![](bg3.png)
enhancement filters have been introduced in litera-
ture. Frangi filter has been reported to be one of
the most effective vessel enhancement filter (Frangi
et al., 1998). In light of that, we introduced a no-
vel multiscale blob filtering method derived from the
Frangi filter for the enhancement of neuron somata.
Cell bodies are then segmented by a localizing region-
based active contour algorithm (Lankton and Tannen-
baum, 2008) followed by a watershed-based step to
split groups of neurons and to separate cells from den-
drites and axons.
The remainder of the paper is organized as fol-
lows. In Sec.2 details on the adopted retinal images
are provided. We present the pipeline of our method
in Sec.3. In Sec.4 results are discussed and conclusi-
ons are provided in Sec.5.
2 MATERIALS
Mouse retinal samples were imaged using Leica SP5
upright confocal microscope. Images were acquired
at (sub)cellular resolution and at high averaging num-
ber to reduce the noise level due to limited light pe-
netration in deep layers of the tissue where RGCs
are located. A total of 5 images (2048 × 2048 and
1024 × 1024 pixels), containing some hundreds of
cells, were selected from 3 different retina samples
including: i) three images coming from samples with
genetic fluorescence expression, (i.e., Im1 from PV-
EYFP and Im2 and Im5 images from Thy1-EYFP
mouse), and ii) two images from samples with immu-
nofluorescence staining using the Calretinin calcium-
binding protein (Im3 and Im4) (Fig.1-2). The samples
were selected in order to best capture the variability
in terms of fluorescence expression, cell and axonal
bundle density and background.
3 METHOD
There are mainly three steps in our pipeline as shown
in Fig.2: Multiscale Blob enhancement filtering (Fig
2.b), Localizing Region-Based Active Contour (Fig
2.c) and Watershed Transform (Fig 2.d).
The blob enhancement filtering is used to initia-
lize the high performance active contour method, he-
avily dependent on the initialization mask. Thanks
to this filter, the processing pipeline can proceed wit-
hout user intervention and manual adjustment. Af-
ter blob filtering, the detected blob-shaped objects are
binarized and used as initialization ROIs for a loca-
lizing region-based active-contour that segments cell
borders. In the most challenging images, the active
contour can result in cell clusters due to fuzzy cell
boundaries and occlusions. In order to overcome this
issue, we use the watershed transform.
3.1 Multiscale Blob Enhancement
Filtering
The aim of blob enhancement is to improve the in-
tensity profile of RGC bodies and reduce the contri-
bution of dendritic and axonal structures. It is based
on the multiscale analysis of the eigenvalues of the
Hessian matrix to determine the local likelihood that
a pixel belongs to a cell, i.e. to a blob structure. The
proposed approach is inspired by the work of Frangi
et al. (Frangi et al., 1998) on multiscale vessel en-
hancement filtering. The Frangi filter essentially de-
pends on the orientational difference or anisotropic
distribution of the second-order derivatives to deline-
ate tubular and filament-like structures. We start from
this idea and modify the filtering process (in particular
equation (15) in (Frangi et al., 1998)) in order to have
a reduction of line-like patterns in favor of blob-like
structures (as (Liu et al., 2010)). Instead of a vessel-
ness measure, we define a blobness measure as fol-
lows:
B(x
o
) =
0, if λ
x
o
1
< 0
e
1
2β
2
·
λ
x
o
2
λ
x
o
1
2
, otherwise
(1)
where λ
x
o
1
and λ
x
o
2
are the eigenvalues of the Hessian
matrix at point x
o
and β is a threshold which con-
trols the sensitivity of the blob filter. Both β and
the Hessian scale have been selected in the range of
the average neuron radius. Eq.(1) is given for bright
structures over dark background. In case of dark ob-
jects conditions should be reversed.
3.2 Localizing Region-based Active
Contour
Localizing region-based active contour (Lankton and
Tannenbaum, 2008) is an improved version of traditi-
onal active contour models (Chan et al., 2001), (Yezzi
et al., 2002) where objects characterized by heteroge-
neous statistics can be successfully segmented thanks
to localized energies, differently from the correspon-
ding global ones which would fail. This framework
allows to remove the assumption that foreground and
background regions are distinguishable based on their
global statistics. Indeed the working hypothesis is
that interior and exterior regions of objects are lo-
cally different. Within this framework, the energies
Segmentation of Retinal Ganglion Cells From Fluorescent Microscopy Imaging
19