Evaluation of Normalization Methods in a Cerebral Artery Atlas for Automatic Labeling

An existence probability atlas has been used for automatic labeling of cerebral arteries. However, the number of arteries varies frequently because of image quality and individual variation of the artery structure. To moderate the influence of number imbalance on labeling accuracy, we propose a new normalized atlas for automatic labeling of cerebral artery centerlines. The number of arteries, which was obtained from magnetic resonance angiography, varies from 11 to 46 among the artery sites. Based on the centerline and diameter, the arterial volume was reconstructed into a voxel space for each subject. After superimposing arteries from 46 subjects, three normalization methods were compared: dividing by the number of subjects (N), by N and the arterial length (L), and by N and the arterial volume (V). To compare the labeling accuracy and precision, the summation of probability and labeling method was also used. The accuracy of all normalization methods was > 85% in all arteries. The precision improved in some parts, with the atlas normalized by N-L and by N-V. The use of N-L and N-V changed the relative value of the existence probability among the parts. Consequently, some normalization methods changed the tendency toward misclassification, which changed the precision.