|Table of Contents|
This page contains workflow diagrams for simple use cases of the four core Pydpiper applications. As the number of core Pydpiper applications increases, we will create more example workflow diagrams.
To view a larger version of each workflow, please click the thumbnail image. The png files may also be downloaded for easier resizing and viewing.
Iterative Model Building (MBM 2.0)
This example workflow is for iterative model building with 3 subjects (subj_1, subj_2, and subj_3). The parameters of the workflow are as follows:
- An LSQ6 alignment towards an initial model, using large rotation parameters.
- Pairwise LSQ12 registration between all three pairs of brains, using three generations of minctracc calls
- Three generations of iterative, non-linear registration using mincANTS.
- Statistics calculations from the final non-linear average back to each individual subject, using two different blurring kernels.
This example workflow is for the entire MAGeT procedure, starting with three atlases (A1, A2, A3) and three inputs (subj_1, subj_2, subj_3). The workflow is as follows:
- Each atlas is registered to each input using two linear minctracc calls and six non-linear minctracc calls.
- The labels from each input atlas are resampled to the space of each input, using the final resulting transform from the registration in the previous step. This creates three sets of input labels for each of subj_1, subj_2 and subj_3, from A1, A2 and A3 respectively.
- Each of the inputs is registered to every other input, using the same registration parameters that were used for the atlas to input registration.
- Each set of input labels is then resampled to the space of each target. For example, for subj_1, labels 1A1, 1A2 and 1A3 are resampled to subj_2 and subj_3. This creates six labels for each input.
- For each input, a voxel voting procedure is run on its six labels, creating a final voted set.
This example workflow is for a simple registration chain application. Two subjects, each scanned at three time points, are registered together. The registration proceeds as follows:
- For each subject, scan T1 is registered to scan T2 using two linear minctracc calls (12-parameter) and one mincANTS call.
- For each subject, scan T2 is registered to scan T3 using the same parameters as the T1 to T2 registration.
- Statistics are calculated from each T1 to T2 scan and each T2 to T3 scan, using two different blurring kernels.
We explicitly note here that we do NOT show the portion of the workflow where the T3 scans from both subjects are registered together to create a common space using the iterative model building procedure shown above. We omitted this step to aid in workflow readability, but should one wish to visualize adding this component, it would include the entire diagram shown above for iterative model building, minus subject three. In addition, additional transform concatenation steps and statistics would be added, for the calculation of deformation fields from the common space back to each individual timepoint.
Two-level Model Building
This example workflow is for two-level model building. Two subjects, each scanned twice, are assumed to have been aligned previously with the LSQ12 module. They are registered with the following protocol:
- Scans T1 and T2 for subject 1 are averaged together to create an initial target.
- The individual subject scans are non-linearly aligned using a three-stage mincANTS optimization, using the average from the previous stage as the initial target. This corresponds to the final stage of the iterative model building procedure described above and builds a consensus average for subject 1.
- The above two steps are executed for subject 2, creating a consensus average for subject 2.
- For each subject, statistics are calculated from its respective consensus average to each individual scan.
- The consensus averages for subject 1 and subject 2 are averaged together to create an initial target. These individual averages are then iteratively aligned to the target in the same procedure that was used to create the subject averages.
- Statistics are calculated from the inter-subject average back to each individual subject average.
- Using the transform from each subject average to the inter-subject average, resample all of the within-subject statistics to the common space created by the inter-subject average.