Abstract
Purpose :
Retinopathy of prematurity (ROP) is the leading cause of avoidable blindness in children. Plus disease is defined by increased venous dilation and arteriolar tortuosity of the posterior retinal vessels, and is a sign that ROP is active and can progress. We present an automatic method combining image enhancement, vessel segmentation and tortuosity measurement to detect the presence of plus disease in retinal images.
Methods :
Image Enhancement: To reduce the noise, we first apply a median filter, then we apply a shade correction method to remove non-uniform illumination, followed by anisotropic diffusion. Finally, we apply contrast limit adaptive histogram equalization (CLAHE) using uniform distribution correction.
Vessel Segmentation: We use a multiscale enhancement technique and then perform CLAHE using the exponential distribution. The threshold is found by calculating the vessel density at a fixed threshold value (0.5), and adjusting based on that density. Finally, the vasculature is skeletonized and crossing/branching points are removed.
Tortuosity: Four measurements are calculated for 30-pixel vessel segments: 1) arc/chord ratio, 2) total curvature using numerical differentiation, 3) combination of method 1, magnitude of curvature and curvature sign changes, and 4) method 1 without dividing the vessel in segments. Tortuosity is measured for each image in an imaging sequence. The top 3 tortuous segments found in each case are then averaged.
Results :
Our dataset consists of 86 imaging sessions of preterm babies <31 weeks gestational age and <1kg birth weight. Clinical ground truth was the dilated examination by a retina specialist that showed 25 sessions had plus disease and 61 did not. We tested each tortuosity method for the classification of plus disease. Best results were achieved by method 2 with an AUC of .86, maximum accuracy of 72% and sensitivity/specificity of 0.92/0.61. Example in Fig. 1.
Agreement was calculated (Cohen’s kappa) between: clinical ground truth and the gradings from a reader (.20), and clinical ground truth and the algorithm output (.57).
Conclusions :
We demonstrated that our automatic approach can detect cases with plus disease with high sensitivity and better agreement with the clinical ground truth than the reader, making the system ideal as a screening tool.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.