1. Y Liu, et al. Vanderbilt University, TN
Journal Club
Y Liu, et al. Vanderbilt University, TN
“Thalamic Nuclei Segmentation in Clinical 3T T1-weighted Images Using High-Resolution 7T Shape Models”
Oct 12, 2015
Jason Su

2. Motivation Most recent publication for thalamic nuclei segmentation
Notably using high-resolution T1w images instead of DTI
Shape-based methods are also applicable to our dataset (since we have 3D VTK models)
Perhaps another way of creating a segmentation candidate for label fusion instead of nonlinear registration

3. Background
At ISMRM 2014, Newton, D’Hause, and Dawant presented “Visualizing Intrathalamic Structures with Combined Use of MPRAGE and SWI at 7T”
The “magic image” vs the “rainbow approach”
5 subjects, MPRAGE at 4 TIs and SWI, manual seg.
Traditionally, DTI and fMRI based methods
These are low resolution and lack a ground truth

4. Aims
Segment thalamic nuclei on conventional 3T T1w using models segmented from 7T
Acquiring multiple contrasts is costly and 7T is not widely available
Even WMnMPRAGE may not be easy to spread due to non-product RFB?
Use shape-based methods because low intra-thalamus contrast for registration-based methods

5. Methods: Acquisition 7 healthy subjects (5M:2F)
Philips 7T – for development of shape model
Ax T1w: 0.7mm-iso, no sequence details
Sag MPRAGE: 0.7mm-iso, TR=4.74, TIs=[400, 640, 960, 1120], TS=4500, fa? ETL?
WMn: 1mm-iso, TR=10, TI=680, TS=6000, fa=4, ETL=200
Obl. Ax/Cor SWI: 0.24x0.24x1mm, 60 slices, fa=45
Philips 3T – what the method will segment
T1w: 0.7mm-iso, TR=7.92

6. Methods: Acquisition 3T and 7T T1w seem to be fairly matched
Each image is registered to T1w (b.1)
Use all these contrasts to perform manual seg.
I found this display very disorienting, relies on left and right showing the same structures
Should show improved contrast of some nuclei at different TIs as well as how good their registration was, don’t think it accomplishes either

7. Nuclei 19 nuclei were segmented compared to our 12 or 15
Some key differences, not sure if these combinations are right
AV = AVD + AM
MD-Pf = MD + Pf
Pul = PuMI + PuA
We’re missing CeM, CL, Li, LP
They’re missing LGN, MGN, RN Sth

8. Methods: Shape Model
Want to establish vertex correspondences between training subjects’ nuclei shapes
Pick a random subject as the reference atlas
This leaves 5 training examples and 1 test
Register whole thalamus from reference to each training example using image then surface registration
Apply these to next level of substructures, then use more nonlinear and shape-based reg.
Recursively do this for all nuclei
Want variance of a vertex point to only be due to intersubject variability so that shape model models this

9. Thin Plate Splines Spline-based interpolation/transformation method
Like bending a thin sheet of metal
Enforce smoothness on fitted surface (integral of sq. of 2nd deriv.)
Has a closed-form solution
Like having K control points that you can tweak around with spline interpolation
Local non-affine point correspondences and global affine parameters
But how do we decide which points will go to which on the truth?
i.e. the ordering of i?
In-house “3D snake” algorithm

10. Methods: Join hierarchical modeling
Only segmented each individual nuclei, need to process to get a hierarchy
For each level, remove inner boundaries of group to arrive at the higher-level shape

11. Methods: Variation estimation
Now that point correspondences have been determined, estimate intersubject variability of these points
Register to reference nuclei using 7dof affine to all training samples (why not 12dof or nonlinear?)
Average across training samples to get a mean shape/set of points for each nuclei
Compute a covariance matrix of these points from mean
Eigen-decomposition to determine the principal axes/modes of variation of points
3 axes explains 80% of shape variation
Don’t we really want the eigenvectors to be tuned for modeling intra nuclei surfaces?

12. Methods: Variation estimation

13. Methods: Segmentation
Using the 3T T1w images
Segment thalamus using FreeSurfer
Shape initialization
Image and shape-based registration to reference
Determine point correspondences of whole thalamus contour
Then take 7dof affine registration of reference surface only?
Feels like throwing away all the work we did for non-linear reg.
At each nuclei level, fit for b using wtd. least squares
Find the combination of eigenvectors that best fits the surface that we know
Ignore errors on inner points since they’re unknown in the new subject
Then this combination of eigenvectors provides the inner points
Repeat recursively for heirarchy
Small fudge factor TPS transformation for whole thalamus mis-alignment at the end of this procedure

14. Validation Do six rounds of leave-one-out cross validation
One is left out because it is the reference
Could have alternatively built the model many times for different reference, then maybe choose the reference that’s most similar to new subject

15. Results

16. Results

17. Results

18. Results Median Dice AV VA VLa VLP VPL Pul LGN MGN CM MD-Pf Hb MTT
Manual
0.842
0.779
0.715
0.815
0.734
0.891
0.837
0.720
0.830
0.885
0.678
0.690
THOMAS
0.774
0.701
0.633
0.865
0.729
0.736
0.784
0.884
0.691
0.556

19. Discussion
A thorough thalamic nuclei segmentation method that works with conventional 3T T1w images
Notes these problems:
Manual segmentation error is unquantified
Limited data to capture intersubject variation
No mention of registration accuracy of 7T to 3T esp. important for ground truths
Can do even better with images that actually have intra-thalamic contrast