[caret-users] new graph topology correction truncated occipital lobe

Donna Dierker donna at brainvis.wustl.edu
Fri Jan 9 12:46:19 CST 2009 

Hi all,

Recently, John added a graph-based topology correction algorithm to 
Caret (see reference below), and we changed the default segmentation 
setting to SUREFIT_THEN_GRAPH, because this combination gave the most 
anatomically correct segmentations in the cases we tested.

The attached capture shows a chunk of occipital cortex missing after 
correction using the graph method.  When error correction is set to 
SUREFIT only, the problem disappears (but potentially several small 
handles reappear;-).  Note that you can run a segmentation both ways -- 
with and without graph -- and then use the Edit Voxels : Copy Subvolume 
to paste in the portions of the no-graph segmentation where graph 
misbehaved.

Overall, we're happy with how the SUREFIT_THEN_GRAPH combination is 
performing, but thought it worth passing on the information.

Donna

-------- Original Message --------
Subject: 	Re: odd segmentation
Date: 	Fri, 9 Jan 2009 10:26:50 -0600
From: 	Alexander Li Cohen <alexc at npg.wustl.edu>
To: 	Donna Dierker <donna at brainvis.wustl.edu>
CC: 	Erin Reid <erin at brainvis.wustl.edu>
References: 	<E11C0B75-2E9D-4241-AB26-AD337DA8BB08 at npg.wustl.edu> 
<496662A7.1000505 at brainvis.wustl.edu> 
<496666B6.4010702 at brainvis.wustl.edu> 
<F76DB883-669E-4D18-A731-08628DD144C2 at npg.wustl.edu> 
<496670F3.7050107 at brainvis.wustl.edu> 
<BAF51C41-76F7-4FC6-900B-32C3DA8C0E6A at npg.wustl.edu> 
<496679D8.2090209 at brainvis.wustl.edu> 
<9E9C1A32-A8F1-43B8-8908-4950DC7E4416 at npg.wustl.edu> 
<49667DFE.3010703 at brainvis.wustl.edu> 
<49668362.2000006 at brainvis.wustl.edu> 
<49676BF1.4040808 at brainvis.wustl.edu> 
<60C67933-14AE-4377-8C52-68D2DF17DDD5 at npg.wustl.edu> 
<496779D6.2080508 at brainvis.wustl.edu>



yup, based on what it did, i can see why it did it for that part of  
occipital cortex...

-Alex
~0-------------

----------------------------------------------------------------------
Alexander Li Cohen
alexc at npg.wustl.edu (WORK Email)
alexcohen at gmail.com (Non-secure / rapid-response Email)
Washington University in St. Louis School of Medicine
Medical Scientist Training Program
----------------------------------------------------------------------




On Jan 9, 2009, at 10:22 AM, Donna Dierker wrote:

> Alex
>
> I think graph = graph theory; per John:
>
>> I have added a new volume topology error correction algorithm to  
>> the segmentation process that produces topologically correct  
>> segmentation volumes and surfaces. This error correction algorithm  
>> is based off an algorithm from Automated Graph-Based Analysis and  
>> Correction of Cortical Volume Topology by David W. Shattuck and  
>> Richard M. Leahy (http://ieeexplore.ieee.org/search/srchabstract.jsp?arnumber=963819&isnumber=20807&punumber=42&k2dockey=963819@ieeejrns&query=%28%28shattuck%29%3Cin%3Eau+%29&pos=2&access=no 
>>  PDF link near left, top).
>>
>> This algorithm first examines each slice and each connected region  
>> in the slice becomes a vertex in a graph (http://en.wikipedia.org/wiki/Graph_theory 
>>  ). The edges in the graph are built by determining the connections  
>> from each slice region to regions in adjacent slices. The graph is  
>> then searched for cycles that indicate the presence of handles.
>>
>> The correction algorithm I implemented is not very sophisticated.  
>> It examines each cycle in the graph and either fills in the handle  
>> or cuts the handle off. These corrections are ordered so that the  
>> algorithm always makes the smallest correction available based upon  
>> the current topology. I believe the algorithm could be enhanced by  
>> examining anatomy volume intensities, stereotaxic coordinates, a  
>> probabilistic volume, or surface properties. For example, in the  
>> image below, there is a handle in the calcarine sulcus that should  
>> be cut rather than filled. The algorithm has been test on six  
>> hemispheres, always producing topologically correct volumes and  
>> surfaces as shown in the following table.
> Donna
>
> On 01/09/2009 10:07 AM, Alexander Li Cohen wrote:
>> sure, it seems we've narrowed it down to the graph error correction  
>> as well. (which actually makes sense, if graph = graph theory...  
>> but i doubt that...)
>>
>> -Alex
>> ~)----------------
>>
>> ----------------------------------------------------------------------
>> Alexander Li Cohen
>> alexc at npg.wustl.edu (WORK Email)
>> alexcohen at gmail.com (Non-secure / rapid-response Email)
>> Washington University in St. Louis School of Medicine
>> Medical Scientist Training Program
>> ----------------------------------------------------------------------
>>
>>
>>
>>
>> On Jan 9, 2009, at 9:23 AM, Donna Dierker wrote:
>>
>>> Alex,
>>>
>>> Do you mind if I forward this thread to caret-users with your  
>>> capture?
>>>
>>> Donna
>>>
>>> On 01/08/2009 04:51 PM, Erin Reid wrote:
>>>> ok ... i got a full surface *with* running error correction  
>>>> (surefit only tho).
>>>>
>>>> Donna Dierker wrote:
>>>>> If overall error correction did good things, then remember you  
>>>>> can use Copy Subvolume on the Edit Voxels to copy a portion of a  
>>>>> volume into another volume.
>>>>>
>>>>> On 01/08/2009 04:16 PM, Alexander Li Cohen wrote:
>>>>>> that did it. it was the error correction screwing up...  
>>>>>> probably couldn't find a path along white matter connecting the  
>>>>>> occipital lobe to the rest........
>>>>>>
>>>>>> -Alex
>>>>>> ~)-------------
>>>>>>
>>>>>> ----------------------------------------------------------------------
>>>>>> Alexander Li Cohen
>>>>>> alexc at npg.wustl.edu <mailto:alexc at npg.wustl.edu> (WORK Email)
>>>>>> alexcohen at gmail.com <mailto:alexcohen at gmail.com> (Non-secure /  
>>>>>> rapid-response Email)
>>>>>> Washington University in St. Louis School of Medicine
>>>>>> Medical Scientist Training Program
>>>>>> ----------------------------------------------------------------------
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Jan 8, 2009, at 4:10 PM, Erin Reid wrote:
>>>>>>
>>>>>>> just to do the quick version I only had the following selected:
>>>>>>> generate raw and fid surfaces
>>>>>>> limit polygons
>>>>>>> generate inflated (which wasn't necessary of course)
>>>>>>> generate cerebral hull
>>>>>>> create curvature, depth ...
>>>>>>> auto save
>>>>>>>
>>>>>>> E
>>>>>>>
>>>>>>> Alexander Li Cohen wrote:
>>>>>>>> what are the bare minimum of seg vol options? I thought I had  
>>>>>>>> it, and it screwed up again...
>>>>>>>>
>>>>>>>> -alex
>>>>>>>> ~0----------
>>>>>>>>
>>>>>>>> ----------------------------------------------------------------------
>>>>>>>> Alexander Li Cohen
>>>>>>>> alexc at npg.wustl.edu <mailto:alexc at npg.wustl.edu> (WORK Email)
>>>>>>>> alexcohen at gmail.com <mailto:alexcohen at gmail.com> (Non- 
>>>>>>>> secure / rapid-response Email)
>>>>>>>> Washington University in St. Louis School of Medicine
>>>>>>>> Medical Scientist Training Program
>>>>>>>> ----------------------------------------------------------------------
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On Jan 8, 2009, at 3:32 PM, Erin Reid wrote:
>>>>>>>>
>>>>>>>>> I ran it with the following settings:
>>>>>>>>> gray = 118
>>>>>>>>> white = 187
>>>>>>>>>
>>>>>>>>> disconnect hindbrain = low thresh
>>>>>>>>> and bare minimum in the seg vol options
>>>>>>>>>
>>>>>>>>> and managed to get a surface of an entire hemisphere, so  
>>>>>>>>> your white setting might just need to be brought down some.
>>>>>>>>>
>>>>>>>>> Hope this helps
>>>>>>>>> Erin
>>>>>>>>>
>>>>>>>>> <moz-screenshot-2.jpg>
>>>>>>>>>
>>>>>>>>> Alexander Li Cohen wrote:
>>>>>>>>>> Left is correctly identified, and using a low thresh for  
>>>>>>>>>> hindbrain didn't change it. I've uploaded the relevant  
>>>>>>>>>> files, help!!!
>>>>>>>>>>
>>>>>>>>>> file is: Alexs_missing_occipital.zip
>>>>>>>>>>
>>>>>>>>>> -Alex
>>>>>>>>>> ~0---------------
>>>>>>>>>>
>>>>>>>>>> ----------------------------------------------------------------------
>>>>>>>>>> Alexander Li Cohen
>>>>>>>>>> alexc at npg.wustl.edu (WORK Email)
>>>>>>>>>> alexcohen at gmail.com (Non-secure / rapid-response Email)
>>>>>>>>>> Washington University in St. Louis School of Medicine
>>>>>>>>>> Medical Scientist Training Program
>>>>>>>>>> ----------------------------------------------------------------------
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On Jan 8, 2009, at 2:48 PM, Erin Reid wrote:
>>>>>>>>>>
>>>>>>>>>>> How sad ... I really don't know click and clack. :-(
>>>>>>>>>>> And yes, stumped on the results. I would be interested in  
>>>>>>>>>>> trying it myself to see if it recreation is possible.
>>>>>>>>>>> E
>>>>>>>>>>>
>>>>>>>>>>> Donna Dierker wrote:
>>>>>>>>>>>> Alex,
>>>>>>>>>>>>
>>>>>>>>>>>> First, check two things:
>>>>>>>>>>>>
>>>>>>>>>>>> 1) Make sure you have identified the hemisphere correctly  
>>>>>>>>>>>> in the spec file.
>>>>>>>>>>>>
>>>>>>>>>>>> 2) When running initial segmentation, turn hindbrain  
>>>>>>>>>>>> thresh to LOW; see it it helps.
>>>>>>>>>>>>
>>>>>>>>>>>> If neither of these helps, then zip the archive and  
>>>>>>>>>>>> upload it here:
>>>>>>>>>>>>
>>>>>>>>>>>> http://pulvinar.wustl.edu/cgi-bin/upload.cgi
>>>>>>>>>>>>
>>>>>>>>>>>> I love click and clack, but I bet Erin has no idea what  
>>>>>>>>>>>> you're talking about.
>>>>>>>>>>>>
>>>>>>>>>>>> Donna
>>>>>>>>>>>>
>>>>>>>>>>>> On 01/08/2009 02:01 PM, Alexander Li Cohen wrote:
>>>>>>>>>>>>> Hi,
>>>>>>>>>>>>> ok, next stumper (I feel like click and clack, each week  
>>>>>>>>>>>>> I've got a new stumper for ya'll)... first step of  
>>>>>>>>>>>>> segmentation, initial result gives me the attached  
>>>>>>>>>>>>> segmentation volume/fiducial... what happened to the  
>>>>>>>>>>>>> missing slices of occipital cortex? I tried running it  
>>>>>>>>>>>>> twice, same result...
>>>>>>>>>>>>>
>>>>>>>>>>>>> -Alex
>>>>>>>>>>>>> ~)---------------
>>>>>>>>>>>>>
>>>>>>>>>>>>> ----------------------------------------------------------------------
>>>>>>>>>>>>> Alexander Li Cohen
>>>>>>>>>>>>> alexc at npg.wustl.edu (WORK Email)
>>>>>>>>>>>>> alexcohen at gmail.com (Non-secure / rapid-response Email)
>>>>>>>>>>>>> Washington University in St. Louis School of Medicine
>>>>>>>>>>>>> Medical Scientist Training Program
>>>>>>>>>>>>> ----------------------------------------------------------------------
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>
>>>>>>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: missing_occipital.jpg
Type: image/jpeg
Size: 145018 bytes
Desc: not available
Url : http://brainvis.wustl.edu/pipermail/caret-users/attachments/20090109/e56ba180/attachment-0001.jpg

More information about the caret-users mailing list