Page 214 - Computational Retinal Image Analysis

210 CHAPTER 11 Structure-preserving guided retinal image filtering

3.2 Evaluation metrics
To evaluate the performance of SGRIF, we first compute how it affects the image
contrast in the area of the optic disc. Two evaluation metrics, namely the histogram
flatness measure (HFM) and the histogram spread (HS) are used to evaluate
performance [63]:

i (∏ n x i) 1/ n (27)
HFM = =1 ,
1 ∑ n x
n i=1 i

where x i is the histogram count for the ith histogram bin and n is the total number of
histogram bins.
)
(3rd quartile-1st quartile of histogram
HS = . (28)
(Maximum-minimum)) of the pixel value range
Another measurement is the mean variability of the local luminosity (VLL) [64]
throughout the optic disc. Given an image I, divided into N × N blocks B i, j , i, j = 1,
…, N with equal sizes. VLL is computed as
1 1 i=1 j=1 − (29)
2
µ
VLL = ∑ ∑(( ij − ),
,)
I
N I 2 N N
−
where I stands for the mean intensity of the entire image and μ(i, j) for the mean of
block B i, j . For all the above metrics, a high value indicates a better result.
3.3 Results
Table 1 summarizes the results observe that the proposed method improves the HFM,
HS, and VLL by 5.9%, 4.3%, and 134.8%, respectively, compared with original
images. GIF improves VLL, but does not increase HFM and HS. This is because GIF
oversmooths some regions close to flat-contrast ones, and reduces the dynamic range
of the histograms. Fig. 5 shows results from five sample images of optic disc. As we
can see, the proposed SGRIF enhances the contrast between the optic cup and the
neuroretinal rim while the improvement by GIF is less clear. Visually, it is difficult
to tell if GIF has oversmoothed some regions but we will show the differences from
subsequent analysis in the next section.

3.4 Application
To evaluate how the declouding benefits the retinal analysis tasks, we conduct the
following experiments: (1) deep learning-based optic cup segmentation; (2) sparse
learning-based CDR measurement.