Useful Papers - Segmentation
Fischl et al. (2002), "Whole brain segmentation: Automated labeling of neuroanatomical structures in the human brain," Neuron 33, 341-355 PDF
Summary: Summary of some of the problems facing intensity-based segmentation and presentation of an algorithm that segments whole brains into various anatomical structures in completely automated fashion. The algorithm uses not just intensity information but a good deal of atlas information about prior probability of anatomical structure location. Fischl et. al show it has very good correlation with labeling by hand, both in normals and in patients with mild Alzheimer's. The software is commercial, and we may or may not be able to lay our hands on it here...
Bottom line: Intensity-based segmentation is hard, and anatomical labeling is harder still. If you can compile a good atlas of tissue types and anatomical locations, though, you can use prior probability information to do a pretty good automated job of segmentation.
Ashburner & Friston (1997), "Multimodal image coregistration and partitioning - a unified framework," NeuroImage 6, 209-217 PDF
Summary: The original paper defining the old SPM (pre-SPM2) way of doing coregistration. The authors suggest defining within-modality templates that are already coregistered and using least-squares methods to coregister the experimental images to those templates. Using segmentation during the coregistration can help improve the success and accuracy of that registration.
Bottom line: A bit obsolete these days; SPM2 has moved to MI coregistration, which is simpler and shows better success rates.
Marroquin et al. (2002), "An accurate and efficient bayesian method for automatic segmentation of brain MRI," IEEE Transactions on Medical Imaging 21(8), 934-945 PDF
Summary: Technical paper demonstrating another Bayesian-based probability atlas for doing automatic brain segmentation (Fischl et. al above use another). Some of the important problems in segmentation are surveyed in mathematical detail - things like partial voluming, etc. Different aspects of the algorithm - the values of the priors, the assumed spatial coherence, the noise, etc. - are varied to test their effects on the algorithm.
Bottom line: Bayesian parameter estimation can do a good job handling segmentation - with the right atlas. Far and away the biggest effect on the success of these atlases is what the value of the priors are, so a good atlas is essential...
Ashburner & Friston (2000), "Voxel-based morphometry — the methods," NeuroImage 11, 805-821 PDF
Summary: Didn't get to these next two, but this paper summarizes the whole analysis path in doing detailed analyses of structural images, say for a between-group study of anatomy. A number of the big problems in the path - inhomogeneity, segmentation, etc. - are discussed.
Yushkevich et al. (2006), "User-guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability," NeuroImage 31, 1116-1128 DOI
Summary: Presents ITK-SNAP (http://www.itk-snap.org), an open-source free automated segmentation algorithm, and demonstrates that SNAP segmentations of the caudate agree highly with manual segmentations by trained raters, and are even more reproducible (in overlap terms) than manual segmentations. SNAP segmentation also went significantly faster and required a hell of a lot less training.
Bottom line: ... It's almost... like a real piece of software... written by software engineers or something? <gasp> Honestly, this seems like really a pretty good tool for making individualized anatomical ROIs quickly and reproducibly.