Unsupervised Segmentation of Choroidal Neovascularization for Optical Coherence Tomography Angiography by Grid Tissue-Like Membrane Systems

Abstract

Accurate segmentation of choroidal neovascularization (CNV) patterns is vital for precise lesion size quantification in age-related macular degeneration. In this paper, we develop a method for unsupervised and parallel segmentation of CNV in optical coherence tomography based on a grid tissue-like membrane (GTM) system. A GTM system incorporates a modified Clustering In QUEst (CLIQUE) algorithm into tissue-like membrane systems. Exploiting CLIQUE's aptitude for unsupervised clustering, GTM systems can detect CNV of different shapes, positions and density without the need of a training stage. The average dice ratio is 0.84±0.04, outperforms both baseline and the state-of-the-art methods. Besides, being a parallel computational paradigm, GTM systems can handle all scans under analysis simultaneously and therefore they are less time consuming, completing CNV detection on 48 scans in 0.56 seconds.