Page 90 - Computational Retinal Image Analysis

82 CHAPTER 5 Automatic landmark detection in fundus photography

were captured using a Topcon TRC NW6 non-mydriatic camera at 45° field of
view at pixel resolutions of 1440 × 960, 2240 × 1488 or 2304 × 1536 [10]. http://
www.adcis.net/en/Download-Third-Party/Messidor.html
STARE: Structured Analysis of the Retina—The STARE database (2000)
consists of ~400 images taken on a TopCon TRV-50 at 35° field of view. The
film was digitized at 605 × 700 pixels per color plane. The database contains
ground truth marking for the optic disc center if present and boasts a wide
variety of disease and image quality levels to contend with [11]. http://cecas.
clemson.edu/~ahoover/stare/
Other databases in use include the DRIONS-DB: Digital Retinal Images
for Optic Nerve Segmentation Database which contains 110 images and mul-
tiple ground truth segmentations of the optic nerve head (http://www.ia.uned.
es/~ejcarmona/DRIONS-DB.html) [12]. The DIARETDB0 and DIARETDB1 con-
tain 130 and 89 images respectively and contain mostly images with at least mild
DR (http://www.it.lut.fi/project/imageret/diaretdb0/) [13, 14]. The e-ophtha data-
base contains 463 images with DR lesions and ground truth segmentations for each
lesion (http://www.adcis.net/en/Download-Third-Party/E-Ophtha.html) [15]. The
Kaggle DR database was made available for a DR labeling competition in 2015.
The large dataset from EyePACs contains images from multiple cameras at multiple
pixel resolutions. Over 80,000 retinal images with DR grades were made avail-
able to train and validate deep learning models [16] (https://www.kaggle.com/c/
diabetic-retinopathy-detection/data).

4 Algorithm accuracy
In order to test an algorithm’s accuracy, there must be some level of ground truth
available for the image. For OD and fovea detection, an ophthalmologist or experi-
enced image grader generally marks the center point pixels for both. These values are
recorded and used to compare against. In this case, results for an algorithm may be
stated as the distance from the ground truth pixel. Average and standard deviations
can then be measured for a dataset. For binary accuracy, an acceptable distance from
the ground truth pixel is used as a threshold. Usually 1 disc radius is used as an ac-
ceptable value. In other cases, the boundary of the OD may be delineated or an OD
mask will be available. If the detected OD is within the ground truth boundary, it is
considered correctly found.
When researchers compare their algorithm against others, they usually do so by
including results on one or more of the open datasets. Processing time is also consid-
ered along with the detection accuracy. There has generally been a tradeoff between
these two factors, but more recent methods have shown that both can be achieved
[17]. Also, older methods that self-report running time would be sped up by the pro-
cessing power of today’s computers.