Segmentation of Retinal Ganglion Cells From Fluorescent Microscopy

Imaging

Silvia Baglietto

1,2

, Ibolya E. Kepiro

, Gerrit Hilgen

, Evelyne Sernagor

, Vittorio Murino

1,5

and Diego Sona

1,6

Pattern Analysis and Computer Vision, Istituto Italiano di Tecnologia, Genova, Italy

Department of Naval, Electric, Electronic and Telecommunication Engineering, University of Genova, Genova, Italy

Nanophysics, Istituto Italiano di Tecnologia, Genova, Italy

Institute of Neuroscience, Newcastle, Newcastle-upon-Tyne, U.K.

Department of Computer Science, University of Verona, Verona, Italy

NeuroInformatics Laboratory, Fondazione Bruno Kessler, Trento, Italy

Keywords:

Cell Segmentation, Active Contour, Multiscale Blob Filtering, Retinal Ganglion Cell.

Abstract:

The visual information processing starts in the retina. The working mechanisms of its complex stratiﬁed

circuits, in which ganglion cells play a central role, is still largely unknown. Understanding the visual coding

is a challenging and active research area also requiring automated analysis of retinal images. It demands

appropriate algorithms and methods for studying a network population of strictly entangled cells. Within this

framework, we propose a combined technique for segmenting retinal ganglion cell (RGC) bodies, the output

elements of the retina. The method incorporates a blob enhancement ﬁltering in order to select the speciﬁc

cell shapes, an active contour process for precise border segmentation and a watershed transform step which

separates single cell contours in possible grouped segmentations. The proposed approach has been validated

on ﬂuorescent microscopy images of mouse retinas with promising results.

1 INTRODUCTION

The retina is a photosensitive membranous tissue

lying at the back of the eye. Its role is to process

the light stimuli and to transmit the information to

the brain through the optic nerve. Although it is of-

ten compared to a photographic ﬁlm on which the

images are imprinted and encoded into electrical sig-

nals with a columnar processing, the retina has a very

complex function and structure, composed of several

layers of cells in a tangled network (Gregory, 2015):

photoreceptors feed into bipolar and amacrine cells,

which pass the input to an assorted set of retinal gan-

glion cells (RGCs). Recent studies show that retina

performs sophisticated non-linear computation, ex-

tracting spatio-temporal visual features with high se-

lectivity. This is reﬂected by the fact that distinct

RGC types have been found to encode speciﬁc visual

features for transmission to the brain (Neumann et al.,

2016).

According to the current knowledge, there is a cor-

relation between RGC morphology and function and

this is usually studied at the single cell level (Ber-

son et al., 2010). Recently, an intense debate is ri-

sing around the importance of studying the retina at

the circuit level and the actual challenge is to tackle

the problem at a different scale, analyzing populati-

ons of neurons at a network level (Roska et al., 2006),

(Kim et al., 2010). Several studies are trying to mo-

del responses of RGC populations heading to a bet-

ter understanding of the role of RGCs (Baden et al.,

2016). To support neuroscientists in this study, auto-

mated tools for the large-scale segmentation of RGC

bodies are required.

In this work we propose an automatic algorithm

for cell segmentation from ﬂuorescent images recor-

ded with a confocal microscope. This process would

allow the characterization of different RGC types dis-

tinguished by different soma morphologies at the po-

pulation level, inspiring studies addressing the corre-

lation between soma morphology and functional be-

havior.

Such datasets are complex for many reasons. They

are affected by low contrast at the cell boundaries,

Baglietto S., Kepiro I., Hilgen G., Sernagor E., Murino V. and Sona D.

Segmentation of Retinal Ganglion Cells From Fluorescent Microscopy Imaging.

DOI: 10.5220/0006110300170023

In Proceedings of the 10th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2017), pages 17-23

ISBN: 978-989-758-215-8

Im1 (PV-EYFP)

Im2 (Thy1-EYFP)

Im3 (Calretinin)

Im4 (Calretinin)

Im5 (Thy1-EYFP)

Detail of Im5

Figure 1: 5 different images containing RGCs used for testing the proposed method. The images show high variability

across samples. In the bottom right, there is a magniﬁed crop of Im5, showing the complexity of images, where the analyzed

structures are mixed with background and other structures.

high cell density and shape deformations, occlusions

among neurons and the presence of a large mass of ot-

her structures such as dendrites and axons on the same

channel. Indeed, the ﬂuorescence in retinal mice ima-

ges, as shown in Fig.1, is expressed by the whole cell

including processes (i.e. dendrites and axons) and it

is non-uniform leading to fragmented appearance of

the objects and to unreliable separation between so-

mas, dendrites and axons. Last but not least, there is a

high variability of samples condition across different

acquisitions.

This global setting introduces to new challenging

computational tasks for image segmentation. Indeed,

state-of-the-art methods usually work on single neu-

ron reconstruction (Gulyanon et al., 2016) and can

hardly be adequate for separating neurons from the

background. The automated segmentation is still a

critical open problem. On the other hand, the ma-

nual interaction to generate the morphological recon-

struction is time consuming and expensive. Tradi-

tional segmentation approaches which use only ba-

sic techniques, such as morphological operators and

thresholding, are not powerful enough and lead to

wrong segmentations (Meijering, 2012). Learning ap-

proaches, such as (Arteta et al., 2013) and (Zhang

et al., 2014), require hand-labelled neurons for trai-

ning and testing. In addition, they cannot manage to

extract the precise segmentation of cells because of

the difﬁculties dealing with the high variance in cell

appearance. In contrast, active contour methods have

demonstrated good performance in image segmenta-

tion dealing with challenging data (Chan et al., 2001),

(Yezzi et al., 2002). Their main limitation is rela-

ted to the strong sensitivity to the model initialization,

which usually requires variable degrees of user inter-

vention. To this end, recent years have witnessed the

spread of active contour models in different formu-

lations, aiming at hybrid approaches for automating

the initialization process (Ge et al., 2015), (Wu et al.,

2015).

Within this scenario, we designed a method based

on active contour initialized on speciﬁc ROIs, which

are automatically identiﬁed by a multiscale blob ﬁlter

emphasizing only cell bodies. Several shape-based

BIOIMAGING 2017 - 4th International Conference on Bioimaging

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

) =











0, if λ

< 0

2β









, otherwise

(1)

where λ

and λ

are the eigenvalues of the Hessian

matrix at point x

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

Figure 2: Pipeline applied to two examples (from the top, Im2 (Thy1-EYFP) and Im1 (PV-EYFP)) with a crop in the bottom

line, showing a problem due to contiguous cells. In column, (a): Original Fluorescent Microscopy Images; (b) Results of the

blob ﬁlter binarization; (c) Results of the active contour segmentation in blue transparency over the original image for getting

the suitable qualitative performance; (d) Results of the watershed transform and of the ﬁnal threshold.

are constructed locally at each point along the curve

in order to allow the analysis of local regions. The

choice of the localization radius is driven by the size

of the object to be segmented. In our case, for each

image, we used a radius equal to the average soma

radius, which depends on the image size and on the

microscope lens.

Thanks to this efﬁcient technique, we obtain a seg-

mentation mask which tightly ﬁts real cell bodies.

3.3 Watershed Transform and Size

Filter

The above active contour fails to separate groups of

overlapping or contiguous cells, hence we exploit the

simplicity and computational speed of the waters-

hed transform, introduced by Beucher and Lantu

ejoul

(Beucher and Lantu

ejoul, 1979).

As a ﬁnal step, we need to delete components

which are too small or too large for being cell so-

mata (a given example is in Fig.2.c (middle ﬁgure)

by applying a size ﬁlter to remove structures with size

outside an acceptable range of somata dimensions.

4 RESULTS AND DISCUSSION

We applied our pipeline to 5 different retinal ima-

ges representative of possible variations on the retinal

samples, such as brightness, intensity, size and num-

ber of cells, presence of axonal structures and pro-

cesses, strong background signals, etc. We generated

the ground truth manually segmenting all cells in each

image (around 280 cells). To give a qualitative evalu-

ation, we report different examples in Fig.2-3 where

it is possible to see that our approach works in diffe-

rent sample conditions. To quantify the performance

of our method, we adopt the Dice Coefﬁcient (DC),

a widely used metric for comparing the ground truth

to the computer-aided segmentation. DC is deﬁned as

follows:

DC =

2(A ∩ B)

(A + B)

where A is the binary ground truth mask and B is

the binary segmentation result. The DC value ran-

BIOIMAGING 2017 - 4th International Conference on Bioimaging

Crop of Im5 Resulted Crop Segmentation

Figure 3: Some cells are not easily visible to the human eye just visualizing the retina images, but they are discovered and

segmented by our algorithm (for example, in this cropped ﬁgure, pink and blue cells were hardly detectable). Adding contrast

to the image makes these somata clearer but it increases noise and cell heterogeneity.

Table 1: Segmentation process results. Dice Coefﬁcient has been computed after all steps in the pipeline (Blob Filter, Active

Contour and Watershed Transform) and shows improvements after each step. For the ﬁnal stage of the pipeline, there is also

the percentage of detected cells computed assuming as detected a cell with minimum overlap with ground truth ﬁxed at 50%.

Image # of cells Blob Filter Active Contour Final

DC DC DC detected cells

Im1 (PV-EYFP) 95 0.60 0.69 0.81 86.32%

Im2 (Thy1-EYFP) 37 0.43 0.58 0.64 89.19%

Im3 (Calretinin) 64 0.62 0.82 0.83 75.00%

Im4 (Calretinin) 29 0.57 0.71 0.79 82.76%

Im5 (Thy1-EYFP) 48 0.51 0.62 0.70 85.42%

ges between 0 (absence of agreement) and 1 (perfect

agreement). A DC higher than 0.70 usually indicates

a satisfactory segmentation (Zijdenbos et al., 1994).

As an additional index of performance, we also

provide the percentage of detected cells for each

image. We consider a cell as detected if it is correctly

segmented for more than 50% of its total area, com-

paring the segmentation mask to the ground truth for

each annotated RGC. Fig.4 shows that 50% threshold

is a good trade off between the certainty of a cell de-

tection and a satisfactory retrieval.

Table 1 shows the quantitative results of these me-

trics on our samples. We compute the DC for each of

the three steps. Each stage clearly improves the seg-

mentation, reaching satisfactory results for all ima-

ges. In Im3 (Fig.1), the ﬂuorescence is mainly ex-

pressed by the body cells; for this reason, we reach

good scores right after the ﬁrst two steps. The weaker

DC values on images Im2 and Im5 are due to a strong

presence of axonal structures which can be hardly re-

moved.

5 CONCLUSIONS

In this paper we have proposed a new algorithm for

the large-scale segmentation of cells in challenging

retinal images. First, a novel and effective multiscale

blob ﬁlter is employed for cell enhancement which

selects ROIs for the initialization of an active con-

tour step, addressing the known weakness of these

methods. Active contour reaches suitable results but

needs a further segmentation in case of multiple cell

aggregations, which has been addressed using a wa-

tershed transform followed by a ﬁlter guided by the

size of structures.

We validated our approach against manual seg-

mentations on 5 images in which there are some hund-

reds of neurons representative of a variety of cell ap-

pearances and image conditions.

Thanks to its generality, this approach could be

applied to similar cell segmentation scenarios and

opens new perspectives for the analysis and the cha-

racterization of the retinal morphology at a population

level.

Segmentation of Retinal Ganglion Cells From Fluorescent Microscopy Imaging

Figure 4: Variation of the % of detected cells as a function of the % of overlap between detected cell and the corresponding

annotated ground truth.

ACKNOWLEDGEMENTS

The research received ﬁnancial support from the 7

Framework Programme for Research of the European

Commision, Grant agreement no. 600847: RENVI-

SION project of the Future and Emerging Technolo-

gies (FET) programme.

We would like to thank Dr. Luca Berdondini, Dr.

Alessandro Maccione, and Dr. Stefano Di Marco

(Neuroscience and Brain Technologies, Istituto Itali-

ano di Tecnologia, Genova) for providing the CMOS

devices and for their support in the sample prepara-

tion. We would also like to thank Dr. Francesca Cella

Zanacchi (Nanophysics, Istituto Italiano di Tecnolo-

gia, Genova) for her support in the microscopy acqui-

sitions.

REFERENCES

Arteta, C., Lempitsky, V., Noble, J. A., and Zisserman, A.

(2013). Learning to detect partially overlapping in-

stances. In Computer Vision and Pattern Recognition

(CVPR), 2013 IEEE Conference on.

Baden, T., Berens, P., Franke, K., Ros

on, M. R., Bethge,

M., and Euler, T. (2016). The functional diver-

sity of retinal ganglion cells in the mouse. Nature,

529(7586):345–350.

Berson, D. M., Castrucci, A. M., and Provencio, I. (2010).

Morphology and mosaics of melanopsin-expressing

retinal ganglion cell types in mice. Journal of Com-

parative Neurology, 518(13):2405–2422.

Beucher, S. and Lantu

ejoul, C. (1979). Use of waters-

heds in contour detection. In International workshop

on image processing, real-time edge and motion de-

tection.

Chan, T. F., Vese, L., et al. (2001). Active contours without

edges. Image processing, IEEE transactions on, 10.

Frangi, A. F., Niessen, W. J., Vincken, K. L., and Viergever,

M. A. (1998). Multiscale vessel enhancement ﬁlte-

ring. In Medical Image Computing and Computer-

Assisted InterventationMICCAI98, pages 130–137.

Springer.

Ge, Q., Li, C., Shao, W., and Li, H. (2015). A hybrid active

contour model with structured feature for image seg-

mentation. Signal Processing, 108:147–158.

Gregory, R. L. (2015). Eye and Brain: The Psychology of

Seeing: The Psychology of Seeing. Princeton univer-

sity press.

Gulyanon, S., Sharifai, N., Kim, M. D., Chiba, A., and Tse-

chpenakis, G. (2016). Crf formulation of active con-

tour population for efﬁcient three-dimensional neurite

tracing. In 2016 IEEE 13th International Symposium

on Biomedical Imaging (ISBI), pages 593–597. IEEE.

Kim, I.-J., Zhang, Y., Meister, M., and Sanes, J. R. (2010).

Laminar restriction of retinal ganglion cell dendrites

and axons: subtype-speciﬁc developmental patterns

revealed with transgenic markers. The Journal of Neu-

roscience, 30(4):1452–1462.

Lankton, S. and Tannenbaum, A. (2008). Localizing region-

BIOIMAGING 2017 - 4th International Conference on Bioimaging

based active contours. Image Processing, IEEE Tran-

sactions on, 17(11):2029–2039.

Liu, J., White, J. M., and Summers, R. M. (2010). Auto-

mated detection of blob structures by hessian analysis

and object scale. In Image Processing (ICIP), 2010

17th IEEE International Conference on, pages 841–

844. IEEE.

Meijering, E. (2012). Cell segmentation: 50 years down

the road [life sciences]. Signal Processing Magazine,

IEEE, 29(5):140–145.

Neumann, S., Hser, L., Ondreka, K., Auler, N., and Haver-

kamp, S. (2016). Cell type-speciﬁc bipolar cell input

to ganglion cells in the mouse retina. Neuroscience,

316:420 – 432.

Roska, B., Molnar, A., and Werblin, F. S. (2006). Paral-

lel processing in retinal ganglion cells: how integra-

tion of space-time patterns of excitation and inhibition

form the spiking output. Journal of Neurophysiology,

95(6):3810–3822.

Wu, P., Yi, J., Zhao, G., Huang, Z.and Qiu, B., and Gao,

D. (2015). Active contour-based cell segmentation

during freezing and its application in cryopreserva-

tion. Biomedical Engineering, IEEE Transactions on,

62(1):284–295.

Yezzi, A., Tsai, A., and Willsky, A. (2002). A fully glo-

bal approach to image segmentation via coupled curve

evolution equations. Journal of Visual Communica-

tion and Image Representation, 13(1):195–216.

Zhang, C., Yarkony, J., and Hamprecht, F. A. (2014). Cell

detection and segmentation using correlation cluste-

ring. In Medical Image Computing and Computer-

Assisted Intervention–MICCAI 2014, pages 9–16.

Springer.

Zijdenbos, A. P., Dawant, B. M., Margolin, R., Palmer,

A. C., et al. (1994). Morphometric analysis of white

matter lesions in mr images: method and validation.

Medical Imaging, IEEE Transactions on, 13(4):716–

724.

Segmentation of Retinal Ganglion Cells From Fluorescent Microscopy Imaging