Improved method for automatic cerebrovascular labelling using stochastic tunnelling

The complexity and high morphological variation of cerebral vasculature make comparison and analysis of the vessel patterning difficult and laborious. A framework for automatic labelling of the cerebral vessels in high resolution 3D images has been introduced in the literature that addresses this need. The segmented vasculature is represented as an attributed relational graph. Each vessel segment is an edge in the graph with local attributes such as diameter and length, as well as relational features representing the connectivity of the vessel segments. Each edge in the graph is automatically labelled with an anatomical name through a stochastic relaxation algorithm. In this paper, we compare the performance of four different optimization schemes, including stochastic tunnelling, for automatic labelling. We validated our method on 7 micro-CT images of C57Bl/6J mice with a leave-one-out test. The mean recognition rate of complete cerebrovasculature using stochastic tunnelling is 80% and shows a 2% (>60 vessel segments) improvement compared to simulated annealing optimization.