Comparison and Validation

In this section, we compared several methods based on their volumetric segmentation accuracy when applied to singlechannel as well as two-channel MR data. For that purpose, we employed digital MR phantoms of neuroanatomy [12] that simulated the appearance and image characteristics of the Tl-weighted images and double-echo images in elderly populations, which generally have lower contrast-to-noise ratio than younger subjects. Because tissue volumes are typically measured across a wide range of ages, this phantom provides a conservative measure of segmentation performance. There are many advantages for using digital phantoms rather than real image data for comparing segmentation methods. These advantages include prior knowledge of the true compartment volumes (i.e., WM, GM, and CSF volumes) and control over image parameters such as mean intensity values, noise, slice thickness, partial volume effects, and magnetic field intensity inhomogeneities.

Tables l and 2 summarize our findings. In Table l, several techniques were compared using a 3D high-resolution Tl-weighted phantom (in-plane resolution = 0^94 mm2, slice thickness = 1^5 mm, Gaussian noise with a = 6^0, 3D linear shading 7% in each direction). The results revealed that the adaptive Bayesian method presented the best overall accuracy for segmentation of WM, GM, and CSF volumes. The adaptive

TABLE 1 Tl-weighted phantom segmentation errors"

Absolute volumetric error (%)'

Absolute volumetric error (%)'

TABLE 1 Tl-weighted phantom segmentation errors"

Technique

Was this article helpful?

0 0

Post a comment