Page 288 - Computational Retinal Image Analysis

286 CHAPTER 14 OCT fluid detection and quantification

The recently organized RETOUCH challenge [52] aimed at evaluating the state-
of-the-art performance of both automated detection and segmentation of individual
fluid types. After training on 70 OCT scans, the teams provided the detection results
of 42 test cases for each fluid individually. The performance of the teams was within
the interobserver variability. Even though detecting fluid manually can be difficult
and subjective, the reported interobserver agreements were high. Chakravarthy et al.
[53] compared the accuracy of four graders and found the agreement of each with the
majority of the other graders was 93%. Agreement between the gradings of the two
medical centers in the RETOUCH fluid detection challenge was similar with 95%
across all three fluid types. Thus, the resulting gold standard gradings are considered
to be of high quality, which is crucial for both training and validating supervised
data-driven machine and deep learning fluid detection methods.

4 Clinical applications
A number of clinically relevant applications are enabled by having fluid detection
and quantification tools available. Prediction of VA outcomes after a period of anti-
VEGF treatment from a set of spatiotemporal OCT features and clinical biomarkers
was proposed for all three main exudative macular diseases: nAMD [66], DME [67],
and RVO [68]. The OCT biomarkers corresponding to the retinal layer thicknesses,
volume, and area covered by the retinal fluid were spatially described by their mean
ETDRS grid values. Similarly, Bogunovic et al. [69] proposed prediction of anti-
VEGF treatment requirements in the following 2 years from a set of spatiotemporal
OCT biomarkers obtained during the initiation phase and showed that the automated
prediction performance was comparable to or better than that of a retina specialist.
DeepMind developed a clinically applicable system for diagnosis based on OCT that
relies on semantic segmentation of OCT volumes, with all three fluid types, IRF,
SRF, and PED, included [27]. In this section, we describe two applications of clinical
decision support that rely on fluid being accurately quantified.

4.1 Structure function

VA or with corrections (e.g., for short-sightedness) best-corrected visual acuity
(BCVA) is a common measure of retinal function in terms of the resolving power of
the eye. BCVA is one of the most important measures to assess disease progression
and therapy response in clinical routine. It is measured by counting the total number
of letters a person is capable of reading from a standardized chart with the size of
letters decreasing per line. The logMar chart is an example [70].
Clinical studies have already demonstrated that the qualitative characterization
of fluid types (presence/absence) affects individual VA and changes in VA during
treatment are strongly affected by individual baseline levels due to a ceiling effect [6,
71]. Thus, a capability to quantify in addition to detect such fluid would enable such
quantitative values to be correlated with visual function, superseding the traditional