Page 284 - Computational Retinal Image Analysis

282 CHAPTER 14 OCT fluid detection and quantification

Having an accurate fluid segmentation method available greatly facilitates the
fluid detection task as detection can be achieved by either thresholding the number
of segmented fluid voxels or training another classifier to learn to detect fluid from
the segmented fluid maps. When fluid segmentation is not available, machine learn-
ing and deep learning are well-suited methods for such a binary image classification
task, which determines whether fluid disease activity is present or not. Several meth-
ods have been proposed to differentiate retinas with macular edema from normal
retinas or retinas with nonexudative diseases, indirectly detecting the fluid. However,
the main downside of such image classification methods is that they only detect
general fluid activity, without distinguishing the various fluid types. This limits the
clinical findings of the classification as different fluid types are associated with dif-
ferent prognoses. In the following, we review the three fluid detection approaches as
mentioned previously.

3.1 Detection using image segmentation
A fluid segmentation method proposed by Xu et al. [34] was additionally evaluated
on the task of fluid detection in 30 OCT scans (Topcon) from 10 patients with nAMD
undergoing anti-VEGF treatment. It achieved an area under the curve (AUC) of 0.8
and 0.92, taking two different expert annotations as the gold standard. Chakravarthy
et al. [53] proposed a method using graph-based optimization and region growing to
identify fluid pockets but the implementation details were scarce, as the method is
part of a commercial system by Notal Vision for at-home monitoring. They reported
a diagnostic accuracy of 91% compared with the majority grading by three retinal
specialists on 142 OCT scans (Zeiss Cirrus) of eyes with nAMD.
The two methods mentioned previously aimed at detecting disease activity in
general without the possibility to detect the presence of each fluid type individu-
ally. As part of their IRF and SRF segmentation development, direct application in
the fluid detection task was demonstrated by Schlegl et al. [21]. ROC curves were
computed by varying the threshold over the number of fluid pixels segmented. In
the largest evaluation of individual fluid detection to date, it was evaluated against a
volume-level manual grading of 1200 OCT scans of eyes with nAMD (400), RVO
(400), and DME (400) with equal distribution of fluid presence and imaged with two
different OCT devices, that is, Cirrus (600) and Spectralis (600). AUCs for IRF and
SRF ranged from 0.91 to 0.97 and 0.87 to 0.98, respectively (Fig. 4). Detection of
SRF in DME cases was the only scenario where detection AUC fell below the 0.90
level, due to SRF being a rare occurrence in patients with DME and thus harder to
train and detect.
Recently, De Fauw et al. [27] developed a two-stage deep learning system for
diagnosis and referral based on OCT. In the first stage, they segment multiple imag-
ing biomarkers including IRF, SRF, and PED. The second stage uses the segmented
maps to train a CNN classifier to provide a differential diagnosis. The system has
been shown to be clinically applicable. Its performance indicating the need for refer-
ral was comparable to the one of retinal specialists but the detection performance of
individual fluid types was not evaluated.