Difference between revisions of "Initial atlas construction workflow"

From NAMIC Wiki
Jump to: navigation, search
m (moved Current atlas construction workflow to Initial atlas construction workflow: The atlas construction workflow has been revised. New workflow has been adopted.)
 
(7 intermediate revisions by 2 users not shown)
Line 1: Line 1:
== Objective ==
+
== Content ==
  
Develop an atlas construction workflow for vervet segmentation, which will maximally (completely) consist of the NA-MIC tools. Learn from the previous experience and try to mimic as closely as possible the workflow developed by Styner et al. [1]
+
This page describes the current working pipeline of atlas construction for vervet data.
  
 
== Data ==
 
== Data ==
Line 17: Line 17:
 
1.1. Select the subject that is best oriented as the template image  
 
1.1. Select the subject that is best oriented as the template image  
  
1.2. Semi-automatically identify the ICC in the template subject.
+
1.2. Manually identify the ICC in the template subject.
  
1.3. 6+9+12 DOF registration of each of the subjects to the template.
+
1.3. Perform intensity calibration for all other subjects to the template.
  
1.4. Skull-strip each subject using dilated ICC of the template.
+
1.4. 12 DOF registration of each of the subjects to the template.
  
1.5. Perform intensity calibration to the template.
+
1.5. Skull-strip each subject using dilated ICC of the template.
  
1.6. Use affine-registered, skull-stripped images to form initial affine average image.
 
 
||
 
||
 
1.1. n/a
 
1.1. n/a
  
1.2. Slicer Editor module, SkullStripper Slicer module
+
1.2. manual in SNAP.
  
1.3. Slicer BRAINSFit extension
+
1.3. Slicer Histogram Matching module
  
1.4. Slicer Mask, Editor modules (verify the mask encloses actual ICC!)
+
1.4. FSL-Flirt
  
1.5. Slicer Histogram Matching module
+
1.5. Slicer Mask, Editor modules (verify the mask encloses actual ICC!)
  
1.6. ''Tool to be developed'': compute average image from the set of input images
 
 
|-
 
|-
 
! 2. Atlas construction
 
! 2. Atlas construction
 
|  
 
|  
  
2.1. Use the affine registered image from Step 1.6 as the template to repeat Steps 1.1-1.6.
+
2.1. Use the skullstriped image from Step 1.2 as the template to repeat Step 1.4.
  
 
2.2. Perform non-rigid alignment of the subjects affinely registered to the template.
 
2.2. Perform non-rigid alignment of the subjects affinely registered to the template.
Line 51: Line 49:
 
2.1 see above  
 
2.1 see above  
  
2.2. Slicer BRAINSFit extension (BSpline), possibly Groupwise BSpline registration by Balci et al. [2]
+
2.2. Diffeomorphic demons
  
2.3. Tool from step 1.6.
+
2.3. Some tool in Slicer under development
  
 
|-
 
|-
Line 63: Line 61:
 
3.2. Slightly smooth the segmentations Styner et al. recommend kernel of 0.4 mm variance.
 
3.2. Slightly smooth the segmentations Styner et al. recommend kernel of 0.4 mm variance.
  
3.3. Back-propagate WM/GM/CSF segmentations to the training subjects using the deformation field from Step 2.2.
+
3.3. Probabilistic atlas is the result of smoothing of the hard segmentation of the averaged template with Gaussian kernel.
 
 
3.4. Construct the probabilistic atlas by averaging of the tissue maps, normalize locally to 1, construct rejection class by inverting the sum of WM+GM+CSF.
 
  
 
|  
 
|  
  
3.1. Slicer editor module (thresholding, manual editing)
+
3.1. Kmeans + manual
  
3.2. Gaussian smoothing module (check availability)
+
3.2. Gaussian smoothing module in Slicer, 0.6 and 0.65 kernel size used for Valentino and Tommy, respectively. '''AF: I do not understand what is the relation of Valentino and Tommy to this, if the operation was performed on hard segmentation from the average template'''
  
3.3. Slicer applyDeformation module (command line accessible only!)
+
3.3. Slicer Gaussian smoothing module
  
3.4. Tool from Step 1.6, tool to be developed, Slicer Editor module
+
*We did not perform back propagation as in Styner07 due to the large structural differences in the dataset.
  
 
|}
 
|}
 
== Standing questions ==
 
 
* when to perform bias correction? After Step 1.4?
 
* deep brain structures -- similar approach for identification in subjects as for GM/WM/CSF in Step 3.1?
 
* Styner et al. used T2 to segment CSF. We only have T1
 
  
 
== References ==
 
== References ==

Latest revision as of 17:51, 3 December 2009

Home < Initial atlas construction workflow

Content

This page describes the current working pipeline of atlas construction for vervet data.

Data

10 vervet subjects (2 subjects possibly have unacceptable image quality), T1 sequence

Workflow / Tools

Workflow step Description Tools
1. Preparation

1.1. Select the subject that is best oriented as the template image

1.2. Manually identify the ICC in the template subject.

1.3. Perform intensity calibration for all other subjects to the template.

1.4. 12 DOF registration of each of the subjects to the template.

1.5. Skull-strip each subject using dilated ICC of the template.

1.1. n/a

1.2. manual in SNAP.

1.3. Slicer Histogram Matching module

1.4. FSL-Flirt

1.5. Slicer Mask, Editor modules (verify the mask encloses actual ICC!)

2. Atlas construction

2.1. Use the skullstriped image from Step 1.2 as the template to repeat Step 1.4.

2.2. Perform non-rigid alignment of the subjects affinely registered to the template.

2.3. Compute the atlas as the average.

2.1 see above

2.2. Diffeomorphic demons

2.3. Some tool in Slicer under development

3. Probabilistic atlas construction

3.1. Segment WM/GM/CSF from the averaged template, manually edit to ensure accuracy

3.2. Slightly smooth the segmentations Styner et al. recommend kernel of 0.4 mm variance.

3.3. Probabilistic atlas is the result of smoothing of the hard segmentation of the averaged template with Gaussian kernel.

3.1. Kmeans + manual

3.2. Gaussian smoothing module in Slicer, 0.6 and 0.65 kernel size used for Valentino and Tommy, respectively. AF: I do not understand what is the relation of Valentino and Tommy to this, if the operation was performed on hard segmentation from the average template

3.3. Slicer Gaussian smoothing module

  • We did not perform back propagation as in Styner07 due to the large structural differences in the dataset.

References

  1. M. Styner, R. Knickmeyer, S. Joshi, C. Coe, S. J. Short, and J. Gilmore. Automatic brain segmentation in rhesus monkeys. Proc SPIE Medical Imaging Conference, Proc SPIE Vol 6512 Medical Imaging 2007, pp 65122L-1 - 65122L-8 pdf
  2. Balci S.K., Golland P., Wells W.M. Non-rigid Groupwise Registration using B-Spline Deformation Model. Insight Journal - 2007 MICCAI Open Science Workshop. link
  3. Previous descriptions of the atlas construction workflow: Summary by Ginger Li, Description off BSL atlas page