Page 263 - Computational Retinal Image Analysis

3 Automatic analysis of drusen and amd-related pathologies 261

AMD. The authors analyzed 12 eyes of 12 AMD patients with three clinical SD-
OCT devices: the Zeiss Cirrus (Carl Zeiss Meditec, Dublin CA), Topcon 3DOCT-
1000 (Topcon, Tokyo, Japan), and Heidelberg Spectralis (Heidelberg Engineering,
GmbH, Heidelberg, Germany) and five different scan patterns. It was concluded that,
while SD-OCT demonstrated excellent performance for visualizing drusen-related
RPE disease the automated segmentation showed distinct limitations. The authors
further investigated the potential of polarization-sensitive optical coherence tomog-
raphy (PS-OCT) for assessing drusen in AMD patients [78]. They examined 15 eyes
from 13 patients presenting with drusen with ophthalmoscope, fundus photography,
autofluorescence, and PS-OCT. They developed a segmentation algorithm based on
the polarization scrambling characteristics of the RPE to evaluate the drusen loca-
tion, area, and volume. Two independent graders identified the drusen in each B-scan
were identified by two independent expert graders. 91.4% of the drusen were de-
tected manually by both expert graders whereas the automated segmentation identi-
fied 96.5%. It was concluded that PS-OCT segmentation was superior to fundus
photography.

3.5 Analysis of other AMD lesions
The automated analysis of other features that are characteristic of AMD is under-
studied compared to the analysis of drusen. We provide a brief overview here for
completeness. There have been efforts to detect and segment areas of geographic atro-
phy in fundus auto-fluorescence (FA) images and SD-OCT. Deckert et al. proposed a
region-growing technique for the segmentation of GA in AF images [79] while Chen
et al. proposed a semi-automatic method for the detection of GA in SD-OCT images.
A restricted summed-voxel projection (RSVP) image is first produced, to which a
geometric active contour model is applied to segment the GA. The authors demon-
strated good performance on 111 SD-OCT images when compared with the propri-
etary method currently implemented in the Cirrus software (Carl Zeiss Meditec, Jena,
Germany) [80]. Schütze et al. proposed an automated lesion size detection method
based on Stokes vector analysis in patients with GA using PS-OCT [81]. Results on
29 eyes of 22 patients showed significant correlation with established fundus imag-
ing techniques such as infrared imaging, AF, and SD-OCT fundus. It is in general
challenging to analyze retinal layers in OCT images, however, Chiu S et al. recently
propose a segmentation framework based on graph theory and dynamic programming
for the segmentation of retinal layers in SD-OCT images with drusen and GA [82].
It is imperative for us to highlight some new progress in the analysis of CNV in a
new imaging technique called OCT angiography (OCTA). Zhang Q et al. proposed
a technique based on morphology analysis to segment CNV from OCTA images and
showed promising results on 27 eyes [83]. This may highlight the value of OCTA in
the management of AMD. Techniques for the analysis of some other common patho-
logical features of AMD, including leakage and neovascularization, have been covered
in other chapters, so we will not review these here. In addition, some recent reviews on
automated algorithms for the analysis of AMD have been published [84, 85].