Research On Accuracy-Oriented Medical Image Registration Algorithms

Medical image registration is an important research topic of medical image processing. Multiple medical images are registered and fused to provide more information for medical analysis and treatment. Although registration is highly focused, still there are deficiencies such as accuracy in existing registration algotithms.In rigid registration algorithms based on normalized mutual information, the similarity metric function gets trapped into local minimums easily during searching process, so an improved particle swarm optimization algorithm is proposed, non-linear decreasing weight is incorporated into the algorithm to enhance searching ability, besides, aggregation analysis and mutation are also adopted to avoid local minimums.Consistent registration error exists when independently estimated forward and reverse transformations are composited in intensity-based nonrigid registration algorithms, so a new consistent registration cost function with local correlation coefficients,consistency constraint and regularization constraint is proposed here. By minimizing the new cost function, forward and reverse transformations are jointly estimated and guaranteed to be inverse of each other, and in this way, the consistent registration error is highly decreased.It is hard to determine corresponding landmarks and match them exactly in landmark-based nonrigid registration algorithms, so a novel method which combines normalized mutual information and gradient information is proposed here to match corresponding landmarks precisely, besides, cubic B-spline basis functions are used to get the smooth transformation.