Dynamic speckle illumination microscopy with
Cathie Ventalon,1Rainer Heintzmann,2and Jerome Mertz1,*
1Boston University, Department of Biomedical Engineering, 44 Cummington Street, Boston,
Massachusetts 02215, USA
2Randall Division of Cell and Molecular Biophysics, King’s College, London, London SE1 1UL, UK
*Corresponding author: firstname.lastname@example.org
Received January 18, 2007; revised March 7, 2007; accepted March 15, 2007;
posted March 23, 2007 (Doc. ID 79148); published April 27, 2007
Dynamic speckle illumination (DSI) provides a simple and robust technique to obtain fluorescence depth
sectioning with a widefield microscope. We report a significant improvement to DSI microscopy based on a
statistical image-processing algorithm that incorporates spatial wavelet prefiltering. The resultant gain in
sectioning strength leads to a fundamentally improved scaling law for the out-of-focus background rejection.
© 2007 Optical Society of America
OCIS codes: 180.1790, 180.2520, 180.6900.
In its standard configuration , confocal microscopy
is based on single-beam scanning and detection
through a simple pinhole mask. Alternatives have
been proposed that use multipinhole masks [2–4] or
no mask at all [5,6]. These techniques all require pre-
defined light patterns (e.g., structured illumination)
that must be highly contrasted at the focal plane, and
depth sectioning is obtained by varying these pat-
terns in a controlled manner. We recently proposed
an entirely new fluorescence imaging technique
based on dynamic speckle illumination (DSI) . In
this technique, the exact illumination patterns are
not known, and depth sectioning is obtained from a
knowledge of only the illumination pattern statistics.
A key advantage of DSI microscopy is that it is re-
markably simple to implement, requiring no detec-
tion mask and only a diffuser, a microscope objective,
and a CCD camera.
The general layout of a DSI microscope is illus-
trated in Fig. 1a. Diffuse laser light is projected into
a sample via a microscope objective, producing a fine
hail of speckle grains. The average lateral size of the
speckle grains is the same as the diffraction-limited
microscope resolution. The intensity distribution of
the grains obeys what is known as negative-
exponential statistics , which is inherently highly
contrasted. An important feature of these statistics is
that they are invariant even in a scattering medium,
since unpredictable phase shifts provoked by the me-
dium only further randomize an already randomized
laser phase front (we assume that the medium does
not significantly alter the beam polarization). Be-
cause the laser illumination is granular, so too is the
resulting fluorescence, which is imaged by the CCD
camera. However, even though the fluorescence is
generated with high contrast everywhere in the
sample, it only appears to be highly contrasted when
it is in focus. DSI microscopy involves acquiring a se-
quence of fluorescence images with randomly chang-
ing (i.e., dynamic) speckle patterns. A postprocessing
algorithm that extracts the varying components from
the resultant image sequence therefore preferentially
extracts in-focus signal from out-of-focus background.
Our initial implementations of DSI microscopy were
based on algorithms that extracted image variations
in time only [7,9]. We report here a significantly im-
proved wavelet-based algorithm that extracts image
variations in both time and space.
The performance of any 3D imaging system is char-
acterized by its lateral resolution and its axial sec-
tioning strength (capacity to reject out-of-focus blur).
In standard widefield microscopy with uniform illu-
mination, the image intensity Idrecorded at the CCD
camera is given by the convolution Id??PSFd?C),
where PSFdis the detection point spread function
and C is the fluorophore concentration in the sample.
While PSFdprovides high in-focus lateral resolution,
it does not provide good axial sectioning because of
its weak dependence on defocus z.
DSI microscopy is essentially a standard widefield
microscope but with speckle illumination. We have
Id?r ?d? =?PSFd?r ?d− r ??C?r ??Is?r ??d3r ?,
where Is?r ?? is the intensity distribution of a given
speckle pattern inside the sample, and r ?=?? ?,z? and
r ?d=?? ?d,0? are the 3D coordinates inside the sample
and at the detector plane, respectively. By acquiring
a series of images with independent speckle patterns
(e.g., Fig. 1b), the image components that vary in
time are extracted by a simple calculation of the root-
mean-square (rms) of the intensity variations at each
pixel (Fig. 1e). The final DSI image intensity IDSIis
given by the rms (i.e., the square root of the temporal
variance) of the detected raw images, defined from
?r ?d?=Id?r ?d?2−Id?r ?d?2. To evaluate this rms, we use
a property of fully developed speckle statistics :
Is?? ?,z?Is?? ??,z? = Is2?1 + PSFi?? ? − ? ??,0??
? Is2?1 + As??? ? − ? ????,
where PSFirepresents the illumination PSF, Ascor-
responds to the area of a speckle grain, and the over-
bar indicates an average over many independent
June 1, 2007 / Vol. 32, No. 11 / OPTICS LETTERS
0146-9592/07/111417-3/$15.00© 2007 Optical Society of America
speckle patterns. The delta-function approximation
corresponds to the ideal case of very small speckle
grain areas. Because the speckle grains are also con-
fined along the axial dimension to a dimension Lsof
the order of the axial width (Rayleigh length) of the
illumination PSF , we can extend this approxima-
tion to three dimensions:
Is?r ??Is?r ??? ? Is2?1 + AsLs??r ? − r ????.
From Eqs. (1) and (3), we readily obtain
?r ?d? ? Is2AsLs?PSFd?r ?d− r ??2C?r ??2d3r ?.
Expression (4) was derived purely on the basis of
speckle statistics and does not depend on the particu-
lar sequence of speckle patterns. In effect, it leads to
been applied to PSFdand C prior to the application of
the convolution. This has important ramifications,
dependence on defocus than PSFd. DSI microscopy,
as described above, therefore exhibits both high lat-
2?C2?1/2. We emphasize that squares have
2is more localized and exhibits a stronger
eral resolution and a degree of depth sectioning .
We note that, inasmuch as PSF2is highly localized,
IDSIscales locally linearly with C.
In this Letter, we report a fundamental improve-
ment in sectioning strength by going one step further
in first extracting the spatial variations from each
image prior to extracting their temporal variations.
To do this, we convolve each raw image with a 2D
wavelet filter of the type illustrated in Fig. 1c. The
overall effect of such a convolution is to replace PSFd
in Eqs. (1) and (4) with the effective PSF
PSFw?? ?,z? =?W?? ???PSFd?? ?? − ? ?,z?d2? ??,
where W is the 2D wavelet filter function. Properties
of this filter are that it should remove uniform back-
ground (i.e., ?W?? ??d2? ?=0) and be localized to approxi-
mately the size of a speckle grain. As such, the filter
serves as an in-focus speckle grain finder, imparting
an overall stronger dependence on defocus to PSFw
than was exhibited by PSFd. Wavelet prefiltering
leads to a striking improvement in DSI imaging (Fig.
1f) that is not only qualitative but quantitative. The
enhancement in sectioning strength can be evaluated
theoretically by considering a thin uniform fluores-
cent planar sample. If the microscope aperture is cir-
cular, then IDSIis found to scale as 1/z when no wave-
let filter is applied  and as 1/z3/2when wavelet
filtering is applied. These scaling laws are corrobo-
rated by experiment (Fig. 1g). Recalling that the sec-
tioning strength for an ideal confocal microscope
scales as 1/z2, our new spatiotemporal DSI algorithm
yields indeed very close to ideal confocal perfor-
mance. We note that our evaluation of sectioning
strength based on a uniform fluorescent plane is
valid only in the limit of strong out-of-focus back-
ground relative to in-focus signal. Because the rms is
an inherently nonlinear operation, the separation be-
tween background and signal is not as clear-cut in fa-
vorable regimes of weak background or large defocus
and can lead to an even stronger effective sectioning
strength than reported here.
To demonstrate the effectiveness of DSI microscopy
in deep tissue imaging, we display images of an ex-
cised mouse olfactory bulb labeled with green fluores-
cent protein (GFP). The illumination source was an
argon-ion laser (JDS Uniphase 2214-20SL, 488 nm);
the diffusing element was a Holoeye LC-R-768 liquid
crystal modulator used in reflection mode (see Ref.
 for details); and the CCD camera was a QImaging
Retiga 2000R. Figures 2 and 3 compare temporal and
spatiotemporal DSI imaging in thick tissue. The sec-
tioning strength of the spatiotemporal algorithm is
manifestly superior, particularly at larger depths. As
illustrated, 10 raw images can yield acceptable DSI
imaging. This can be improved by extending the
number to 30 raw images (see Fig. 3d); however,
there is little advantage to increasing the image
An important feature of DSI microscopy is that, as
opposed to confocal microscopy, it does not involve a
detection mask, which presents disadvantages and
croscope objective; BS, dichroic beamsplitter. b–f, Images of
a fluorescent pollen grain obtained with an Olympus 40?,
1.3 NA oil objective. b, Widefield image obtained with an
arbitrary speckle pattern. Spatial prefiltering is performed
by convolving with a 2D wavelet filter, c, that removes uni-
form background and extracts in-focus speckle grains, d.
Final DSI images are obtained by computing the rms of an
image sequence illuminated with random speckle patterns
for each image, without, e, or with, f, wavelet prefiltering.
DSI images e and f are obtained from the same data set.
Scale bar, 2 ?m. For the gradients in images b–f, green and
orange (bottom of scale d) correspond to positive and nega-
tive values, respectively; black corresponds to zero. g, Tem-
poral (blue, topmost plots) and spatiotemporal (red) DSI
signals measured from a thin fluorescent plane as a func-
tion of defocus z, acquired with an Olympus 40?, 0.65 NA
objective. Traces are shown on linear (left) and logarithmic
scales (right). In the right-hand panel, the experimental
traces are fitted with straight lines of slope −1 and −3/2 for
temporal and spatiotemporal DSI (respectively).
(Color online) a, DSI microscope layout. MO, mi-
OPTICS LETTERS / Vol. 32, No. 11 / June 1, 2007
advantages. A disadvantage is that, because there is
no detection mask, shot noise from out-of-focus light
can cause DSI image degradation. However, because
shot noise obeys independent intensity statistics
than speckle, the bias it introduces in the calculation
of image variance [Eq. (4)] can be removed. For unfil-
tered images, the shot-noise variance is given by
V?? ?d?shot,unfiltered=gId, where g is the CCD camera
gain and Id?? ?d? is the time average of the unfiltered
raw images (corresponding to the standard widefield
image). For wavelet-filtered images, spatial averag-
ing decreases the overall shot-noise variance, and we
have V?? ?d?shot,filtered=gId?? ?d?Ap?W2?? ??d2? ?, where Apis
the area of a camera pixel, assumed to be smaller
than the area of imaged speckle grains. To remove
this extra shot-noise variance, we subtract it from
the variance of the filtered or nonfiltered fluorescence
images derived from Eq. (4). This provides an unbi-
ased estimate of IDSI
, and hence of IDSI.
We previously reported a near-confocal sectioning
strength of z−3/2using a temporal DSI algorithm that
required speckle pattern translation . Our new
spatiotemporal DSI algorithm provides the same sec-
tioning strength but without this requirement. This
has significant advantages when performing thick
tissue imaging, since it obviates the need for speckle
motion control. Because randomized speckle illumi-
nation is particularly easy to achieve with a simple
diffuser , it can be readily implemented in any mi-
croscope or imaging endoscope.
We thank Matt Wachowiak, Nicolas Pirez, and
Ryan Carey for providing olfactory bulb samples. C.
Ventalon is partially supported by the Délégation
Générale pour l’Armement (France).
1. T. Wilson and C. Sheppard, Theory and Practice of
Scanning Optical Microscopy (Academic, 1984).
2. G. Q. Xiao, T. R. Corle, and G. S. Kino, Appl. Phys.
Lett. 53, 716 (2000).
3. R. Juskaitis, T. Wilson, M. A. A. Neil, and M. Kozubek,
Nature 383, 804 (1996).
4. P. J. Verveer, G. S. Hanley, P. W. Verbeek, L. J. Van
Vliet, and T. M. Jovin, J. Microsc. 189, 192 (1998).
5. M. A. A. Neil and T. Wilson, Opt. Lett. 22, 1905 (1997).
6. M. A. A. Neil, A. Squire, R. Juskaitis, P. I. H.
Bastiaens, and T. Wilson, J. Microsc. 197, 1 (2000).
7. C. Ventalon and J. Mertz, Opt. Lett. 30, 3350 (2005).
8. J. W. Goodman, Speckle Phenomena: Theory and
Applications (Roberts & Company, 2006).
9. C. Ventalon and J. Mertz, Opt. Express 14, 7198
Fig. 2. (Color online) Images of an excised mouse olfactory
bulb labeled with synapto-pHluorin, a pH-sensitive GFP
targeted to presynaptic terminals of sensory neurons, ac-
quired with an Olympus, 20?, 0.95 NA water objective.
Axial sections of a glomerulus are acquired at depths
50 ?m, a–c, and 90 ?m, d–f, below the tissue surface, ob-
tained with standard widefield, a,d, and DSI microscopy
using temporal, b,e, and spatiotemporal, c,f, postprocessing
with 10 images (same data sets for each row). Standard
widefield images are obtained from the average (as opposed
to the rms) of the DSI raw image sequence. Scale bar,
tory bulb (same sample and experimental conditions as in
Fig. 2). Images are obtained with standard widefield, a,
and DSI microscopy using temporal, b, and spatiotemporal,
c,d, postprocessing, using 10, a, b, c, and 30, d, raw images.
The exposure time per raw image was 500 ms, imposed by
the low maximum output power delivered by our laser
(?3 mW at the sample). Scale bar, 20 ?m.
(Color online) Axon fibers in a GFP-labeled olfac-
June 1, 2007 / Vol. 32, No. 11 / OPTICS LETTERS