To FreeSurfer ListServ: Dr. Iglesius recently responded to the questions below and asked that the question and response (also included below) be posted on the Freesurfer ListServ…Thanks very much.
Questions on 13 Feb 2017:
In the manuscript Iglesius et al, “A computational atlas of the hippocampal formation…”, Neuroimage. 2015 July 15; 115: 117–137, there is a Table 7, repasted below. I wondered if you could help me to make sure that I am correctly interpreting what each one of the permutations involved. The first line (aseg, 5.3) was clear. I believe that the second line refers to the summation of the subregion volumes generated using version 5.3. I was unsure as to how “in vivo 5.3” of the third line differs from the first two lines? In the final line, was “this study and ex vivo” a different data set or were all using the exact same data? The impression that I received was that there is benefit in calculating hippocampal volume by adding up the subfields, but that the subfield accuracy is better in 6.0 which further increases the value of subfield addition for calculating overall hippocampal volume.
Accuracy and area under the curve for the AD discrimination task. AtlasAccuracy at elbowAUROC Whole hippocampus (“aseg”, FreeSurfer v5.3)82.1%0.887 Whole hippocampus (adding up the volumes of the subregions)84.0%0.901 In vivo (FreeSurfer v5.3)86.3%0.917 Ex vivo (this study and FreeSurfer v6.0)88.0%0.931
Freesurfer may be used to generate hippocampal volumes for an enrichment study of MCI patients, to identify which subjects are most likely to worsen clinically at greater rates. I am trying to understand whether the performance of Freesurfer in this regard will be affected by whether the version used is 5.3 vs. 6.0.
One further question is how the high resolution atlas used in 6.0 compares to multi atlas techniques. I believe that your paper did make a comparison to at least one multi-atlas approach, and performance was superior for certain subfields. However, others have found that multi-atlas approaches may perform better than FS (versions prior to 6.0). If there is a point to be made about FS 6.0 and its comparison to multi-atlas techniques, this could be helpful as well.
Any insights on these questions would be appreciated.
Responses from E Iglesius:
Thank you for your email. I really appreciate that people seem to be finding this atlas useful.
I agree that the table is not as clear as it should have:
Whole hippocampus (“aseg”, FreeSurfer v5.3): this is just the estimate from the main FS recon-all script (“aseg”), stored in aseg.stats Whole hippocampus (adding up the volumes of the subregions): adding up the subfield volumes. It works a bit better than the aseg version (there’s actually 2-3 cases in which aseg fails badly but the hippocampal module manages to recover). In vivo (FreeSurfer v5.3): multivariate classification using the volumes of the subfields, as estimated by the in-vivo atlas in FS5.3 (Van Leemput 2009) Ex vivo (this study and FreeSurfer v6.0): multivariate classification with the ex vivo atlas.
Regarding your specific question: people have written papers about how terrible the 5.3 subfield segmentation was. So, if you’re doing subfields, you should use 6.0. If you’re doing whole hippocampal volumes by adding subfields, it doesn’t matter much if you’re using 5.3 or 6.0, but I’d still use 6.0 because with 5.3 you’re looking for trouble for the reason mentioned above. And, of course, if you’re using the whole hippocampal volume, you can always use the aseg volume.
My two cents regarding Freesurfer 6.0 vs. multi-atlas techniques:
- If you have training data that is matched (in terms of MRI acquisition and contrast) to you test data, go for multi-atlas. It’s slow, but will work well.
- If you do not, then generative models like ours do better- because they are independent of the modality of the test scan. This is important when you publicly release your code, e.g., as part of FreeSurfer (you need algorithms to be robust).
- In our case: if the training data is ex vivo, there’s no choice. You have to do generative - there’s no way to register the intensities with good accuracy.
Cheers,
/Eugenio