A comparison of nonlinear least-square optimization methods in affine registration of SPECT images

The complexity of brain structures and the differences between brains of different subjects or across image modalities make necessary the normalization of the images with respect to a common template. The general affine model, with 12 parameters, is usually chosen as the first normalization algorithm before to proceed with a more complex non-rigid spatial transformation model. Usually, a cost function which presents an extreme value when the template and the image are matched together is optimized. In this work, the objective function is chosen to be the mean squared difference between both images, and the performance of different optimization procedures are compared. Namely, the Levenberg-Marquardt and four Gauss-Newton algorithms.