Atlas Based Methods

A recent automatic approach for segmentation of sagittally acquired spoiled gradient echo pulse sequence head images has been reported by Hartmann et al. for the study of brain atrophy in alcoholic patients [16] (see Fig. 13). This method starts with a rigid registration to register an atlas volume with each patient volume and brings the volumes into global correspondence. Then it uses a nonrigid transformation based on Thirion's method [30] to deform the volumes into local correspondence. In this method, binary volumes of the brain and cerebellum in the atlas are generated with manual segmentation. The deformation field that warps the atlas to the patient is applied to these binary volumes to create intradural and cerebellum masks in that patient volume. Contours have to be drawn on the atlas once, and all subsequent segmentations are fully automatic. These automatic brain and cerebellum segmentations are seen to be accurate enough for measuring the brain volume. The threshold for the elimination of CSF must be carefully chosen and is done automatically in this case. This method has been shown to work well on several patients, using an atlas volume acquired in the same manner as those of the patients. Hartmann's technique shows great promise for future development and use with sagittal T1-weighted MR images, and possibly for conventional spin-echo pulse sequence images.

Another atlas-based method has been developed by Collins et al. for automatic segmentation of gross anatomical structures of the brain [10]. Their method uses a probabilistic atlas and is particularly effective for isolating the gyri. It has been validated on 20 MRI volumes.

FIGURE 13 Atlas-based automatic brain segmentation results.
0 0

Post a comment