Tikhonov-regularization-based projecting
sparsity pursuit method for fluorescence
molecular tomography reconstruction
Jiaju Cheng (成家驹) and Jianwen Luo (罗建文)*
Department of Biomedical Engineering, School of Medicine, Tsinghua University, Beijing 100084, China
*Corresponding author: luo_jianwen@tsinghua.edu.cn
Received September 9, 2019; accepted September 29, 2019; posted online December 23, 2019
For fluorescence molecular tomography (FMT), image quality could be improved by incorporating a sparsity
constraint. The L1 norm regularization method has been proven better than the L2 norm, like Tikhonov regu-
larization. However, the Tikhonov method was found capable of achieving a similar quality at a high iteration
cost by adopting a zeroing strategy. By studying the reason, a Tikhonov-regularization-based projecting sparsity
pursuit method was proposed that reduces the iterations significantly and achieves good image quality. It was
proved in phantom experiments through time-domain FMT that the method could obtain higher accuracy
and less oversparsity and is more applicable for heterogeneous-target reconstruction, compared with several
regularization methods implemented in this Letter.
Keywords: fluorescence molecular tomography; sparsity pursuit; Tikhonov regularization; good image
quality; high efficiency.
doi: 10.3788/COL202018.011701.
Fluorescence molecular tomography (FMT) based on
fluorochrome emission is capable of detecting in-vivo
biological activities
[1]
and it is a sensitive functional imag-
ing method that could quantitate fluorochrome concentra-
tions at femtomolar levels
[2]
. However, due to its limited
penetration in biological tissues, there are not many clini-
cal applications of FMT but instead extensive applications
in drug development
[3,4]
and cancer studies like the tumor-
igenicity study of glioblastoma xenografts in immunodefi-
cient mice
[5]
. But it may be applied to imaging of the
human wrist like in photoacoustic tomography
[6]
.
Time-domain (TD) FMT utilizing pulsed excitation
sources is one of the three approaches in FMT while
the other two are continuous-wave (CW) FMT utilizing
steady excitation sources and frequency-domain (FD)
FMT utilizing modulated excitation sources. Fluorescence
yield and fluorescence lifetime are the two primary param-
eters while lifetime is proved an excellent probe for local
micro-environmental sensing
[7]
. Simultaneously recovering
the fluorescent yield and lifetime distributions is only
possible in TD-FMT and there are methods developed
for that purpose
[8,9]
. Moreover, only for TD-FMT, it is
possible to use early photons to achieve higher spatial
resolution and stability
[10,11]
.
Because of the strong scattering of the light in biological
tissues, the reconstruction of FMT would be extremely
ill-conditioned, limiting the quality of the reconstructed
images. To improve the image quality, a conventional
way is to incorporate a sparsity constraint in the inverse
problem, adopting regularization methods such as Tikho-
nov regularization and L1-norm regularization. Although
it is straightforward to utilize Lp-norm (0 ≤ p < 1) regu-
larization, the non-convex Lp-norm regularization could
not be directly used in FMT reconstruction. But by
applying a certain strategy, non-convex Lp-norm regulari-
zation could be solved by transferring into L1-norm to
achieve better sparsity
[12–14]
. Among these, Tikhonov regu-
larization is widely used and can be easily solved by iter-
ative algorithms, while its solution is usually not sparse
enough. However, according to our previous work
[15]
, with
a large number of iterations, an iterative Tikhonov regu-
larization method that projects negative values to zero in
each iteration could achieve sparse reconstruction results
not worse than other regularization methods.
By studying how zeroing strategy achieves sparsity in
the solutions, in this Letter, a Tikhonov-regularization-
based projecting sparsity pursuit method (PrSP-Tk) is
proposed that can achieve a better image quality than
the Tikhonov regularization method with zeroing strategy
(zeroing-Tk) with much fewer iterations. In this case, the
method could achieve both good image quality and effi-
ciency, making it a practical method for TD-FMT com-
pared with other conventional regularization methods.
For TD-FMT, the forward model was obtained through
the telegrapher equation based on the finite element
method (FEM)
[16]
. To reconstruct the fluorescent yield,
the inverse problem could be formulized by a method based
on the first-order derivative of the measurement data,
i.e., the slope of early photon tomography (s-EPT)
[15]
,as
the following equation
A
S
X ¼ Y
S
; (1)
where X denotes the fluorescence distribution. Y
S
is the
slope of the time-resolved measurement data at a particular
time. A
S
is the matrix derived from the forward model cor-
responding to Y
S
. s-EPT is proved to be able to achieve good
reconstructed image quality
[15]
. Therefore, the simulations
COL 18(1), 011701(2020) CHINESE OPTICS LETTERS January 2020
1671-7694/2020/011701(6) 011701-1 © 2020 Chinese Optics Letters