Detection of Retinopathy of Prematurity using multiple instance learning

This paper proposes a new method for detecting Retinopathy of Prematurity (ROP) using multiple instance learning (MIL) approach from retinal images captured by RetCam, a digital retinal camera. In this work, a set of features having significant relevance to capture ROP characteristics, are extracted and miGraph MIL method is used as the classifier to learn from the extracted features. The diagnostic image is split into a grid of patches, and instances are constructed from each grid element by extracting a set of features from it. All the feature sets or group of instances belonging to the same image are grouped into a bag. Labels are assigned for instances and for the bags as a whole. Finally, the bags along with their labels are fed into a MIL classifier for classification. A good performance of miGraph on the ROP retinal images is observed and the initial experimental results are promising. In our literature survey, we observed that current research on detection of ROP using MIL has not been reported till now. Our results indicate that MIL offers an easy, yet effective, paradigm for ROP screening.