Page 137 - Computational Retinal Image Analysis

130 CHAPTER 7 OCT layer segmentation

5.3 Boundary detection methods
The second large set of methods predicts cell layer boundaries first, instead of direct
semantics segmentation. Boundaries are usually optimized with graph-based meth-
ods and in some cases transformed to pixel-wise segmentations.
An example of this approach was conducted by Gopinath et al. [21], who use a
Fully-Connected—CNN to predict the pixel-wise class for nine retinal layers. They
use an additional FCNN to identify edges, cut it into eight vertical bands and refine
the segmentation with a BLSTM to reproduce a final consistent segmentation. In
their work Fang et al. [33] find boundaries of nine retinal layers. A CNN is trained
on patches to infer a boundary probability map. Applying Dijkstra’s algorithm on the
map they create a partition for each boundary layer. The results are evaluated on 117
OCT from 39 participants with non-exudative AMD.
Similarly, Liu et al. [34] propose a method that considers accurate boundary
probabilities in order to perform the task of seven retinal layer segmentation at mul-
tiple scales and multi-dimensional features. The boundaries are then provided to a
graph segmentation algorithm that refines the final result with Dijkstra’s algorithm
(see also Fang et al. [33]).
Hamwood et al. [35] evaluate whether patch size affects the quality of boundary
segmentation in a significant way, or not. Patch size indeed affects the outcome, but
also architecture and number of target classes. This suggests that even for boundary-
based methods, specific results for the segmentation improve with more classes.
Enough classes guide the learning process to make a better distinction, reducing the
number of false positives.
Over the last few years segmentation of retinal layers has also served as the
precursor for segmentation of other regions in retinal OCT, such as the region of
the choroid, critical in the development of diseases such as diabetic macular edema
(DME). In this area we would find methods such as developed by Masood et al.
[36], where the efforts went in the direction of providing a segmentation of the
Bruch’s membrane (BM) using baseline thresholding and choroid using a combi-
nation of deep learning based neural network for patchwise classification. These
approaches together with traditional medical image processing techniques and stan-
dard pre-processing techniques facilitate the task of segmentation. The inhomoge-
neity of the choroid in some cases makes the problem of segmentation of choroid a
real challenge.
Sui et al. [37] focused on tackling the same problem using a graph-search seg-
mentation technique focused on identifying edges combined with deep learning.
In this work the authors learn the optimal graph weight via a data-driven strat-
egy which is reached by using a new multi-scale end-to-end CNN architecture.
Alonso-Caneiro et al. [38] used a CNN (ReLayNet) providing a probability map
that marked the predicted location of the boundaries. The map was subsequently
refined and adjusted using a graph-search approach to extract the boundaries
from the probability maps. In this case the authors highlighted that the graph-
search algorithm used to extract boundaries would often decrease the quality of
the results.