help
help > RE: Implicit vs explicit masking and its effect on denoising
Dec 23, 2022 12:12 AM | Alfonso Nieto-Castanon - Boston University
RE: Implicit vs explicit masking and its effect on denoising
Dear Pavel,
Regarding (1) BOLD data that is outside of the FOV of a particular session is treated as 0-values (this is true irrespective of the choice of implicit- or explicit- masking)
Regarding (2), implicit masking uses SPM procedure to define the analysis mask (analysis mask defined as voxels with BOLD signal above 0.80*A, where A is the average BOLD signal across voxels with values above M/8, and M is the average BOLD signal across the entire volume)
Regarding (3), QC-FC analyses are typically defined from a subset of voxels within the (explicit) analysis mask (that is needed so that a given voxel points to the "same" location across different subjects). In the context of implicit-masking (which is often used in the context of subject-space analyses) you should have got a warning saying:
warning: 1000-node network sample for histogram displays created using subject-specific grey-matter mask voxels. Quality Control QC-FC correlation plots may be inaccurate (if grey-matter masks are different across subjects). Enter a subject-independent Grey Matter ROI in order to avoid this warning
which means that CONN is switching to using voxels within subject-specific gray-matter masks for QC plots, and while that will be fine for other QC plots/analyses this choice will invalidate the assumptions behind QC-FC analyses. If you want to fix this and still be able to use QC-FC analyses in the context of implicit masking you may do so by entering in the 'Gray matter' ROI a mask common across all subjects identifying likely gray matter voxels (and CONN will use that mask to extract the 1000-node network samples used for QC plots).
Last, regarding (4), there is no explicit option for that, but I imagine you could mask your functional data directly by your choice of explicit mask during preprocessing (e.g. using the 'functional_mask' preprocessing step), and then use 'implicit masking' to complement that with SPM's implicit masking procedure. That said, I think in your case it is probably simpler/better to manually define your choice of mask from the data (e.g. take the 'art_mean_*.nii' images from each subject, threshold those images and then compute the intersection of all of them; e.g. you could do both of those things on a single command-line using spm_imcalc), and then simply enter that mask as a explicit mask in CONN's analysis mask option.
Hope this helps
Alfonso
Originally posted by Pavel Hok:
Regarding (1) BOLD data that is outside of the FOV of a particular session is treated as 0-values (this is true irrespective of the choice of implicit- or explicit- masking)
Regarding (2), implicit masking uses SPM procedure to define the analysis mask (analysis mask defined as voxels with BOLD signal above 0.80*A, where A is the average BOLD signal across voxels with values above M/8, and M is the average BOLD signal across the entire volume)
Regarding (3), QC-FC analyses are typically defined from a subset of voxels within the (explicit) analysis mask (that is needed so that a given voxel points to the "same" location across different subjects). In the context of implicit-masking (which is often used in the context of subject-space analyses) you should have got a warning saying:
warning: 1000-node network sample for histogram displays created using subject-specific grey-matter mask voxels. Quality Control QC-FC correlation plots may be inaccurate (if grey-matter masks are different across subjects). Enter a subject-independent Grey Matter ROI in order to avoid this warning
which means that CONN is switching to using voxels within subject-specific gray-matter masks for QC plots, and while that will be fine for other QC plots/analyses this choice will invalidate the assumptions behind QC-FC analyses. If you want to fix this and still be able to use QC-FC analyses in the context of implicit masking you may do so by entering in the 'Gray matter' ROI a mask common across all subjects identifying likely gray matter voxels (and CONN will use that mask to extract the 1000-node network samples used for QC plots).
Last, regarding (4), there is no explicit option for that, but I imagine you could mask your functional data directly by your choice of explicit mask during preprocessing (e.g. using the 'functional_mask' preprocessing step), and then use 'implicit masking' to complement that with SPM's implicit masking procedure. That said, I think in your case it is probably simpler/better to manually define your choice of mask from the data (e.g. take the 'art_mean_*.nii' images from each subject, threshold those images and then compute the intersection of all of them; e.g. you could do both of those things on a single command-line using spm_imcalc), and then simply enter that mask as a explicit mask in CONN's analysis mask option.
Hope this helps
Alfonso
Originally posted by Pavel Hok:
Dear CONN maintainers,
I have searched the forum (found especially this thread: https://www.nitrc.org/forum/message.php?... ), but I could not find answers to the following questions.
Background: I am running ROI-to-ROI, SBC analysis, and ICA. Some subjects have parts of the cerebellum missing due to an inappropriate FoV setting. Except for smoothing, the whole pre-processing was performed outside the CONN (e.g., normalized unsmoothed BOLD images were imported).
Questions:
1. How does CONN account for missing voxels when explicit mask is used? With explicit mask, the results viewer for SBC analysis still shows voxels in the cerebellum.
2. How is the implicit mask created? With implicit masking, cerebellum is (correctly) no longer visible in the results, but the resulting group brain mask contains non-brain tissue, such as the eyes and parts of the skull.
3. How does the explicit mask affect denoising? With the explicit mask, QC-FC associations indicate much worse performance (e.g., 88% match with NH compared to 97% match with NH when the implicit mask is used). Does the masking only affect the plots (since they also show voxels not covered by FoV) or does it influence the denoising procedure itself (e.g., by extracting CSF/WM signal from voxels where there are no data)?
4. Is there a way to force an intersection of implicit and explicit mask? That would probably solve all my issues.
Kind regards
Pavel Hok
I have searched the forum (found especially this thread: https://www.nitrc.org/forum/message.php?... ), but I could not find answers to the following questions.
Background: I am running ROI-to-ROI, SBC analysis, and ICA. Some subjects have parts of the cerebellum missing due to an inappropriate FoV setting. Except for smoothing, the whole pre-processing was performed outside the CONN (e.g., normalized unsmoothed BOLD images were imported).
Questions:
1. How does CONN account for missing voxels when explicit mask is used? With explicit mask, the results viewer for SBC analysis still shows voxels in the cerebellum.
2. How is the implicit mask created? With implicit masking, cerebellum is (correctly) no longer visible in the results, but the resulting group brain mask contains non-brain tissue, such as the eyes and parts of the skull.
3. How does the explicit mask affect denoising? With the explicit mask, QC-FC associations indicate much worse performance (e.g., 88% match with NH compared to 97% match with NH when the implicit mask is used). Does the masking only affect the plots (since they also show voxels not covered by FoV) or does it influence the denoising procedure itself (e.g., by extracting CSF/WM signal from voxels where there are no data)?
4. Is there a way to force an intersection of implicit and explicit mask? That would probably solve all my issues.
Kind regards
Pavel Hok
Threaded View
| Title | Author | Date |
|---|---|---|
| Pavel Hok | Dec 1, 2022 | |
| Alfonso Nieto-Castanon | Dec 23, 2022 | |