Page 131 - Computational Retinal Image Analysis

124 CHAPTER 7 OCT layer segmentation

day. Some manual segmentation datasets such as that found in Chiu et al. [1] or Chiu
et al. [2] are publically available, but remain few at this time. Critically, most papers
that use manual delineations are validated on datasets, public or not, with tens or
hundreds of volumes.
Yet, from their anatomical definitions, retinal layers viewed in OCT images have
been subject to interpretation ever since OCT devices were introduced in ophthalmol-
ogy. This is mainly due to obfuscations engendered by poor signal-to-noise and image
artifacts. Beyond this, the ability of consistently define each retinal layer is extremely
challenging when considering pathological cases (e.g., diabetic and especially mac-
ula degeneration patients). For instance, when retinal layers atrophy and retinal layer
structures dissipate, spatial definitions of where layers begin and end are broadly ill-
defined. In addition, inter-observer variability plays an important role in the definition
of such gold standards. The work of Chiu et al. [1] was one of the first to explore how
inter-grader variability could impact performance of automated methods and the work
of Dubose et al. [3] explores a theoretical framework to bound the accuracy of any
such derived gold standard. Hence the notion of consistent standards for retinal layer
segmentation in OCT images remains a challenge and an open problem.
Given a gold standard, or ground truth, layer segmentation and an automatically
generated segmentation, a variety of different performance measures have been used
in recent years to validate the quality of produced layers. One of the most com-
mon measures of performance involves the per-pixel differences between layer sur-
face produced and its corresponding ground truth. From this, the mean unsigned or
signed errors can be computed for each layer [4–6]. Others have also computed the
pixelwise mean absolute difference (MAD) and root mean squared error (RMSE)
between layer boundaries [7]. Alternatively, the Dice score and Chamfer distance
between predicted and reference layers have also been used [7a, 8].

3 Intensity based methods
In its infancy, research focused on retinal layer segmentation methods primarily dealt
with significant noise levels in OCT images and at its core with finding the boundar-
ies or edges of the layers taking into account different constraints and anatomical
properties. For most methods prior to 2010, this involved developing methods that
could segment layers while overcoming signal-to-noise challenges that appeared in
individual OCT cross sections, or B-scans. Such methods heavily relied on model
based approaches, as opposed to more recent data-driven methods that have become
common today.
Simple approaches that attempted to segment layers by finding layer boundar-
ies in individual columns, or A-scans, of OCT images using pixel thresholds, hand-
crafted filtering and correlation exploitation techniques [9–11] were perhaps the first
of their kind conceptually to segment layers. A number of different works then built
on top of this, by providing more anatomically specific methods for the macula [12,
13] or the optic nerve [14].