| Deletions are marked like this. | Additions are marked like this. |
| Line 7: | Line 7: |
| Improvements to the new longitudinal stream (only on volume related issues, surface improvements to come). | Improvements to the new longitudinal stream (only on volume related issues). |
| Line 10: | Line 10: |
|
* only '''one''' template estimation on the norm.mgz files of all timepoints * a new block '''-base-init''' added to recon-all for the initial template creation * this creates the maps from each TP to base and the '''norm_template''' * then create an orig_template as orig/001.mgz to initialize the base run * the norm_template is used as brainmask * in gcareg and canorm the norm_template is used instead of the nu * this setup reduced the difficulties of dealing with two spaces (two template estimations) * should lead to significant run time improvements especially with several time points |
* Only '''one''' template estimation on the norm.mgz files of all timepoints, to prevent difficulties with two possibly different template locations (nu and norm as in 4.4). * A new block '''-base-init''' added to recon-all for the initial template creation, to be able to call later parts of the script individually without re-running the initialization everytime. * The -base-init block creates the maps from each TP to base and the '''norm_template''' (using the norm.mgz of all TPs). * Then an orig_template is created as orig/001.mgz to initialize the base run. * The norm_template is used as brainmask (as it contains only brain). * In -gcareg and -canorm the norm_template is used instead of the nu.mgz. * This setup (with only one estimation) should lead to significant run time improvements especially with several time points. |
| Line 20: | Line 19: |
|
* the talairach.lta is now created by concatenation (tpN -> base -> talairach) * the brainmask is copied/mapped from the base (which is basically an OR of all TPs) * mri_ca_normalize has a new -long algorithm using the base aseg as init * mri_ca_label now uses the correct intensity scaling factors of the base |
* The talairach.lta is now created by concatenation (tpN -> base -> talairach) * The brainmask is copied/mapped from the base (which is basically an OR of all TPs) * mri_ca_normalize has a new -long algorithm using the base aseg as init, fixing the bias towards smaller volumes in subcortical structures, as seen in 4.4. * mri_ca_label now uses the correct intensity scaling factors of the base. |
| Line 30: | Line 29: |
|
* We have one '''-base run''' and '''-long runs''' for each TP * A base template (median) is created and used to initialize the longitudinal runs * The base is '''unbiased''' and can be viewed as an initial guess where things are |
* We have one '''-base run''' and '''-long runs''' for each TP. * A base template (median) is created and used to initialize the longitudinal runs. * The base is '''unbiased''' and can be viewed as an initial guess where things are. |
| Line 34: | Line 33: |
|
* and [[mri_robust_template]] (unbiased robust template estimation) * Probabilistic fusion was added [[mri_fuse_segmentations]] to incorporate label information from the other TPs at a specific location * All TPs have to be processed cross sectionally (independently) first |
* and [[mri_robust_template]] (unbiased robust template estimation). * Probabilistic fusion was added [[mri_fuse_segmentations]] to incorporate label information from other TPs at a specific location. * All TPs have to be processed cross sectionally (independently) first. |
| Line 42: | Line 41: |
| The stream is described here: [[LongitudinalProcessingPreV4.4]]. | The old stream is described here: [[LongitudinalProcessingPreV4.4]]. |
| Line 45: | Line 44: |
|
* One of the TPs was used to initialize the others (TP1 by default) * This lead to a bias wrt to TP1 * selecting a different TP for the initialization completely changed the results * A fix to init TP1 with itself improved things, but did not remove the bias |
* One of the TPs was used to initialize the others (TP1 by default). * This lead to a bias wrt to TP1. * Selecting a different TP for the initialization completely changed the results. * A quick fix to init TP1 with itself improved things, but did not remove the bias. |
Longitudinal Stream Change Log
The description of the current longitudinal stream can be found here: LongitudinalProcessing
FS Version 4.5
Improvements to the new longitudinal stream (only on volume related issues).
Base Stream:
Only one template estimation on the norm.mgz files of all timepoints, to prevent difficulties with two possibly different template locations (nu and norm as in 4.4).
A new block -base-init added to recon-all for the initial template creation, to be able to call later parts of the script individually without re-running the initialization everytime.
The -base-init block creates the maps from each TP to base and the norm_template (using the norm.mgz of all TPs).
- Then an orig_template is created as orig/001.mgz to initialize the base run.
- The norm_template is used as brainmask (as it contains only brain).
- In -gcareg and -canorm the norm_template is used instead of the nu.mgz.
- This setup (with only one estimation) should lead to significant run time improvements especially with several time points.
Long Stream:
The talairach.lta is now created by concatenation (tpN -> base -> talairach)
- The brainmask is copied/mapped from the base (which is basically an OR of all TPs)
- mri_ca_normalize has a new -long algorithm using the base aseg as init, fixing the bias towards smaller volumes in subcortical structures, as seen in 4.4.
- mri_ca_label now uses the correct intensity scaling factors of the base.
FS Version 4.4
First version of a working longitudinal stream, without optimizing each step.
Main Differences:
We have one -base run and -long runs for each TP.
- A base template (median) is created and used to initialize the longitudinal runs.
The base is unbiased and can be viewed as an initial guess where things are.
New tools: mri_robust_register (symmetric registration)
and mri_robust_template (unbiased robust template estimation).
Probabilistic fusion was added mri_fuse_segmentations to incorporate label information from other TPs at a specific location.
- All TPs have to be processed cross sectionally (independently) first.
FS pre Version 4.4
These old versions of the longitudinal stream should not be used!
The old stream is described here: LongitudinalProcessingPreV4.4.
Short Info:
- One of the TPs was used to initialize the others (TP1 by default).
- This lead to a bias wrt to TP1.
- Selecting a different TP for the initialization completely changed the results.
- A quick fix to init TP1 with itself improved things, but did not remove the bias.
Original Author: MartinReuter