Obtaining Consensus Annotations For Retinal Image Segmentation Using Random Forest And Graph Cuts

We combine random forest (RF) classifiers and graph cuts (GC) to generate a consensus segmentation of multiple experts. Supervised RFs quantify the consistency of an annotator through a normalized consistency score, while semi supervised RFs predict missing expert annotations. The normalized score is used as the penalty cost in a second order Markov random field (MRF) cost function and the final consensus label is obtained by GC optimization. Experimental results on real patient retinal image datasets show the consensus segmentation by our method is more accurate than those obtained by competing methods.