Purpose:
To examine the accuracy of fractal analysis for plus diseasediagnosis in retinopathy of prematurity (ROP).
Methods:
34 retinal images were diagnosed as "plus" or "not plus" byconsensus of 22 experts. These images were segmented by a computer-basedimage analysis system (see sample RetCam and segmented images).Fractal dimension, Df, was computed with the box-counting methodusing Matlab (Natick, MA). Images were separated into retinalartery and vein subsets for analysis, and fractal dimensionswere computed for all subsets. The Mann-Whitney test was usedto compare "plus" vs. "not plus" groups. Receiver operatingcharacteristic areas under the curve (AUC) for plus diseasediagnosis by artery and vein fractal dimensions were calculated,and compared to a reference standard of expert diagnosis.
Results:
Among the 34 images, 21 were identified as "not plus" and 13as "plus" by expert consensus. Mean (SD) retinal artery Df was1.1874 (0.0974) in the "not plus" group and 1.2404 (0.08682)in the "plus" group (no statistically significant differencebetween the groups [p=0.151]). Mean (SD) retinal vein Df was1.2691 (0.04677) in the "not plus" group and 1.3011 (0.04626)in the "plus group" (statistically significant difference betweenthe groups [p=0.045]). AUC (SD; p-value) for retinal arteryDf was 0.650 (0.099; 0.151) and for retinal vein Df was 0.707(0.096; 0.045).
Conclusions:
Fractal dimension of retinal veins from ROP retinal images maydiagnose plus disease with high accuracy.
Keywords: retinopathy of prematurity • image processing • imaging/image analysis: clinical