4. The investigation of the brain requires several challenges in
imaging technology, since brain activity spans many orders of
magnitude both spatially (from nanometers to centimeters) and
temporally (from milliseconds to months).
Thanks to their flexibility in terms of spatial and temporal
resolution optical methodologies have become very useful in
neuroscience research.
Whole mouse brain
cm scale
Pyramidal neuron
mm scale
Synapses
μm scale
Optical techniques to explore the brain
5. Light sheet microscopy (LSM)
Key advantages:
from J. Huisken and D. Y. Stainier (Development, 2009)
Optical sectioning with wide field
detection scheme
Fast high resolution 3D imaging
Optical sectioning with low-NA optics
(having longer WD)
Imaging of large specimens without
sample sectioning.
Only the observed plane is illuminated
Reduced photobleaching.
6. Traditional clearing protocols are based on the
substitution of water with a refractive-index-
matching liquid, as Benzyl Alcohol/Benzyl
Benzoate (BABB) or Dibenzyl Ether (DBE)
[Becker et al. 2012].
Chemical clearing of entire brains
Silvestri et al. JoVe 2013.
In large specimens as whole mouse brains, a
substantial amount of light scattering persist
even after clearing. This expands the
illumination beam, leading to out-of-focus
background fluorescence and blurring of
images.
7. We recently developed a new
implementation of confocal slit
detection (confocal light sheet
microscopy, CLSM) in which the
out-of-focus background rejection
is assured by a spatial filter.
Confocal light sheet microscopy (CLSM)
Silvestri et al. Optics Express 2012
8. SV CV
TV
10×
10×
Cerebellum from a P10 L7-GFP mouse cleared in BABB
Total volume 73 mm³, voxel size 0.8×0.8×1 µm³, acquisition time ≈ 24 h (1.3 MegaVoxels/s)
Whole brain imaging
11. Ex vivo light sheetIn vivo reflectanceIn vivo two-photon
Light sheet microscopy of mouse brains is essentially an ex vivo technique. It can be combined with in vivo
two-photon imaging to gain a more comprehensive multi-scale view of the brain.
To find back in the cleared brain the same field of view imaged in vivo, blood vessels can be used as a
reference map.
Correlative two-photon and light sheet microscopy
12. The same neuron imaged in vivo with two-photon microscopy can be found again
in the large-scale optical tomography obtained with CLSM
The side of the red cube is 100 μm
Correlative two-photon and light sheet microscopy
Silvestri et al. Methods 2014
15. A 2 mm thick block of a formalin-
fixed tissue of a patient with
hemimegalencephaly (HME),
treated with passive CLARITY
protocol immunostained with
different antibody and cleared with
47% TDE/PBS
0
200
400 µm
600 µm
800 µm
1000 µm
Human Brain Imaging
Costantinie et al. Sci Report 2015
16. Adapted from Kasthuri and Lichtman (2007)
A data flood
Data production: about 5 TB per week
- 10 Gb/s dedicated connection from LENS to CINECA
- Connection from LENS to Juelich via CINECA (using PRACE infrastructure)
17. To image an entire brain many parallel stacks of images are acquired. They are subsequently merged with
a custom-made software suited to work with very large data sets (~ 1 TB)
Tera Stitcher ®
Bria et al., BMC Bioinformatics 2012
19. Very large samples (cm-sized)
Large variability of contrast in
different areas
TB-sized datasets
Naïve methods (as thresholding or clustering)
are poorly effective as usually depend on
sensitive parameters
Advanced methods usually requires the
calculation of multiple features per each
image voxel, leading to a data multiplication
which is not manageable with large datasets
Automatic 3D cell localization
20. Semantic deconvolution uses a supervised neural network, to enhance selected
features of the image reducing the intensity of other structures.
This method creates a more uniform
image where significant structures
(hence the name semantic) are well
visible.
In the deconvolved image, easy low-
cost localization algorithms (e.g.
clustering) can achieve very good
performances.
Original image
Ideal image
Deconvolved image
Manualsoma
localization
Neural
network
with 2
hidden
layers
Supervision
Only on a small subset
Semantic deconvolution
Frasconi et al., Bioinformatics 2014
21. 224’222 Purkinje cells automatically localized in the cerebellum of an L7-GFP mouse. Precision (true
positives/all localized cells) is 95%, recall (true positives/all real cells) 97%
~ 1 day of computation on a 16-cores workstation to analyze 120 Gvoxels
Whole brain quantitative neuroanatomy
Silvestri et al., Frontiers in Neuroanatomy 2015