COL 9(6), 061002(2011) CHINESE OPTICS LETTERS June 10, 2011
Photoacoustic image reconstruction based on
Bayesian compressive sensing algorithm
Mingjian Sun (孙孙孙明明明健健健)
∗
, Naizhang Feng (冯冯冯乃乃乃章章章), Yi Shen (沈沈沈 毅毅毅), Jiangang Li (李李李建建建刚刚刚),
Liyong Ma (马马马立立立勇勇勇), and Zhenghua Wu (伍伍伍政政政华华华)
Department of Control Science and Engineering, Harbin Institute of Technology, Harbin 150001, China
∗
Corresp onding author: sunmingjian@hit.edu.cn
Received December 17, 2010; accepted January 14, 2011; posted online May 6, 2011
The photoacoustic tomography (PAT) method, based on compressive sensing (CS) theory, requires that,
for the CS reconstruction, the desired image should have a sparse representation in a known transform
domain. However, the sparsity of photoacoustic signals is destroyed because noises always exist. Therefore,
the original sparse signal cannot be effectively recovered using the general reconstruction algorithm. In
this study, Bayesian compressive sensing (BCS) is employed to obtain highly sparse representations of
photoacoustic images based on a set of noisy CS measurements. Results of simulation demonstrate that the
BCS-reconstructed image can achieve superior performance than other state-of-the-art CS-reconstruction
algorithms.
OCIS codes: 100.3020, 110.5120, 170.5120.
doi: 10.3788/COL201109.061002.
Photoacoustic imaging, in recent times, has emerged
as a promising imaging technique for biomedical
applications
[1]
. Several physiologically important
molecules, such as hemoglobin, possess a high charac-
teristic absorption; therefore, photoacoustic imaging pro-
vides superlative quality images of vasculature and hemo-
dynamic functions in vivo
[2−5]
. In photoacoustic imag-
ing, a pulsed broad laser beam illuminates the biologi-
cal tissue to generate a rapid increase in temperature.
The resultant thermoelastic expansion leads to the emis-
sion of short-wavelength pulsed ultrasonic waves. These
acoustic waves are detected by ultrasonic transducers,
and an image is then reconstructed from signals recorded
at different locations surrounding the tissue.
In photoacoustic tomography (PAT) imaging, the re-
construction algorithms existing for circular tomography
require a great number of measurements, which require
complex and expensive electronic equipments. In addi-
tion, it is almost impossible to cover the entire surface
of tissue in practice; therefore, the data can often be
acquired from limited view angles. To resolve such limit-
ing factors, Provost et al. demonstrated that the theory
of compressive sensing (CS) can be used for reconstruc-
tion in PAT by using a small number of angles
[6]
. Liang
et al. applied the CS theory to address the issue of arti-
facts in limited-view imaging and to reduce the number
of random illuminations for fast data-acquisition
[7]
. Guo
et al. incorporated the CS theory in the PAT reconstruc-
tion. Both phantom and in vivo results showed that the
CS method can effectively reduce the number of under-
sampling artifacts
[8]
.
Nonetheless, in practice, photoacoustic signals are of-
ten polluted by noises
[9]
. Noisy signals are not strictly
sparse signals, but they are compressible signals. In the
abovementioned CS-based PAT theory, basis functions
that are once added are never removed. A successful ap-
plication of CS requires that the desired image must have
a sparse representation in a known transform domain;
however, the noise destroys the sparsity of photoacoustic
signals. In such noisy conditions, the original sparse sig-
nal cannot be effectively recovered.
In this letter, a Bayesian compressive sensing (BCS)
method is employed to obtain highly sparse representa-
tions of photoacoustic images based on a set of noisy
CS measurements. It has been demonstrated, in the
sparse Bayesian learning literature, that utilization of the
relevance vector machine (RVM)
[10]
can facilitate more
effective resolution of problems in CS
[11]
.
Based on the CS method, if a given image f is com-
pressible in a transform basis function Ψ, it is possible to
perform a compressed set of measurements y: in case of
CS measurements corrupted by an approximated zero-
mean Gaussian noise n with unknown variance σ
2
, CS
measurements may be represented as
y = Φω + n, (1)
where Φ = [ϕ
1
, · · · , ϕ
N
] is an M×N matrix, based on the
assumption that M random CS measurements are made.
Therefore, if ω represents weights with the smallest N-M
(M ¿ N) coefficients set to zero, the reconstruction of f
from y reduces to estimation of the sparse weight vector
ω.
A typical method for solving such an ill-posed problem
is via the l
p
norm of ω:
˜ω = arg min
ω
{k y − Φω k
2
2
+ρ k ω k
p
}, (2)
where the scalar ρ (0 6 p 6 1) controls the relative im-
portance applied to the Euclidian error and the sparse-
ness term.
Under the common assumption of a zero-mean Gaus-
sian noise, the Gaussian likelihood model can be obtained
as
p(y
¯
¯
ω
,
σ
2
) = (2πσ
2
)
−M/2
· exp(−
1
2σ
2
k y − Φω k
2
). (3)
In the above analysis, the CS problem of inverting for
the sparse weights ω is converted into a linear-regression
1671-7694/2011/061002(4) 061002-1
c
° 2011 Chinese Optics Letters