Page 278 - Computational Retinal Image Analysis

276 CHAPTER 14 OCT fluid detection and quantification

these hurdles to reach the level of clinically applicable performance. There are sev-
eral scenarios where automated detection and quantification of fluid may become
clinically relevant. These include, for instance, the automated screening and triage of
OCT scans for the presence of macular disease, the assessment of fluid response and
recurrence during a course of therapy, or the evaluation of fluid in order to provide
prognostic information.
In the remainder of this chapter, we review the methods and state of the art in au-
tomated OCT fluid quantification (Section 2) and detection (Section 3), present two
clinically relevant analyses that result from having fluid quantification tools avail-
able (Section 4), and, finally, summarize the main findings and conclude the chapter
(Section 5).

2 OCT fluid quantification
Being able to quantitatively measure the amount of fluid present is expected to play
an increasingly important clinical role in effective managements of patients with
exudative macular diseases. There have already been retreatment protocols sug-
gested that rely on knowing the fluid size, for example, Arnold et al. [12] tested a
relaxed retinal fluid regimen that tolerated incomplete resolution of SRF ≤200 μm.
Furthermore, the ability to quantify the three fluid types individually is an important
step to enable prognostic decision support systems, as presented in Section 4.
Fluid quantification relies on having accurate OCT image segmentation methods.
In addition, due to the large amount of fluid pockets typically present in a scan, the
methods should be fully automated as even semiautomated approaches would be
too labor intensive and consequently of very limited practical use. Nowadays, such
image segmentation models are exclusively based on convolutional neural networks
(CNN) due to their ability to learn without feature handcrafting, and their continual
improvement with the increasing number of annotated datasets. Fast execution of the
trained CNN models is also essential for their deployment in the clinic, where only
limited time is available for planning between the OCT acquisition and the treatment.

2.1 Segmentation using supervised learning
As soon as CNN deep learning models were discovered to be highly successful at
image recognition [14], they were redesigned to address the semantic segmentation
task by making CNN learn the mapping from raw intensity images to dense pixel-
wise tissue class labels. A popular initial approach was based on patch classification,
where each pixel was classified separately using an image patch surrounding it. The
main reason for using patches was CNN’s components made of fully connected lay-
ers that required a fixed input size image. Schlegl et al. [15] used a CNN early on
with a patch-based pixel classification for the task of fluid segmentation on OCT. In
this pioneering work, the multiscale CNN model was able to differentiate between
IRF and SRF types and operate in a weakly supervised setting.