Hi Mark,
Thanks for your insight and apologies for the late reply.
I have been experimenting with some parameters in FLIRT. In particular, I tried the following four options:
1) correlation ratio cost function, which is the default setting in FLIRT;
2) correlation ratio cost function with searchrx, searchry, and searchrz all -75 to +75;
3) mutual information cost function;
4) mutual information cost function with searchrx, searchry, and searchrz all -75 to +75.
I agree with your second point (and have read from a few resources) that mutual information should be the cost function that is used when registering from CT to MRI space. But to my (untrained) eye, the CT brains and stroke lesions registered to MNI space for options #2 and #3 above look very similar (see download link below), even though both options use different cost functions. Would you have any insight as to why this may be the case?
https://1drv.ms/u/s!AgwBDHvQMsQqawvP61Zgbp1LZL8
In regards to your third point about brain extraction, both the CT brains and MNI template that I used are skull-stripped for FLIRT. I have showed some of the MNI-registered brains to a colleague who is more knowledgeable in registration. Although some lesions overlap the ventricles in the MNI brain (also shown in the attached link), he believed that the results thus far with FLIRT are satisfactory (given the difficulty of this situation). However, he suggested that non-linear registration may help to improve the registration to the MNI template.
From my understanding, FNIRT requires us to input raw brain images (i.e., non-skull-stripped brains). Given that you mention in your third point that non-brain structures in the CT are very prominent and likely causing registration issues, does this mean it would not be possible to use FNIRT to register CT images since I would be including the skull and (inevitably) other artifacts that would disrupt proper non-linear registration?
Would FNIRT potentially work if I rescale the voxel intensities in the CT scan such that the negative voxel intensities (which mainly correspond to artifacts in the CT image) are no longer affecting the registration?
I appreciate any insight you may have.
Thanks for your help,
Sam
On Wed, 5 Jul 2017 08:31:42 +0000, Mark Jenkinson <[log in to unmask]> wrote:
>Dear Sam,
>
>Firstly, this kind of registration will be quite difficult, so you probably have to try several approaches in order to find one that works well.
>
>Secondly, if you are not using the mutual information cost function (or normalised mutual information) then switch to that, as this is crucial for CT to MRI registration.
>
>Thirdly, are you brain extracting both images? If not, then do this as the non-brain structures in the CT are so prominent that they are likely to be causing major problems.
>
>Lastly, your approach sounds fine (given the above). You may need a very large up-weighting on your ventricle mask in order for it to have an effective (e.g., values of 100 or 1000) but the principle is sound. Alternatively, if you want to specify individual point landmarks then you might find that the pointflirt tool is more useful for you.
>
>All the best,
> Mark
>
>
>> On 20 Jun 2017, at 05:45, Sam Choi <[log in to unmask]> wrote:
>>
>> Dear FSLers,
>>
>> I am working with CT scans of stroke patients and would like to register the brains and stroke lesions in CT space into MNI space. Currently, the registration methods I have been trying have not been working well. The registered lesions often overlap the ventricles in the MNI template.
>>
>> A suggestion that I received is to try manual registration. Specifically, I would select certain areas/landmarks around the ventricles in the CT scan and register these areas to the MNI template. The expectation of this manual registration process is to improve registration around the ventricles such that the stroke lesions do not overlap the ventricles in the MNI template.
>>
>> With this in mind, how do I perform manual registration using FSL tools? Would the following steps make sense:
>> 1) Using the CT scan, create a mask of the areas/landmarks around the ventricles that I want to improve registration (for brevity, I will refer to this as the "ventricle mask" herein).
>>
>> 2) Given that I am dealing with stroke patients, I also need to create an inverted lesion mask such that FLIRT ignores the lesion areas during registration (i.e., lesion areas are weighted 0, whereas rest of brain is weighted 1).
>>
>> 3) Use FSLMATHS to add the ventricle mask and the inverted lesion mask together. The resulting mask file would lead to the lesion areas still having a weight of 0, while the regions part of the ventricle mask will have a weight >1.
>>
>> 4) Use FLIRT to register the CT scan to the MNI template. Use the -inweight option to input the mask created from Step 3 (above).
>>
>> I would appreciate any insight you may have.
>> Thanks for your help!
>>
>> Sam
|