Combining multiple expert annotations using semi-supervised learning and graph cuts for Crohn’s disease segmentation

Abstract

We propose a graph cut (GC) based approach for combining annotations from multiple experts and segmenting Crohns disease (CD) tissues in magnetic resonance (MR) images. Random forest (RF) based semi supervised learning (SSL) predicts missing expert labels while a novel self consistency (SC) score quantifies the reliability of each expert label and also serves as the penalty cost in a second order Markov random field (MRF) cost function. The final consensus label is obtained by GC optimization. Experimental results on synthetic images and real CD patient data show our final segmentation to be more accurate than those obtained by competing methods. It also highlights the effectiveness of SC score in quantifying expert reliability and accuracy of SSL in predicting missing labels.