Page 281 - Computational Retinal Image Analysis

2 Oct fluid quantification 279

2.1.1 Preprocessing and postprocessing
Denoising
OCT scans have low SNRs due to the high level of speckle noise. Standard efficient
preprocessing approaches, such as Gaussian and gamma noise filtering, median and
bilateral filtering [29], or total variation-based [30] denoising, are used for noise
suppression. However, even without any preprocessing CNN generally showed very
good robustness to noise and the degradation of image quality caused by severe reti-
nal abnormalities attenuating the OCT signal has a stronger detrimental impact than
the speckle.
Retina and layer segmentation
The retina can be effectively segmented due to the strong signal and image gradient
of the inner limiting membrane (ILM) at the top of the retina and the bottom RPE
interface [31, 32]. These interfaces are also often made available by built-in OCT
device algorithms. Thus, a region of interest denoting the retina can be used as both a
preprocessing step to limit the segmentation to this region and a postprocessing step
to remove false-positive labels.

Data augmentation
Data augmentation is an effective technique to inject invariance to a desired set of geo-
metric and intensity transformations into the deep learning models. It typically involves
random cropping of a subimage from a B-scan, random rotation, mirroring and intro-
ducing multiplicative speckle noise as well as applying affine and elastic geometric
transformations jointly over the image inputs and gold standard segmentations.

2.1.2 Traditional machine-learning and nonmachine-learning approaches
Before the appearance of deep learning segmentation models, traditional machine
learning was applied to provide a classification of OCT voxels based on a set of
hand-crafted features describing the local appearance. Such voxel classes then pro-
vided a prior for subsequent graph-based or variational-based image segmentation
methods. Chen et al. [33] presented a fully 3D approach based on voxel classification
followed by a graph-cut-based segmentation. Xu et al. [34] proposed a retinal voxel-
classification approach stratified by presegmented layers. Wang et al. [35] proposed a
combination of a fuzzy C-means for initialization followed by evolution of a level-set
contour in three orthogonal OCT planes.
A simultaneous joint segmentation of retinal layers and fluid was a long-pursued
goal, as it overcomes error propagation when segmenting step-wise retinal layers
first and then fluid regions. Such a voxel classification approach followed by a refine-
ment with graph-cut and dynamic programming was proposed by Chiu et al. [24].
Novosel et al. [36] developed loosely coupled-level sets to jointly segment fluid and
retinal layers by modeling the fluid as an additional space-variant layer. A purely
data-driven model with minimum a priori constraints was proposed by Montuoro
et al. [37] based on auto-context [38] to jointly segment retinal fluid and layers by
learning their mutual interaction properties.