Bayesian transduction and Markov conditional mixtures for spatiotemporal interactive segmentation

In this paper we propose a novel transductive learning machine for spatiotemporal classification casted as an interactive segmentation problem. We present Markov conditional mixtures of naive Bayes models with spatiotemporal regularization constraints in a transductive learning and inference framework. The proposed model extends on previous work to account for non independent and identically distributed (i.i.d.) sequential data by imposing the learning and inference problem w.r.t. time. The multimodal mixture assumption on the class-conditional likelihood for each covariate feature domain in conjunction with spatiotemporal regularization constraints allow us to explain more complex distributions required for classification in multimodal longitudinal brain imagery. We evaluate the proposed algorithm on multimodal temporal MRI brain images using ROC statistics and report preliminary results.