Toward a severity index for ROP: An unsupervised approach

Retinopathy of prematurity (ROP) is a disease affecting low birth-weight infants and is the major cause of childhood blindness. Although accurate diagnosis is important, there is a high variability among expert decisions mostly due to subjective thresholds. Existing work focused on automated diagnosis of ROP. In this study, we construct a continuous severity index as an alternative to discrete classification. We follow an unsupervised approach by performing nonlinear dimensionality reduction. Instead of extracting several statistics of image features, each image is represented by the probability distribution of its features. The distance between distributions are then used in manifold learning methods as the distance between samples. The experiments are carried out on a dataset of 104 wide-angle retinal images. The results are promising and they reflect the challenges of the discrete classification.