Bayesian Transductive Markov Random Fields for Interactive Segmentation in Retinal Disorders
In the realm of computer aided diagnosis (CAD) interactive segmentation schemes have been well received by physicians, where the combination of human and machine intelligence can provide improved segmentation efficacy with minimal expert intervention [1-3]. Transductive learning (TL) or semi-supervised learning (SSL) is a suitable framework for learning-based interactive segmentation given the scarce label problem. In this paper we present extended work on Bayesian transduction and regularized conditional mixtures for interactive segmentation . We present a Markov random field model integrating a semi-parametric conditional mixture model within a Bayesian transductive learning and inference setting. The model allows efficient learning and inference in a semi-supervised setting given only minimal approximate label information. Preliminary experimental results on multimodal images of retinal disorders such as drusen, geographic atrophy (GA), and choroidal neovascularisation (CNV) with exudates and subretinal fibrosis show promising segmentation performance.
KeywordsInteractive Segmentation Naive Bayes Mixture Models Generative Learning Markov Random Fields Bayesian Transduction
Unable to display preview. Download preview PDF.