1. Color photography of the ocular fundus is commonly used to diagnose retinal diseases but manual analysis is time-consuming and computer-aided analysis can help address this.
2. However, color retinal images suffer from non-uniform illumination and contrast variability due to natural variations in the retina and imaging conditions. This can negatively impact machine learning methods for diagnosis.
3. The study proposes a two-stage method to normalize luminosity and contrast in intra- and inter-retinal images with respect to a standard reference image. When applied to a diabetic retinopathy severity classification task using deep learning, it improved accuracy from 57% to 75.6%.
Luminosity and Contrast Normalization in Color Retinal Images
1. 1 Shiley Eye Institute , Department of Ophthalmology, University of California at San Diego, La Jolla, CA
2 Email : eshahria@ucsd.edu , Contact : 619-549-3796
ResultsMethod
Conclusion
Ehsan Shahrian Varnousfaderani,1,2 Siamak Yousefi,1 Akram Belghith,1 and Michael H. Goldbaum1
Color photography of the ocular fundus is a widely used imaging
modality that permits the non-invasive analysis of retina. Images
are routinely acquired by ophthalmologists and trained medical
professionals to diagnose and monitor progression of retinal
diseases, including age-related macular degeneration and diabetic
retinopathy, that are leading cause of blindness worldwide.
Manual analysis of ocular fundus images : time-consuming,
expensive, and sometimes inaccurate.
Computer aided ocular fundus image analysis: fast, accurate and
inexpensive
But
Color retinal images suffer from non-uniform illumination and
contrast variability due to the large natural variations in retinal
pigmentation and vignette and different models of cameras,
magnification, and image quality. This problem can highly affect
performance of machine learning or deep learning methods that
are used for diagnosis.
The proposed two-stage method makes illumination uniform and enhances color in
the image with respect to the standard reference image, which improves
performance of deep learning and other machine learning methods.
[9784-130]
Purpose
Introduction
Proposing a new method to normalize luminosity and contrast
in both intra- and inter- retinal images with respect to standard
reference image.
Image normalization in two stages
1st : Uniform illumination
* For compliance with the way humans perceive color, the illumination normalization is done in perceptually uniform color space like the LUV (L is lightness and U and V are chromatic values) color space.
𝐼 𝑢𝑛𝑖𝑓𝑜𝑟𝑚 𝑥, 𝑦 =
𝐼 𝑥, 𝑦 − 𝜇 𝑙𝑜𝑐𝑎𝑙
∗
𝑥, 𝑦
𝜎𝑙𝑜𝑐𝑎𝑙
∗
(𝑥, 𝑦
𝜇 𝑙𝑜𝑐𝑎𝑙
∗ and 𝜎𝑙𝑜𝑐𝑎𝑙
∗
represent the
local mean and standard
deviation on Background image
𝐼 𝑏𝑎𝑐𝑘 𝑥, 𝑦 = 𝐼(𝑥, 𝑦 𝑖𝑓 𝐷 𝑥, 𝑦 < 𝑡
𝐼 𝑏𝑎𝑐𝑘 𝑥, 𝑦 = 𝜇(𝑥, 𝑦 𝑖𝑓 𝐷 𝑥, 𝑦 ≥ 𝑡
𝐷 𝑥, 𝑦 = |
𝐼 𝑥, 𝑦 − 𝜇 𝑙𝑜𝑐𝑎𝑙(𝑥, 𝑦
𝜎𝑙𝑜𝑐𝑎𝑙(𝑥, 𝑦
|
2nd : Contrast normalization w.r.t Standard Reference Image
𝐼 𝑐𝑜𝑛𝑡𝑟𝑎𝑠𝑡
𝑐
= 𝐼 𝑈𝑛𝑖𝑓𝑜𝑟𝑚
𝑐
×
ℎ𝑖𝑠𝑡𝑝𝑒𝑎𝑘 𝐼 𝑈𝑛𝑖𝑓𝑜𝑟𝑚
𝑐
ℎ𝑖𝑠𝑡𝑝𝑒𝑎𝑘 𝐼 𝑅𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒
𝑐
𝑐 = 𝐿, 𝑈 , 𝑉
Application on Deep Learning ( Diabetic Retinopathy Severity)
Database : 35126 retinal images. ( Train 90% , Test 10%)
Overall classification accuracy 75.6% with normalization and 57% with original images
Luminosity and Contrast Normalization in Color Retinal Images
based on Standard Reference Image