Figure 4.

Comparison of three non-linear diffeomorphic registration methods. First and second column: we see the intensity bias affecting the source and target images. (A) Registration of the head with no confidence mask: strong deformation fields are estimated in the skull and meninges that diffuse to the outer cortex region and bias the results (cf. red ellipse where a non-realistic expansion of 38% is found). (B) Registration of the skull-stripped images (no confidence mask): the use of the skull-stripped images biases the result at the level of the outer cortex (cf. red ellipses) where non-existing high value deformations are found due to the high intensity gradient. In fact, skull-stripping imposes the outside brain intensity to be zero creating a high intensity gradient that biases the registration results (the update δv_i-j is directly proportional to the image gradient). (C) Registration of the head with confidence mask: the registration using the confidence mask enables us to estimate realistic transformations in the outer cortex.