Our Blog
15 NovSfN 2017 Poster
Using a probabilistic atlas to improve manual parcellation of the cerebellum
Andrew Worth1, Jason Tourville2
1 Neuromorphometrics, Inc., Somerville, MA
2 Dept. of Speech, Language, and Hearing Sciences, Boston University, Boston, MA
Abstract (download poster, 6.1MB)
In order to develop automated methods for labeling anatomy in MRI brain scans, regions of interest (ROIs) must first be labeled manually and this must be done in a large representative sample in order to capture variability. This is a tedious task that requires considerable expertise but the overall effort can be reduced and the accuracy and reliability can be increased by using a probabilistic atlas to improve both the individual ROIs and also the labeling method.
The cerebellum was labeled in 56 T1-weighted MRI brain scans. There were 31 unique subjects, 25 of which were scanned twice and labeled blind as to which were repeat scans. We also labeled the famous high-resolution Colin27 scan. Following the Schmahmamn et al. atlas, 33 regions were delineated in coronal slices: brain stem, right and left cerebellum white matter, and right, left and vermis lobules: I-IV, V, VI, CrusI (vermis “VIIAf”), CrusII, (vermis “VIIAt”), VIIb, VIIIa, VIIIb, IX, X. Identification of the relevant cerebellar sulci was done with the aid of 3D renderings of the cerebellum surface at a series of depths between the white matter and exterior surface. The iso-intensity surfaces were defined by various intensities on either side of the gray matter peak intensity found in a histogram over the cerebellum region. “Sulci lines” drawn on these surface were projected into the 3D scan volume space to identify specific sulci. “Parcellation lines” were then drawn based on the sulci lines in coronal slicess to separate the cerebellar cortex into lobes.
All labeled scans were used to create a probabilistic atlas by averaging the presence or absence of each label after a simple 3D warp to normalize the size of each cerebellum. Then, each region in each slice of every scan was compared to this atlas and an “atlas mismatch” score was calculated. In order to flush out labeling errors, large mismatch values were flagged and those outlines were examined and modified as necessary.
In our results, we describe the locations and types of errors that were found using the probabilistic atlas, and we suggest modifications to the labeling method to improve the reliability and decrease labeling time. In our initial work, we noted the lack of a reliable landmark for dividing vermis VIIAt and vermis VIIAf. According to Schmahmann et al., these regions are separated by the horizontal sulcus. This sulcus is one of the more robust landmarks on the lateral surface but rarely extends medially across the vermal surface. As such, the boundary between these two regions is often arbitrary. Reliable labeling of these areas requires either combining these two small vermis ROIs or identifying another, more robust landmark.
Reference
Schmahmann, Jeremy D., Julien Doyon, Michael Petrides, Alan C. Evans, and Arthur W. Toga. MRI atlas of the human cerebellum. Academic Press, 2000.