In my phd work I was heavily involved in hippocampal segmentation, and I can say with confidence that FSL FIRST is not a state-of-the-art segmentation method. It belongs to an earlier generation of segmentation methods with poorer reliability, which have contributed to a lot of contradictory results in my field of hippocampal development.
I would not use it in my research.
Ditto. We developed a nice set of high-res hippocampal atlases[1] and multi-atlas segmentation method to use them[2] and compared it with FSL, and freesurfer and found FSL and freesurfer would often over-estimate segmentations. Good for a first pass because they are both dead-easy to run though.
I have some experience with FreeSurfer. As a whole, FreeSurfer is pretty terrible at hippocampal/subcortical segmentation, but very good at cortical surface reconstruction. One of the links I posted in grandparent presents a validation experiment on Hongzhi Wang's method employing Adaboost learning to vastly improve FreeSurfer segmentations. It is impressive and useful for researchers used to FreeSurfer.
However, the state of the art for image segmentation is dominated by a class of methods called multi-atlas segmentation, which I am using for a large longitudinal study I am currently writing up.
A large problem of earlier versions of FreeSurfer v. < 6 was the resampling of all data to 1mm isotropic voxels.
For hippocampal segmentation, many regions contain intersections between three tissue types (white matter, gray matter, and CSF). This complicates tissue classification of images with large voxels, because a voxel will be an average of multiple tissue types. FreeSurfer version 6 has options to use the native resolution of your structural images, so this might improve things some. I have yet to try it out. However, it will still use a single probability atlas, which tends to have lower performance than the multi-atlas methods.
I should emphasize that FreeSurfer is great for what it does great: Surface reconstruction.
Thanks for all the details. Are those atlas based methods applicable to standard T1 scans or require extras like T2 or flair? Are there available software implementations or it's all research code?
the method is quite general and applies to any image type for which an appropriate atlas is available.
It is simple to see why with a straightforward description of Multiatlas methods. Generally two steps are performed: Warping and Voing.
Step 1. Warp. An atlas brain is spatially warped to match the subject's brain using affine, non-linear and diffeomorphic normalization tools (ANTs being one of the strongest contending registration spatial normalization programs). This warp maps the atlas's ROI onto the subject's brain, producing a candidate segmentation. This is done for each atlas, producing N candidate segmentations.
Step 2. Vote. In its simplest form, candidate segmentations vote at each voxel for an ROI label, majority wins.
This produces the finished segmentation.
Advances in image registration and voting methods over the last decade have really made this the preferred method for accurate image registration. Unfortunately, sometimes good multi atlases are lacking. Historically, most published atlases are single atlases with an average or probability estimate at each voxel. There are some mutliatlases available for specific structures (e.g. hippocampus, see MAGeT comments above). I believe there are some multi ROI multi atlases now available. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4691373/
https://www.researchgate.net/publication/273940329_Assessing...
https://www.researchgate.net/publication/260915135_Volume_of...
https://www.researchgate.net/publication/274252280_Developme...
[edit] I had meant to point a link to my chapter on hippocampal development. https://www.researchgate.net/publication/314194708_Hippocamp...