Effects of Patient Population Specific Atlases on Automatic Segmentation of Subcortical Structures in Freesurfer

Istvan Csapo¹, Jared Price¹, Troy Russell¹, Jeff Dewey¹, Ekta Sem¹, Daniel McCaffrey¹, Charles R. Guttmann¹, Bradford Navia², David F. Tate¹

We set out to investigate whether using manually labeled training data from a specific patient cohort to build probabilistic atlases for Freesurfer improves the automatic segmentation of novel images from the same cohort compared to the segmentations using the default atlas. 106 brain images were segmented and manually corrected. 64 of these images were used to build 6 different atlases. The other 42 images were segmented with Freesurfer using the new atlases. The results were compared to the manual segmentations. Out of the 8 subcortical structures investigated, the hippocampus segmentations were significantly improved with the new population specific atlases.