80
L. Fang and S. Li
SURE, K-SVD, and BM3D methods are chosen to the default values as in [28, 39,
40].
All the methods are tested on datasets from 17 eyes from 17 subjects with and
without non-neovascular age-related macular degeneration (AMD) enrolled in the
A2A SDOCT study. Ten datasets are selected from normal subjects while the remaining datasets are from AMD subjects. All the datasets are acquired using the SDOCT
imaging systems from Bioptigen, Inc. (Research Triangle Park, NC). For each patient,
two kinds of SDOCT scans were acquired. (1) A square (~6.6 × 6.6 mm) volume
scan with 1000 A-scans and 100 B-scans, including the fovea. (2) An azimuthally
repeated scan with 1000 A-Scans and 40 B-Scans targeted at the fovea. Each Ascan is cropped to achieve B-Scans of size 280 × 1000, which only includes the
informative areas (excluding the smooth dark areas deep below the choroid or in
the vitreous). For the repeated 40 B-scans, the ImageJ (software; National Institutes
of Health, Bethesda, Maryland, USA) StackReg Registration plug-in [41] is firstly
used for the registration and then the registered B-scans are averaged to create the
corresponding noiseless B-scan.
The mean-to-standard-deviation ratio (MSR) [42], contrast-to-noise ratio (CNR)
[43], and peak signal-to-noise-ratio (PSNR) are adopted as the quantitative metrics
to evaluate the performance of different denoising methods. The MSR and CNR are
computed as,
MSR
μ f
σ f
,
(4.6)
CNR
|μ f − μ b |
0.5(σ
2
f + σ
2
b )
,
(4.7)
where μ b and σ b are the mean and the standard deviation of the background region
(e.g. red box #1 in Fig. 4.5), while μ f and σ f are the mean and the standard deviation
of the foreground regions (e.g. red box #2–6 in Fig. 4.5). The PSNR is global metric,
which is computed as,
PSNR 20 · log 10
⎛
⎜
⎜
⎝
Max R
1
H
H
h1
R h − ˆ
R h
2
⎞
⎟
⎟
⎠ ,
(4.8)
where R h is the hth pixel in the reference noiseless image R, ˆ
R h represents the hth
pixel of the denoised image ˆ
R, H is the total number of pixels, and Max R is the
maximum intensity value of R. Since there is no ideal “noiseless” image available,
the averaged foveal image is used as a noiseless approximation to the corresponding
foveal B-scan from the noisy volumetric scan.
Figure 4.5 shows two raw SDOCT retinal images (from a normal and an AMD
subject) and their visually denoised results obtained from various denoising methods.
Since the boundaries between retinal layers contain important pathologic information
L. Fang and S. Li
SURE, K-SVD, and BM3D methods are chosen to the default values as in [28, 39,
40].
All the methods are tested on datasets from 17 eyes from 17 subjects with and
without non-neovascular age-related macular degeneration (AMD) enrolled in the
A2A SDOCT study. Ten datasets are selected from normal subjects while the remaining datasets are from AMD subjects. All the datasets are acquired using the SDOCT
imaging systems from Bioptigen, Inc. (Research Triangle Park, NC). For each patient,
two kinds of SDOCT scans were acquired. (1) A square (~6.6 × 6.6 mm) volume
scan with 1000 A-scans and 100 B-scans, including the fovea. (2) An azimuthally
repeated scan with 1000 A-Scans and 40 B-Scans targeted at the fovea. Each Ascan is cropped to achieve B-Scans of size 280 × 1000, which only includes the
informative areas (excluding the smooth dark areas deep below the choroid or in
the vitreous). For the repeated 40 B-scans, the ImageJ (software; National Institutes
of Health, Bethesda, Maryland, USA) StackReg Registration plug-in [41] is firstly
used for the registration and then the registered B-scans are averaged to create the
corresponding noiseless B-scan.
The mean-to-standard-deviation ratio (MSR) [42], contrast-to-noise ratio (CNR)
[43], and peak signal-to-noise-ratio (PSNR) are adopted as the quantitative metrics
to evaluate the performance of different denoising methods. The MSR and CNR are
computed as,
MSR
μ f
σ f
,
(4.6)
CNR
|μ f − μ b |
0.5(σ
2
f + σ
2
b )
,
(4.7)
where μ b and σ b are the mean and the standard deviation of the background region
(e.g. red box #1 in Fig. 4.5), while μ f and σ f are the mean and the standard deviation
of the foreground regions (e.g. red box #2–6 in Fig. 4.5). The PSNR is global metric,
which is computed as,
PSNR 20 · log 10
⎛
⎜
⎜
⎝
Max R
1
H
H
h1
R h − ˆ
R h
2
⎞
⎟
⎟
⎠ ,
(4.8)
where R h is the hth pixel in the reference noiseless image R, ˆ
R h represents the hth
pixel of the denoised image ˆ
R, H is the total number of pixels, and Max R is the
maximum intensity value of R. Since there is no ideal “noiseless” image available,
the averaged foveal image is used as a noiseless approximation to the corresponding
foveal B-scan from the noisy volumetric scan.
Figure 4.5 shows two raw SDOCT retinal images (from a normal and an AMD
subject) and their visually denoised results obtained from various denoising methods.
Since the boundaries between retinal layers contain important pathologic information