This wiki pages describes the research effort for improving the subcortical segmentation accuracy for data acquired across different scanner platforms or imaging sequences.

The subcortical segmentation method within FreeSurfer is an atlas-driven method. It uses a probabilistic brain atlas that encodes the prior intensity model for each brain structure at each atlas location, in addition to the prior model for the spatial distribution of these brain structures. As a result, the accuracy of the method often degrades when processing data acquired using scanner plaforms and/or imaging sequences that differ from the data used for the atlas training, due to changes in underlying image contrast. To resolve this problem and reduce the sensitivity of the method to changes in scanner platforms or acquisition parameters, we have developed an intensity renormalization procedure that automatically adjusts the atlas prior intensity model, in order to better fit the input data.

The intensity renormalization is performed for individual structures. The basic idea is to try to get an estimation of the intensity distribution of each structure in the input image, and then the prior atlas intensity model can be properly normalized to match the given data.