3 Speckle Noise Reduction and Enhancement for OCT Images
49
2.1. Initialization: Choose initial values for a
0
i , μ
0
i , σ
0
i ; where a i is the coefficient of ith component’s pdf and μ i and σ i are the pdf parameters of ith
component,
2.2. E-step: Compute responsibility factors as a N × 1 auxiliary variable that
for each observed data represents the likelihood that the observed data is
produced by component i.
2.3. M-step: Update parameters a
k
i , μ
k
i , σ
k
i using likelihood function maximization
2.4. Iteration: Substitute the updated parameters in the previous step to calculate pdf for each component ( f i (y))
2.5. Continue until the parameters satisfy convergence conditions
3. Calculate the CDF of each Normal-Laplace component using final values of μ i ,
σ i and σ n
4. Gaussianize each Normal-Laplace component using proposed Gaussianization
Transform
5. Use AMAP method to obtain a weighted summation of enhanced components
6. Apply exponential operator to find the contrast enhanced image.
3.3.2 Results
To analyze the proposed model for OCT contrast enhancement, two datasets from
different OCT imaging systems, Topcon 3D OCT-1000 and Cirrus HD-OCT (Carl
Zeiss Meditec, Dublin, CA), are used and three evaluation measures are applied. The
first two measures are Contrast to Noise Ratio (CNR), and Edge Preservation (EP)
and are computed based on methods described in [51]. In addition, another measure
Evaluation Measure of Enhancement (EME) which is introduced by Agaian et al.
[52] is used as a new measure for evaluating contrast enhancement. The proposed
method is compared with the Contrast-Limited Adaptive Histogram Equalization
(CLAHE) method and Agaian’s method for contrast enhancement (transform histogram shaping) [52]. All of these three methods are tested on both datasets. Both
the visual and numerical results illustrate the superiority of the proposed method.
Figure 3.4 displays an example of the proposed method in comparison to other methods. Furthermore, Table 3.4 shows the results averaged over all 130 Topcon B-scans
and the averaged result over 60 Zeiss B-scans is displayed in Table 3.5.
This improvement in contrast enhancement in compare to other methods stems
from considering a specific pdf for the OCT data and then converting it to a proportionate Gaussian distribution, thus providing the possibility of enhancement for
these images. These computations are useful for pre-processing retinal OCT images
to make them more suitable for visualization or further automated processing like
alignment or segmentation tasks. Hence, using this model as a preprocessing step
improves intra-retinal layer segmentation results and demonstrates the efficacy of
this model.
49
2.1. Initialization: Choose initial values for a
0
i , μ
0
i , σ
0
i ; where a i is the coefficient of ith component’s pdf and μ i and σ i are the pdf parameters of ith
component,
2.2. E-step: Compute responsibility factors as a N × 1 auxiliary variable that
for each observed data represents the likelihood that the observed data is
produced by component i.
2.3. M-step: Update parameters a
k
i , μ
k
i , σ
k
i using likelihood function maximization
2.4. Iteration: Substitute the updated parameters in the previous step to calculate pdf for each component ( f i (y))
2.5. Continue until the parameters satisfy convergence conditions
3. Calculate the CDF of each Normal-Laplace component using final values of μ i ,
σ i and σ n
4. Gaussianize each Normal-Laplace component using proposed Gaussianization
Transform
5. Use AMAP method to obtain a weighted summation of enhanced components
6. Apply exponential operator to find the contrast enhanced image.
3.3.2 Results
To analyze the proposed model for OCT contrast enhancement, two datasets from
different OCT imaging systems, Topcon 3D OCT-1000 and Cirrus HD-OCT (Carl
Zeiss Meditec, Dublin, CA), are used and three evaluation measures are applied. The
first two measures are Contrast to Noise Ratio (CNR), and Edge Preservation (EP)
and are computed based on methods described in [51]. In addition, another measure
Evaluation Measure of Enhancement (EME) which is introduced by Agaian et al.
[52] is used as a new measure for evaluating contrast enhancement. The proposed
method is compared with the Contrast-Limited Adaptive Histogram Equalization
(CLAHE) method and Agaian’s method for contrast enhancement (transform histogram shaping) [52]. All of these three methods are tested on both datasets. Both
the visual and numerical results illustrate the superiority of the proposed method.
Figure 3.4 displays an example of the proposed method in comparison to other methods. Furthermore, Table 3.4 shows the results averaged over all 130 Topcon B-scans
and the averaged result over 60 Zeiss B-scans is displayed in Table 3.5.
This improvement in contrast enhancement in compare to other methods stems
from considering a specific pdf for the OCT data and then converting it to a proportionate Gaussian distribution, thus providing the possibility of enhancement for
these images. These computations are useful for pre-processing retinal OCT images
to make them more suitable for visualization or further automated processing like
alignment or segmentation tasks. Hence, using this model as a preprocessing step
improves intra-retinal layer segmentation results and demonstrates the efficacy of
this model.