Page 282 - Computational Retinal Image Analysis

280 CHAPTER 14 OCT fluid detection and quantification

PED segmentation was initially approached as a layer segmentation problem as
it corresponds to the deformation of a space between the RPE and Bruch’s mem-
brane (BM). PED could then be easily identified from the resulting layer thickness
map. Shi et al. [39] developed a multisurface segmentation using graph search by
specifying different constraints on surface smoothness corresponding to the RPE
and BM. Sun et al. [40] first estimated BM surface from the convex hull of the RPE,
followed by a shape-constrained graph cut to obtain the final PED segmentation. Wu
et al. [41, 42] proposed a 3D method to segment and differentiate between SRF and
PED fluid pockets. They combined texture, intensity, and thickness scores to build a
voxel-level fluid probability map and then applied a continuous max-flow to obtain
the segmentations.

2.2 Segmentation using weakly supervised and unsupervised
learning
A supervised learning approach for semantic segmentation requires substantial ef-
fort in producing a large-scale dataset of manual pixel-wise annotations needed for
the training. As an alternative, weakly supervised techniques focus on achieving
segmentation based on OCT or image region-level information of fluid presence.
Finally, unsupervised learning approaches based around the concept of anomaly
detection require a training set of healthy retinas only. They use a two-step process
where first normal shape and appearance is learned and then anomalies such as
fluid can be detected as deviations from the norm. This reflects the natural study
process of medical students, who first learn what a healthy tissue looks like and
subsequently gain the ability to identify pathologies deviating from this normal
appearance.
An early automated approach addressed the fluid segmentation problem as a lo-
cal anomaly detection based on retinal texture and thickness properties [43]. After
learning normal variability from images of healthy eyes, the method was applied to
determine 2D en-face footprints of fluid-filled regions, although a 3D localization
was missing. Schlegl et al. [15] used the approximate spatial location of fluid in the
form of its retinal layer group and centrality and reached the performance equal to
≈85% of the fully supervised approach. As a by-product of interpreting image clas-
sification results, fluid-related regions have been identified with moderate accuracy
[44–46]. Seeböck et al. [47] trained a multiscale deep denoising autoencoder [48] on
healthy images, and used a one-class support vector machine (SVM) that identified
anomalies in new data. While Schlegl et al. [49] used generative adversarial networks
[50] to embed image patches into a low-dimensional space where the deviations from
the manifold of healthy patches could be measured.

2.3 Evaluation
A frequently used for evaluating segmentation performance is a DSC, corresponding
to the F1 score, the harmonic average between precision and recall. It is a measure of