![](bg4.png)
with reduced dimensions) to its base, where the final
layer is detected in the original OCT image.
When using a 3-D model, the process is analo-
gous: pyramid is built with 3-D OCT images, and
the layer is detected in all of them during the descent
stage. In the base of the structure, only the layer seg-
mented in the central image of the group is preserved,
since it has the contribution of its adjacent images.
4 RESULTS AND DISCUSSION
Different experiments have been designed in order to
find the best pyramidal structures for the segmenta-
tion of first and sixth layers and compare its results
with those obtain with the original graph-based seg-
mentation method. The data set used in the experi-
ments is composed by 18 OCT sequences, including
both healthy and diabetic retinopathy patients. Each
sequence consists of 128 images, which makes a total
of 2304 images.
These images were captured using Cirrus HD-
OCT, with Spectral Domain Technology (Zeiss), at
a resolution of 334x334 pixels. All the results have
been obtained using an Intel Core TM 2 Duo proces-
sor (2,4 GHz) and a RAM of 2 GB.
Three experiments have been designed for each
layer separately, since their detection is independent.
The first one has the purpose of studying the pyrami-
dal height, in order to determine the number of levels
that achieves a successful segmentation with mini-
mum cost. In the second experiment, the feasibility
of a descending method without intermediate levels
is studied, avoiding the segmentation in certain levels
of the structure with the aim of optimizing even more
the pyramidal process. Finally, in the last experiment,
results of the best pyramids obtained previously are
compared with those provided by the original seg-
mentation method.
Given that the aim of this work is optimizing the
efficiency of the graph-based segmentation method,
maintaining the accuracy in the segmentation results
is essential. With that purpose, experiments 1 and
2 are done in two steps: firstly the feasible pyrami-
dal heights are studied, in order to consider pyramids
with enough number of levels (but not excessive) to
achieve appropriate segmentation results. After that,
efficiency is studied for those structures. Thus, at each
experiment, both effectiveness (successful segmenta-
tion) and efficiency (in terms of consumed time and
memory) are considered.
It is necessary to take into account that, in these
experiments, layer 1 is segmented using a 2-D graph,
since its location is evident in the images, and the in-
formation from adjacent images is not needed. Re-
garding layer 6, a 3-D graph is needed, due to the dis-
continuities in this layer, in addition to the fact that the
borderline between RPE and Choroids is not well de-
fined in the images. The 3-D graph for layer 6 is built
with overlapping groups of 3 consecutive images.
Parameters involved in the methodology have
been extracted empirically. The factor scale sc used
in this work is set to 0.5, so dimensions for the image
located in the level i of the pyramid are the half of the
image immediately below (level i − 1).
4.1 Experiment 1: Determining the
Optimal Pyramid Height
In this set of experiments, pyramids of different height
are evaluated, in order to determine the number of le-
vels that allows to obtain a successful segmentation
of each layer with minimum computational cost. This
task is done in two steps: firstly, a coarse study of the
heights than be considered to built pyramids for both
layers is done. After that, time and memory required
by the pyramid-based approach using these heights to
detect layers 1 and 6 is deeply studied.
Regarding layer 1, preliminary experiments show
that segmentation does not present mistakes for pyra-
mids of up to 4 levels, whereas layer 6 can not be cor-
rectly segmented with structures higher than 3 levels.
Once that the maximum number of levels for each
pyramid is known, performance is studied to select
the most appropriate structure. Table 1 shows results
obtained, while Figure 5 reflects them graphically.
Table 1: Time and memory consumed in the segmentation
task using pyramids of different heights, both expressed
as mean (standard deviation). Best time results have been
highlighted.
h = 2 h = 3 h = 4
Layer 1
Time (s) 58.44 (12.88) 30.22 (6.03) 29.67 (5.57)
Memory (MB) 103.22 (1.56) 103.83 (1.86) 104.00 (1.88)
Layer 6
Time (s) 1216.94 (557.36) 562.17 (93.10) –
Memory (MB) 115.11 (4.79) 123.22 (4.87) –
4.1.1 Discussion
First aspect to be observed is that the consumed me-
mory does not change significantly using different le-
vels in the pyramid, so it is not a decisive factor in the
selection of the best structure. However, it is observed
that layer 6 requires more memory than the layer 1.
Obviously, this is because of the 3-D model used to
detect that layer.
The number of levels has more evident repercus-
sions in the computation time. As Figure 5 (a) ex-
poses, time required to detect layer 1 decreases con-
siderably when h = 3 , whereas the small difference
ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence
718