Combining multiple expert annotations using semi-supervised learning and graph cuts for medical image segmentation

Generating consensus ground truth segmentation from multiple experts is important in medical imaging applications such as segmentation. We propose a novel approach to combine multiple expert annotations using graph cuts (GC) and semi supervised learning (SSL). Current methods use iterative Expectation-Maximization (EM) based approaches to estimate the final annotation and quantify annotator's performance. This poses the risk of getting trapped in local minimum and providing inaccurate estimates of annotator performance. A novel self consistency (SC) score quantifies annotator performance based on the consistency of their annotations in terms of low level image features. The missing annotations are predicted using SSL techniques that consider global features and local image consistency. The self consistency score also serves as the penalty cost in a second order Markov random field (MRF) cost function which is optimized using graph cuts to obtain the final consensus label. Graph cut optimization gives a global maximum and is non-iterative, thus speeding up the process. Experimental results on synthetic images, real data of Crohn's disease patients and retinal images show our final segmentation to be accurate and more consistent than those obtained by competing methods. It also highlights the effectiveness of self consistency in quantifying expert reliability and accuracy of SSL in predicting missing labels.