Macular segmentation using an automatic deep learning and graph cut strategy

Age-related Macular Degeneration (AMD) is one of the main diseases affectingvisual health in the world, since it causes severe and irreversible vision loss inelderly people. According to the International Agency for the Prevention of Blind-ness (IAPB), by the year 2040, 288 million people are expected to be affectedworldwide, while in Colombia for the year 2020, approximately 234,200 caseswere estimated. With the improvement of computer vision techniques, automaticdiagnosis has become a reliable tool for screening and determining the degree ofthe disease. Many strategies have been proposed in the last decade, from the firstbased on manual characterization of the disease to current ones based on modeltraining using large databases of images annotated by specialists, with very highcomputational costs and data collection time. The present article proposes a novelmethod based on combining the two trends, using the best of each one. The dis-ease is characterized by segmenting its two fundamental features in cascade: theocular vascular network is segmented with a pre-trained deep learning model andtransformed into a graph for extraction of terminal nodes (which are known todelimit the macular zone); and the macula is segmented through a semi-automatic algorithm called graph cut that requires initial background and foreground infor-mation (which corresponds to the terminal nodes already obtained) to become anautomatic method. Finally, we present a comparison between macular segmen-tation using the proposed algorithm with parametric changes in its architectureand a deep learning model for macular segmentation with a reduced set of aug-mented images. It is confirmed that, with a limited database, the proposed graphcut-deep learning strategy is a viable way to obtain a method to extract the mainfeatures of the disease and generate a subsequent diagnosis.