Page 107 - Computational Retinal Image Analysis

100 CHAPTER 6 Retinal vascular analysis: Segmentation, tracing, and beyond

8
The dataset of VICAVR includes 58 images, acquired by a TopCon NW-100
nonmydriatic camera with a resolution of 768 × 584. Annotated vascular patterns
include the caliber of the vessels measured at different radii from the optic disc as
well as the vessel type (artery/vein) labeled by three experts.
9
INSPIRE-AVR [33] has 40 images of 2392 × 2048 pixels and corresponding
arteriolar-to-venular diameter ratio (AVR) values observed by two ophthalmologists.
The artery/vein vessel centerlines and the vessel types are further labeled by
Dashtbozorg et al. [27].
10
RVTD [34] is for vascular tortuosity evaluation. It contains 60 zoomed images
with 30 arteries and 30 veins that are ranked by vessel tortuosity. These image patches
are extracted from 60 retinal images with 50-degree FOV and 1300 × 1100 pixels.
In addition to the typical CF photography, there are also benchmarks on different
retinal imaging devices, including scanning laser ophthalmoscopy (SLO), ultra-wide
fundus imaging (UWFI), and fluorescein angiogram (FA).
11
• IOSTAR [35] includes 24 images, vessel segmentation, artery/vein, and
OD annotations. Images are acquired with an EasyScan SLO camera with a
45-degree FOV.
12
• AV-WIDE [36] has 30 ultra-wide FOV images, includes both healthy eyes as
well as eyes with AMD. Each image is from a different individual, was obtained
using obtained using an Optos 200Tx UWFI camera, and is of around 900 ×
1400 in size. Annotations of the vessel segmentation and the artery-vein labels
are provided.
13
• VAMPIRE [37] contains eight ultra-wide FOV FA images, obtained from
OPTOS SLO ultra-wide FOV device which may achieve up to 200-degree FOV
with scanning SLO technique. These images were taken from two different FA
image sequences, with vessel segmentation annotation being provided.
Fig. 1 shows a CF image and various vessel annotations, including vascular
junctions, vessel segmented, artery/vein, and tree masks. Meanwhile, there are also
14
efforts in applying multispectral imaging to retinal studies. For example, CMIF
[38] is a dedicated multispectral fundus image dataset of 35 sets of images from
35 healthy young individuals of diverse ethnicities. The multispectral images are
acquired by implementing a filter wheel into fundus camera, which gives rise to a set
of 17 narrow band-pass filters for designated wavelengths in the range of 480–705 nm.
This development is yet to be ready for clinical translation, nevertheless it offers
novel insights into retinopathy.

8
See http://www.varpa.es/research/ophtalmology.html.
9
See https://medicine.uiowa.edu/eye/inspire-datasets.
10
See http://bioimlab.dei.unipd.it/RetinalVesselTortuosity.htm.
11
See http://www.retinacheck.org/datasets.
12
See http://people.duke.edu/~sf59/Estrada_TMI_2015_dataset.htm.
13
See https://vampire.computing.dundee.ac.uk/vesselseg.html.
14
See http://www.cs.bham.ac.uk/research/projects/fundus-multispectral/.