Page 105 - Computational Retinal Image Analysis

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

2 Benchmark datasets and evaluation metrics

2.1 Datasets
Table 1 provides a summary of 12 representative retinal imaging benchmark datasets.
As some URLs of these published datasets are relocated over the years, we strived to
identify the current web addresses.
1
2
DRIVE [24] and STARE [25] are two popular benchmarks for vessel
segmentation. DRIVE contains 40 fundus images of size 768 × 584, obtained using
Canon CR5 nonmydriatic 3CCD camera with field of view (FOV) of 45 degrees.
STARE involves 20 fundus images of size 605 × 700 taken from a TopCon TRV-50
fundus camera at 35-degree FOV. Apart from segmented vessel masks, the artery/
3
vein vessel annotations are further labeled by RITE [26, 27] for DRIVE dataset.
Azzopardi and Petkov [28] provided vessel junction annotations, including vascular
bifurcations and crossovers.
4
HRF [29] consists of 45 images, of which one-third are images from healthy
patients, one-third are patients with diabetic retinopathy (DR), and the rest are
images of glaucomatous patients. Binary gold standard vessel segmentation images,
generated by a group of experts, are available for each image. The size of fundus
images is 3504 × 2336.
5
CHASEDB1 [30] includes 28 retinal fundus scans from 14 children. Images
were recorded using a Nidek NM-200-D fundus camera with a 30-degree FOV, with
a resolution of 999 × 960 pixels. Two experts’ segmentations are available for each
images as ground truth.
6
ARIA [31] consists of 143 images taken from either healthy subjects, diabetics,
or patients with age-related macular degeneration (AMD). Obtained with a Zeiss
FF450+ fundus camera with a 50-degree FOV. In addition to vessel segmentation,
annotations of the OD and fovea are also provided.
7
The DR HAGIS [32] dataset is composed of 40 fundus images, which were
acquired in different screening centers in the United Kingdom for DR screening
purpose, using a Topcon TRC-NW6s, a Topcon TRC-NW8, or a Canon CR DGi
fundus camera with a horizontal 45-degree FOV. The resulting images are of different
sizes, being either 4752 × 3168, 3456 × 2304, 3126 × 2136, 2896 × 1944, or 2816
× 1880 pixels. The dataset is divided into four comorbidity subgroups, that is, DR,
hypertension, AMD, and glaucoma. Each subgroup comprises 10 images, with the
image of only one patient being duplicated into two subgroups. Vessel segmentation
annotation is also provided for each image.

1 See http://www.isi.uu.nl/Research/Databases/DRIVE/.
2 See http://www.ces.clemson.edu/~ahoover/stare/.
3 See https://medicine.uiowa.edu/eye/rite-dataset.
4 See https://www5.cs.fau.de/research/data/fundus-images/.
5 See https://blogs.kingston.ac.uk/retinal/chasedb1/.
6 See https://eyecharity.weebly.com/aria_online.html.
7 See https://personalpages.manchester.ac.uk/staff/niall.p.mcloughlin/.