Abstract
This paper describes and evaluates FMRIB’s nonlinear image registration tool (FNIRT), that is part of the FMRIB software library (FSL). It is a small deformation framework using sum of squared differences (SSD) as its cost function and Gauss-Newton for minimisation. The framework uses a joint shape and intensity model that attempts to explain the observed differences between two images in terms of having different shape and/or contrast, being differently affected by intensity bias-fields etc. Thus the estimation of the warps will be relatively unaffected by intensity differences that would otherwise violate the assumptions behind the SSD cost function. It uses a projection onto a manifold defined by a specified range of allowed Jacobian determinants to ensure that the warps are diffeomorphic. The utility of the model is demonstrated on a variety of simulated and experimental data with good results. FNIRT is also quantitatively evaluated using previously published datasets consisting of scans from multiple subjects, all with anatomically defined brain regions that are manually outlined. In this evaluation FNIRT performs well in comparison to previously published results with other registration algorithms.