Improved Skull Stripping using Graph Cuts

Introduction. Removal of non-brain tissues, particularly dura, is an important step in enabling accurate measurements of brain structures. Previous approaches (2,5,6) can lead to brain loss and/or leave residual dura (4), posing a problem for subsequent cortical thickness estimation. Brain loss cannot be reversed downstream in the segmentation pipeline. Inclusion of dura mater can cause overestimation of cortical thickness. We propose a novel, fully-automatic skull stripping algorithm that is motivated by the earlier work involving intensity thresholding and removal of narrow connections using morphological processing (Error! Reference source not found.). Instead of morphological operations, we employed a superior graph theoretic segmentation with suitably modified edge weight assignment to enable precise removal of narrow connections and dura attachments. We wanted to test whether this approach would lead to cleaner (with fewer dura attachments) brain mask, which would be more suitable for subsequent cortical thickness estimation.