Michal Bernier1, Martin Lepage, Roger Lecomte, Luc Tremblay2, Louis Dor-Savard2, Maxime Descoteaux3
1Dpartement d'informatique et Centre d'imagerie molculaire de Sherbrooke, Universit de Sherbrooke, Sherbrooke, QC, Canada; 2Centre d'imagerie molculaire de Sherbrooke, Universit de Sherbrooke, Canada; 3Computer Science, Universit de Sherbrooke
Automatic and accurate image registration of PET-MRI images is not trivial, since the image intensity content of PET and MRI images is highly uncorrelated. Using normalized mutual information (NMI) and symmetric correlation ratio (SCR) as similarity measures in the registration algorithm, we propose 2 novel elements 1) we introduce a multi-resolution similarity measure combining the intensity histograms at 3 resolution levels simultaneously and 2) we add a PCA initialization to obtain a coarse initialization for a fast, fully automatic accurate registration. These novelties are highlighted for accurate automatic registration of 8 rats in a femur bone tumor study.