Biomedical Engineering Reference
In-Depth Information
=30,asin
[
28
]. The exact same diffusion time (2:4 s) was used which correspond to 10 and
20 iterations with a step in time of 0:24 and 0:124 s to the 2D and 3D traditional
NCDF cases, respectively. For the improve adaptive version we considered the time
step as in (
12
) with a
The 2D-and 3D-NCDF filters were applied using k
D
10 and
D
0:75. However, to guarantee the stability of
the method one needs to consider the stability condition t
D
0:25 and b
D
0:1660.
As remarked in [
2
], it is advisable to consider more a conservative step in time
than the maximum theoretical limit, so we allowed only a maximum step in time
of 0:125.
To perform the quantitative measurement of the filter performance the MSE and
the metrics were computed for each of the 88 volumetric scans. The results from
the quantitative analysis are shown in Table
1
.
The average computing time for each volume of 200
cos =6
1;024 voxels OCT
data is of 97 min for the Lee filter, 2 min for the Perona-Malik (PM) (2D) filter, 3 min
for the 2D-NCDF [
28
] filter and 11 min for the 3D-NCDF here proposed. The total
diffusion time for PM, 2D-NCDF and the 3D filter is of 2:4 s.
These tests were performed using a 2.4 GHz Intel Core 2 Quad Q6600 CPU,
3 GB of RAM to run Matlab on the Ubuntu 10.04 operating system.
Figures
5
-
7
show the result of the application of the 2D-NCDF to a particular
B-scan and the same B-scan after the respective 3D-NCDF filtered data. In Fig.
5
it
is shown an healthy volunteer's retina example where is possible to note the good
performance of the 3D-NCDF filter in reducing speckle while preserving retinal
layers. Figures
6
and
7
show results from a CNV and an AMD cases, respectively.
For easy of comparison a region is zoomed in and shown in insets using a pseudo-
color (Fig.
7
). Please note the good preservation and better visualization of a retinal
membrane after the 3D-NCDF application and the increased homogeneity of the
vitreous.
To quantify the filters performance in homogeneous areas, one B-scan of each
volumetric data was selected and the ENL was computed on a vitreous region. The
results are presented in Table
2
.
Moreover, we also compared the 3D traditional NCDF with the improved version
described in Sect.
2.3
, which is illustrated in Fig.
8
.
Performance metrics for the comparison of the traditional and improved 3D-
NCDF were also performed for three subjects, as a proof-of-concept, for a short
diffusion time of T
200
0:75 s. We obtained a higher ENL for the improved method
(ENL
Impr
) in comparison with the traditional (ENL
Trad
) one, both in high intensity
areas as the nerve fiber layer and the retinal pigment epithelium (ENL
Impr
D
D
176:26
˙
83:49 and ENL
Trad
D
42:59
˙
5:54) and in low intensity areas as the
vitreous (ENL
Impr
0:78). This shows
that in high intensity areas the improved method quadruplicates the smoothness
(measured by the ENL), while for low intensity areas it becomes ten times
smoother. As expected the improved method is more conservative in tissue regions
(high intensities).
D
106:55
˙
3:02 and ENL
Trad
D
10:71
˙
