Automated cerebellar segmentation: Validation and application to detect smaller volumes in children prenatally exposed to alcohol.

Objective: To validate an automated cerebellar segmentation method based on active shape and appearance modeling and then segment the cerebellum on images acquired from adolescents with histories of prenatal alcohol exposure (PAE) and non-exposed controls (NC).

Methods: Automated segmentations of the total cerebellum, right and left cerebellar hemispheres, and three vermal lobes (anterior, lobules I-V; superior posterior, lobules VI-VII; inferior posterior, lobules VIII-X) were compared to expert manual labelings on 20 subjects, studied twice, that were not used for model training. The method was also used to segment the cerebellum on 11 PAE and 9 NC adolescents.

Results: The test-retest intraclass correlation coefficients (ICCs) of the automated method were greater than 0.94 for all cerebellar volume and mid-sagittal vermal area measures, comparable or better than the test-retest ICCs for manual measurement (all ICCs > 0.92). The ICCs computed on all four cerebellar measurements (manual and automated measures on the repeat scans) to compare comparability were above 0.97 for non-vermis parcels, and above 0.89 for vermis parcels. When applied to patients, the automated method detected smaller cerebellar volumes and mid-sagittal areas in the PAE group compared to controls (p < 0.05 for all regions except the superior posterior lobe, consistent with prior studies).

Discussion: These results demonstrate excellent reliability and validity of automated cerebellar volume and mid-sagittal area measurements, compared to manual measurements. These data also illustrate that this new technology for automatically delineating the cerebellum leads to conclusions regarding the effects of prenatal alcohol exposure on the cerebellum consistent with prior studies that used labor intensive manual delineation, even with a very small sample.