenk, GM, et al (1994) Measurement of respiratory acoustic signals: effect of microphone air cavity depth. Chest 106,1140-1144[Abstract/Free Full Text]
-
Kompis, M, Pasterkamp, H, Oh, Y, et al (1998) Spatial representation of thoracic sounds. Proc Annu Int Conf IEEE Eng Med Biol Soc 20,1661-1664
This article has been cited by other articles:
|
|
|
|
 
J. A. Fiz, J. Gnitecki, S. S. Kraman, G. R. Wodicka, and H. Pasterkamp
Effect of Body Position on Lung Sounds in Healthy Young Men
Chest,
March 1, 2008;
133(3):
729 - 736.
[Abstract]
[Full Text]
[PDF]
|
|
|
|
|
|
|
H. Kiyokawa and H. Pasterkamp
Volume-dependent variations of regional lung sound, amplitude, and phase
J Appl Physiol,
September 1, 2002;
93(3):
1030 - 1038.
[Abstract]
[Full Text]
[PDF]
|
|
|
|
|
|
|
T. Bergstresser, D. Ofengeim, A. Vyshedskiy, J. Shane, and R. Murphy
Sound transmission in the lung as a function of lung volume
J Appl Physiol,
August 1, 2002;
93(2):
667 - 674.
[Abstract]
[Full Text]
[PDF]
|
|
|
TOP
Abstract
Introduction
