9 Choroidal OCT Analytics
213
9.2.1 Problem Setup and Solution Approaches
Choroidal thickness distribution estimated from OCT images emerged as an important metric in disease management [14]. Consequently, estimation accuracy has
assumed a vital role in ensuring accurate diagnostic outcome [17]. Choroidal thickness measurements have in turn been used to obtain choroidal volume. So far, such
thickness measurements have been performed by experts by manually delineating
the choroid inner and outer boundaries and then taking the difference. Such manual
analysis of OCT scans is time consuming, laborious as well as susceptible to fatigueinduced error. Manual estimation of choroidal volume is rarely performed in view of
the inordinate time and effort involved. Against this backdrop, automated segmentation of choroid layer could be crucial in reducing professional effort and time per
subject, potentially allowing more subjects to obtain specialized medical attention.
Further, choroidal volume now being routinely used in addition to the usual choroidal
thickness. Automation would also avoid human error induced by fatigue and tedium.
Accordingly, we propose a novel automated algorithm for choroid segmentation and
related thickness and volume measurements.
Since the last few years, automation of choroid segmentation have been attracting
considerable attention. In view of the eye physiology (Fig. 9.1), choroid segmentation
consists of two tasks: detecting (i) choroid inner boundary (CIB) and (ii) choroid outer
boundary (COB). Of these, the first task is relatively well posed because the RPE,
defining the CIB, is significantly brighter than adjacent layers. Indeed, the gradientbased approach in various flavors has proven accurate not only in detecting CIB [18,
19], but also in the related problem of detecting boundaries between successive retinal layers with well-defined brightness transition [20, 21]. Accordingly, we shall also
adopt a gradient-based approach for CIB detection. In contrast, the task of detecting
the COB poses considerable challenge. This happens because the COB is essentially
a notional divide between the choroidal granularity and the scleral uniformity, which
is not defined by marked variation in brightness, and often open to subjective interpretation. Even so, gradient-based deterministic methods have been suggested for
COB detection [22, 23]. However, statistical methods appear more suitable to handle
the inherent uncertainties involved. Accordingly, machine learning [24, 25] as well
as gradient-based probabilistic methods [19] have been attempted. Yet, aforementioned attempts does not directly exploit the structural transition from granularity to
uniformity across the COB. Against this backdrop, SSIM-based method has been
proposed to quantify the structural dissimilarity between choroid and sclera using the
yardstick of structural similarity (SSIM) index to find an initial estimate of the COB,
followed by Hessian analysis to remove the discontinuities in the initial estimate [26].
However, the resulting boundary, although adequately separates the choroidal vessels from scleral uniformity, generally is not smooth and deviates substantially from
smooth boundaries manually drawn by experts [27]. With a view to obtaining close
match with the latter, tensor voting is employed to achieve the desired smoothing.
From a clinical perspective, even accurate estimate of choroidal thickness on its own
could sometimes be inadequate in assessing choroidal involvement in chorioretinal
213
9.2.1 Problem Setup and Solution Approaches
Choroidal thickness distribution estimated from OCT images emerged as an important metric in disease management [14]. Consequently, estimation accuracy has
assumed a vital role in ensuring accurate diagnostic outcome [17]. Choroidal thickness measurements have in turn been used to obtain choroidal volume. So far, such
thickness measurements have been performed by experts by manually delineating
the choroid inner and outer boundaries and then taking the difference. Such manual
analysis of OCT scans is time consuming, laborious as well as susceptible to fatigueinduced error. Manual estimation of choroidal volume is rarely performed in view of
the inordinate time and effort involved. Against this backdrop, automated segmentation of choroid layer could be crucial in reducing professional effort and time per
subject, potentially allowing more subjects to obtain specialized medical attention.
Further, choroidal volume now being routinely used in addition to the usual choroidal
thickness. Automation would also avoid human error induced by fatigue and tedium.
Accordingly, we propose a novel automated algorithm for choroid segmentation and
related thickness and volume measurements.
Since the last few years, automation of choroid segmentation have been attracting
considerable attention. In view of the eye physiology (Fig. 9.1), choroid segmentation
consists of two tasks: detecting (i) choroid inner boundary (CIB) and (ii) choroid outer
boundary (COB). Of these, the first task is relatively well posed because the RPE,
defining the CIB, is significantly brighter than adjacent layers. Indeed, the gradientbased approach in various flavors has proven accurate not only in detecting CIB [18,
19], but also in the related problem of detecting boundaries between successive retinal layers with well-defined brightness transition [20, 21]. Accordingly, we shall also
adopt a gradient-based approach for CIB detection. In contrast, the task of detecting
the COB poses considerable challenge. This happens because the COB is essentially
a notional divide between the choroidal granularity and the scleral uniformity, which
is not defined by marked variation in brightness, and often open to subjective interpretation. Even so, gradient-based deterministic methods have been suggested for
COB detection [22, 23]. However, statistical methods appear more suitable to handle
the inherent uncertainties involved. Accordingly, machine learning [24, 25] as well
as gradient-based probabilistic methods [19] have been attempted. Yet, aforementioned attempts does not directly exploit the structural transition from granularity to
uniformity across the COB. Against this backdrop, SSIM-based method has been
proposed to quantify the structural dissimilarity between choroid and sclera using the
yardstick of structural similarity (SSIM) index to find an initial estimate of the COB,
followed by Hessian analysis to remove the discontinuities in the initial estimate [26].
However, the resulting boundary, although adequately separates the choroidal vessels from scleral uniformity, generally is not smooth and deviates substantially from
smooth boundaries manually drawn by experts [27]. With a view to obtaining close
match with the latter, tensor voting is employed to achieve the desired smoothing.
From a clinical perspective, even accurate estimate of choroidal thickness on its own
could sometimes be inadequate in assessing choroidal involvement in chorioretinal