Matching blood vessel patterns with the generalised EM algorithm

Esther de Ves, Richard C. Wilson, Edwin R. Hancock

Abstract: Describes an algorithm for registering retinal angiograms. The idea is to exploit the network structure of the blood vessel patterns to improve the recovery of the affine registration parameters. The framework for our study is provided by the generalised EM algorithm. There are two novel features of the algorithm. Firstly, in the expectation step of the algorithm we update the probabilities of correspondence matching using contextual constraints provided by the network structure of the blood-vessel pattern in the angiograms. The second novel idea is to recover affine transformation parameters by applying the Gustafson fuzzy clustering procedure to the expected log-likelihood function.