CCB SIG: Pipeline Neuroimaging Solutions (SIG-PNS) - Minimum Distance Template (MDT) Atlasing Pipeline (December 04, 2009; 2:00 PM, DIVE)
This demo will present the redesigned automated Pipeline
workflow for constructing the Minimum Distance Template (MDT) atlas for a given population, based on one specific image modality (e.g., sMRI). There will be two specific validation examples based on 100+ ICBM
- Quantitative: The ICBM data will be registered to the MDT, NRU and ICBM 462 atlases and mean displacement stats across all subjects will be reported for each voxel. The distributions of these mean displacement magnitudes for all 3 atlases will also be shown.
- Qualitative (visual): The resulting MDT atlas will be visually compared to the ICBM 462 atlas and the Danish NRU atlas constructed of the same 100+ subjects.
- All non-linear warping tools use the KL MI registration algorithm, described here. A symmetric optimization described here is applied in place of the original, making the algorithm much more robust and suitable for cross-subject registration.
- The KL MI tool is incorporated in a multiscale bundle which begins with a small working grid of 32^3 and goes on to 64^3 and 128^3. This approach allows for larger initial deformations where necessary, ensuring that the fine resolution step is devoted entirely to local deformations. Currently, no information about the previous step's final displacement field is available to the subsequent step, making this approach equivalent to regriding. This may be changed in the future.
- Input is assumed to be a set of linearly registered subjects, whose mean has been precomputed. The Air reconcile workflow is used for this. It is attached at the bottom. This workflow performs n^2 linear registrations and defines the common space based on the reconcile routine. This can take several hours, depending on the number of images, but is generally much better than simply picking one subject as target.
- This workflow follows the typical MDT algorithm, first using an affine mean to initially warp the images, then deforming the new mean according the inverse of the new warp field average (this creates the MDT), and finally registering all images to the newly created MDT. MDT as well as final Jacobians and registered images are saved.
- The workflow is faster than the iterative version below, but can still take up to 12 hours on a busy grid.
- This workflow selects a median image and iteratively moves it ever-closer to the middle of the data space.
- Median is defined as the image with the lowest L2 distance to the median image of the data set. It is in fact one of the inputs. The advantage of using this instead of a warp mean is that details are fully preserved. The disadvantage is that the warping is more susceptible to local minima. The new warping algorithm seems to overcome this quite well, though.
- Once a target image has been selected, it is used exactly as the MDT, except iteratively: with each new iteration the new field average is inverted and composed with the previous target warp, moving the image a bit closer to the middle.
- There is no limit on the number of iterations. Currently it is set at 3 (plus the final registration), but one can see that the overall registration continues to improve across input images even at the last step. It is also interesting that often the final target looks very similar in general form and large ROIS like the ventricles and the corpus callosum to the MDT result of the previous workflow. The only difference is sharpness is level of detail.
- This workflow registers all images to a selected median, similarly to the iterative workflow. The difference is that here it is only done once, and no inversion is applied.
- This worflow uses the ICBM atlas to align images into a common space. Images are pre-aligned linearly.
- We found that 12-parameter AIR does not work very well here, even with histogram equalization and bias correction. We simply use 9-parameter registration instead.
Results & Validation
Feedback & suggestions
- MDT_TBD.pipe: TBD MDT atlas construction
- MDT with KLMI multiscale symmetric: This workflow uses a symmetric optimization and a multiscale approach to create an MDT from linearly registered data.
- MDT with KLMI multiscale iterative: This workflow applies the multiscale symmetric warping iteratively to an automatically selected median target. The resulting atlas has all the detials of a single subject while lying in the middle of subject space.
- MDT with KLMI 1subject bias: This workflow selects a median image and warps the rest to it once. Linear registration is assumed.
- KLMI existing atlas: This workflow shows an example of using an existing atlas (in this case the ICBM) with this warping. Bias correction and histogram equalization are included.
* AIRreconcile_v2: This workflow uses air reconcile to linearly align all images to a common space.
Tensor-Based Morphometry (TBM)
The Minimum Distance Template (MDT) construction and subsequent Tensor Based Morphometry (TBM) workflows are actively developed and constantly evolving. The TBM protocol is divided into 3 steps:
- TBM_step1: analogous to the existing ‘AIR Reconcile’ workflow; puts all subjects into a common space using a reconciled all-to-all registration.
- TBM_step2: iteratively, non-linearly registers each brain (from the previous step) into an average space, then uses these brains to construct a minimum distance template to be used as an atlas in the next step.
- TBM_step3: non-linearly registers the brains from the first step into the MDT, computes the Jacobian determinants of the displacement fields, and runs two independent group t-test.