09 NovSfN2013 Poster

Mining reliable landmarks for automated understanding of brain structure in MRI

Andrew Worth¹, Jason Tourville²

¹ Neuromorphometrics, Inc., Somerville, MA
² Dept. of Speech, Language, and Hearing Sciences, Boston University, Boston, MA

Abstract (download poster, 17MB)

Automated brain labeling has made great strides lately, but the human brain as seen in MRI still has complicated and detailed anatomy that is being overlooked because of anatomical variation. In this work we mine a database of a large number of manually labeled brain scans to find reliable landmarks that will help bring both the next level of automatic labeling and better understanding of the variation of human brain structure.

This approach can be used for the entire brain, but we focus on gyral and sulcal landmarks for labeling the highly variable cerebral cortex. We know from Ono (Thieme Medical, New York, 1990) that specific sulci have typical configurations. Moreover, we can see this variation in the individual brain scans that we have labeled. These variations are a problem for automated methods. What is needed is a characterization of typical sulcal configurations.

We previously generated a probabilistic atlas of 100+ neuroanatomical regions using 50 individually labeled scans. In a multi-year effort, boundaries were created manually to enclose neuroanatomical areas to comprehensively cover the entire brain. The current anatomical protocol defines structures such as the ventricles, thalamus, cerebellum, etc. The cortex is sub-divided into 49 “parcellation units” (PUs). Since a sulcus is a depression or fissure between the folds of cortical grey matter (gyri) and we already have labeled the PUs, we can algorithmically define a sulcus as the boundary between adjacent pairs of particular PUs. Sulci are the “negative space” of gyri.

In this work, we identify a collection of pairs of PUs using the BrainColor protocol to locate named sulci. For example, the cingulate sulcus “cgs” is defined by 8 pairs of PUs. The pair that defines the anterior region of the cgs is the anterior cingulate gyrus “ACgG” and the superior medial segment frontal gyrus “MSFG”.

To begin to understand typical sulcal configurations, we use this method to generate individual sulci and then their probabilistic average to discover landmarks and topological relationships. Each sulci is warped into a standard whole-brain coordinate system, and then each sulcus is compared with all others using an overlap metric. Groups of similar sulci define categories of sulcal configurations. Within these configurations, landmarks and their connections are identified to describe these categories. The most useful landmarks are those that reliably correlate with surface curvature because that information can be obtained from the scan and then used to locate the landmarks, which then identify the sulcal configuration.