Abstract
Purpose: :
To compare diabetic retinopathy (DR) severity levels graded from compressed stereo color digital retinal images to those graded from uncompressed digital images and film photos.
Methods: :
In diabetic persons, 154 eyes were imaged with a Topcon camera (6 megapixel sensor, 2392x2048 pixels) in 7 standard-field color stereo 35o [7SF] per ETDRS protocol. Digital images (D) were stored both uncompressed (TIFF) and compressed (JPEG2000), the latter at 37:1 - just below the threshold of Just Noticeable Difference (JND), the theoretical smallest difference between two similar photographs detectable by a human, using Color JND-Metrix (Sarnoff Corp., Princeton, NJ). For reference, film images (F) were taken with the same camera. Three graders independently determined ETDRS DR level from digital uncompressed (DU) and compressed (DC) images using calibrated 20" LCD monitors with hand-held stereo viewers in Topcon IMAGEnet. Image tonal parameters were enhanced according to the AREDS2 image model, and images also examined in green-channel breakout. DR was graded from F (slides on light boxes with Donaldson viewers). Order and timing of evaluations were controlled by custom software to minimize bias and recall. DR levels within each medium were defined by central tendency among graders.
Results: :
From DU images, graders classified DR as follows: no DR = 28 eyes, non-proliferative DR (NPDR) = 94 eyes (microaneurysms only/mild/moderate/severe = 7/31/21/35), proliferative DR (PDR) = 31 eyes (mild/moderate/severe = 9/11/11), and ungradable DR = 1 eye. Comparing DC vs. DU for specific ETDRS level (9-step scale) yielded 79.2% exact agreement and 98.7% +1 step (unweighted Κ = 0.76, SEΚ = 0.04; linear weighted Κ = 0.92, SEΚ = 0.02). Agreement with F was similar between DC and DU: exact, 64.3% vs. 66.9%; +1 step, 94.8% vs. 94.8%; unweighted Κ, 0.59 vs 0.61 (SEΚ ≤0.05); weighted Κ, 0.82 vs. 0.83 (SEΚ = 0.03). No systematic difference in sensitivity for detecting DR levels was evident between DU and DC, nor between either variety of D and F.
Conclusions: :
High-quality (albeit lossy) compression of digital images did not have a significant impact on DR severity evaluation compared to uncompressed digital and film images. JPEG2000's wavelet compression achieved higher compression at equivalent JND values, producing smaller files, than did standard JPEG's cosine transform. Smaller image files facilitate data storage/transmission.
Keywords: clinical (human) or epidemiologic studies: outcomes/complications • diabetic retinopathy • imaging methods (CT, FA, ICG, MRI, OCT, RTA, SLO, ultrasound)