help
489 Subscribers 
help > Overlap within and between networks in networks.nii
Showing 1-5 of 5 posts
Jul 8, 2019 04:07 PM | Pravesh Parekh - National Institute of Mental Health and Neurosciences
Overlap within and between networks in networks.nii
Dear Dr. Alfonso,
I was exploring the networks.nii file that is shipped with Conn and I noticed that there are several voxels which have multiple labels. For example, [-58 -38 34] is labeled in volumes 17, 21, and 29. There are a total of two voxels which are labeled in three volumes and 3789 voxels which are labeled by two different volumes (where volume refers to the volume number in the 4D file: vol1 = DefaultMode.MPFC, and so forth). I have attached a list of all such voxels and the corresponding volume numbers in which they have been labeled.
My concern is that overlapping voxels between ROIs will affect the connectivity between these two regions given the perfectly correlated time series. Perhaps a hard threshold after ICA is required so that these voxels get assigned to either of the overlapping regions? What are your thoughts on this?
Warm Regards
Pravesh
I was exploring the networks.nii file that is shipped with Conn and I noticed that there are several voxels which have multiple labels. For example, [-58 -38 34] is labeled in volumes 17, 21, and 29. There are a total of two voxels which are labeled in three volumes and 3789 voxels which are labeled by two different volumes (where volume refers to the volume number in the 4D file: vol1 = DefaultMode.MPFC, and so forth). I have attached a list of all such voxels and the corresponding volume numbers in which they have been labeled.
My concern is that overlapping voxels between ROIs will affect the connectivity between these two regions given the perfectly correlated time series. Perhaps a hard threshold after ICA is required so that these voxels get assigned to either of the overlapping regions? What are your thoughts on this?
Warm Regards
Pravesh
Jul 24, 2019 06:07 AM | Pravesh Parekh - National Institute of Mental Health and Neurosciences
RE: Overlap within and between networks in networks.nii
Bumping this up!
Aug 14, 2019 06:08 PM | Pravesh Parekh - National Institute of Mental Health and Neurosciences
RE: Overlap within and between networks in networks.nii
Bumping this question up!
Aug 19, 2019 04:08 PM | Alfonso Nieto-Castanon - Boston University
RE: Overlap within and between networks in networks.nii
Dear Pravesh,
That's an interesting question. Traditionally, ICA components are hard-thresholded in order to identify those areas most representative of each network (this is the same approach used to define the areas in CONN's 'networks.nii' atlas). But, almost by definition, spatial ICA components are not truly non-overlapping. More specifically, when defining spatially-independent components, the distribution of spatial loadings for two different components must be independent, which, somewhat informally, means that if you know the value at a given voxel of the loading of one component, that knowledge cannot give you any clue about what the loading of a second component will be for that same voxel. Non-overlapping components would fail to be independent, as knowing that a voxel loads heavily in one component (so that, when thresholded, this voxel is identified as part of a given network) would imply that the loading for other components at that same voxel must be low (so that, when thresholded, this same voxel is not identified as part of other networks), contrary to the independence definition.
One possible work-around that is sometimes used when you really want to define truly non-overlapping areas while still using this same underlying ICA framework, is to, instead of using a hard-threshold value to select which areas are associated with each network (or sometimes in addition to using a hard-threshold value), identify a voxel as part of a given network/component only if that network is the one with the highest loading across all networks for this particular voxel. That explicitly precludes one voxel from being part of two "networks", even if that same voxel truly had high load values for two or more networks. This is, for example, what you would do in CONN's 'ICA networks. Summary' window when using the "ICA parcellation" function. My opinion is that while this is generally useful in the context of functional parcellation, where you want to assign a unique label to each voxel, in the more general context of network identification this may impose a somewhat unnecessary restriction/assumption regarding whether two different networks may or may not share any areas in common.
In any way, just my two cents. Please let me know your thoughts!
Best
Alfonso
Originally posted by Pravesh Parekh:
That's an interesting question. Traditionally, ICA components are hard-thresholded in order to identify those areas most representative of each network (this is the same approach used to define the areas in CONN's 'networks.nii' atlas). But, almost by definition, spatial ICA components are not truly non-overlapping. More specifically, when defining spatially-independent components, the distribution of spatial loadings for two different components must be independent, which, somewhat informally, means that if you know the value at a given voxel of the loading of one component, that knowledge cannot give you any clue about what the loading of a second component will be for that same voxel. Non-overlapping components would fail to be independent, as knowing that a voxel loads heavily in one component (so that, when thresholded, this voxel is identified as part of a given network) would imply that the loading for other components at that same voxel must be low (so that, when thresholded, this same voxel is not identified as part of other networks), contrary to the independence definition.
One possible work-around that is sometimes used when you really want to define truly non-overlapping areas while still using this same underlying ICA framework, is to, instead of using a hard-threshold value to select which areas are associated with each network (or sometimes in addition to using a hard-threshold value), identify a voxel as part of a given network/component only if that network is the one with the highest loading across all networks for this particular voxel. That explicitly precludes one voxel from being part of two "networks", even if that same voxel truly had high load values for two or more networks. This is, for example, what you would do in CONN's 'ICA networks. Summary' window when using the "ICA parcellation" function. My opinion is that while this is generally useful in the context of functional parcellation, where you want to assign a unique label to each voxel, in the more general context of network identification this may impose a somewhat unnecessary restriction/assumption regarding whether two different networks may or may not share any areas in common.
In any way, just my two cents. Please let me know your thoughts!
Best
Alfonso
Originally posted by Pravesh Parekh:
Dear Dr. Alfonso,
I was exploring the networks.nii file that is shipped with Conn and I noticed that there are several voxels which have multiple labels. For example, [-58 -38 34] is labeled in volumes 17, 21, and 29. There are a total of two voxels which are labeled in three volumes and 3789 voxels which are labeled by two different volumes (where volume refers to the volume number in the 4D file: vol1 = DefaultMode.MPFC, and so forth). I have attached a list of all such voxels and the corresponding volume numbers in which they have been labeled.
My concern is that overlapping voxels between ROIs will affect the connectivity between these two regions given the perfectly correlated time series. Perhaps a hard threshold after ICA is required so that these voxels get assigned to either of the overlapping regions? What are your thoughts on this?
Warm Regards
Pravesh
I was exploring the networks.nii file that is shipped with Conn and I noticed that there are several voxels which have multiple labels. For example, [-58 -38 34] is labeled in volumes 17, 21, and 29. There are a total of two voxels which are labeled in three volumes and 3789 voxels which are labeled by two different volumes (where volume refers to the volume number in the 4D file: vol1 = DefaultMode.MPFC, and so forth). I have attached a list of all such voxels and the corresponding volume numbers in which they have been labeled.
My concern is that overlapping voxels between ROIs will affect the connectivity between these two regions given the perfectly correlated time series. Perhaps a hard threshold after ICA is required so that these voxels get assigned to either of the overlapping regions? What are your thoughts on this?
Warm Regards
Pravesh
Aug 22, 2019 05:08 AM | Pravesh Parekh - National Institute of Mental Health and Neurosciences
RE: Overlap within and between networks in networks.nii
Dear Dr. Alfonso,
Thank you for your reply. This makes a lot of sense to me. I now see how the idea of independence would be violated when applying a hard threshold. And indeed, assigning the voxel to a network with the highest loading across would not be the ideal solution except when creating functional parcellation. My concern was in the context of functional parcellation: since networks atlas is already added to the Conn project by default, I assume many users would simply use it for functional connectivity (ROI-to-ROI, for example) and not necessarily going down the ICA road to create their own functional parcellation. In such cases, there would be a bias in the calculation of mean connectivity between regions where there is an overlap of voxels.
Thank you got clearing this up!
Warm Regards
Pravesh
Originally posted by Alfonso Nieto-Castanon:
Thank you for your reply. This makes a lot of sense to me. I now see how the idea of independence would be violated when applying a hard threshold. And indeed, assigning the voxel to a network with the highest loading across would not be the ideal solution except when creating functional parcellation. My concern was in the context of functional parcellation: since networks atlas is already added to the Conn project by default, I assume many users would simply use it for functional connectivity (ROI-to-ROI, for example) and not necessarily going down the ICA road to create their own functional parcellation. In such cases, there would be a bias in the calculation of mean connectivity between regions where there is an overlap of voxels.
Thank you got clearing this up!
Warm Regards
Pravesh
Originally posted by Alfonso Nieto-Castanon:
Dear
Pravesh,
That's an interesting question. Traditionally, ICA components are hard-thresholded in order to identify those areas most representative of each network (this is the same approach used to define the areas in CONN's 'networks.nii' atlas). But, almost by definition, spatial ICA components are not truly non-overlapping. More specifically, when defining spatially-independent components, the distribution of spatial loadings for two different components must be independent, which, somewhat informally, means that if you know the value at a given voxel of the loading of one component, that knowledge cannot give you any clue about what the loading of a second component will be for that same voxel. Non-overlapping components would fail to be independent, as knowing that a voxel loads heavily in one component (so that, when thresholded, this voxel is identified as part of a given network) would imply that the loading for other components at that same voxel must be low (so that, when thresholded, this same voxel is not identified as part of other networks), contrary to the independence definition.
One possible work-around that is sometimes used when you really want to define truly non-overlapping areas while still using this same underlying ICA framework, is to, instead of using a hard-threshold value to select which areas are associated with each network (or sometimes in addition to using a hard-threshold value), identify a voxel as part of a given network/component only if that network is the one with the highest loading across all networks for this particular voxel. That explicitly precludes one voxel from being part of two "networks", even if that same voxel truly had high load values for two or more networks. This is, for example, what you would do in CONN's 'ICA networks. Summary' window when using the "ICA parcellation" function. My opinion is that while this is generally useful in the context of functional parcellation, where you want to assign a unique label to each voxel, in the more general context of network identification this may impose a somewhat unnecessary restriction/assumption regarding whether two different networks may or may not share any areas in common.
In any way, just my two cents. Please let me know your thoughts!
Best
Alfonso
Originally posted by Pravesh Parekh:
That's an interesting question. Traditionally, ICA components are hard-thresholded in order to identify those areas most representative of each network (this is the same approach used to define the areas in CONN's 'networks.nii' atlas). But, almost by definition, spatial ICA components are not truly non-overlapping. More specifically, when defining spatially-independent components, the distribution of spatial loadings for two different components must be independent, which, somewhat informally, means that if you know the value at a given voxel of the loading of one component, that knowledge cannot give you any clue about what the loading of a second component will be for that same voxel. Non-overlapping components would fail to be independent, as knowing that a voxel loads heavily in one component (so that, when thresholded, this voxel is identified as part of a given network) would imply that the loading for other components at that same voxel must be low (so that, when thresholded, this same voxel is not identified as part of other networks), contrary to the independence definition.
One possible work-around that is sometimes used when you really want to define truly non-overlapping areas while still using this same underlying ICA framework, is to, instead of using a hard-threshold value to select which areas are associated with each network (or sometimes in addition to using a hard-threshold value), identify a voxel as part of a given network/component only if that network is the one with the highest loading across all networks for this particular voxel. That explicitly precludes one voxel from being part of two "networks", even if that same voxel truly had high load values for two or more networks. This is, for example, what you would do in CONN's 'ICA networks. Summary' window when using the "ICA parcellation" function. My opinion is that while this is generally useful in the context of functional parcellation, where you want to assign a unique label to each voxel, in the more general context of network identification this may impose a somewhat unnecessary restriction/assumption regarding whether two different networks may or may not share any areas in common.
In any way, just my two cents. Please let me know your thoughts!
Best
Alfonso
Originally posted by Pravesh Parekh:
Dear Dr. Alfonso,
I was exploring the networks.nii file that is shipped with Conn and I noticed that there are several voxels which have multiple labels. For example, [-58 -38 34] is labeled in volumes 17, 21, and 29. There are a total of two voxels which are labeled in three volumes and 3789 voxels which are labeled by two different volumes (where volume refers to the volume number in the 4D file: vol1 = DefaultMode.MPFC, and so forth). I have attached a list of all such voxels and the corresponding volume numbers in which they have been labeled.
My concern is that overlapping voxels between ROIs will affect the connectivity between these two regions given the perfectly correlated time series. Perhaps a hard threshold after ICA is required so that these voxels get assigned to either of the overlapping regions? What are your thoughts on this?
Warm Regards
Pravesh
I was exploring the networks.nii file that is shipped with Conn and I noticed that there are several voxels which have multiple labels. For example, [-58 -38 34] is labeled in volumes 17, 21, and 29. There are a total of two voxels which are labeled in three volumes and 3789 voxels which are labeled by two different volumes (where volume refers to the volume number in the 4D file: vol1 = DefaultMode.MPFC, and so forth). I have attached a list of all such voxels and the corresponding volume numbers in which they have been labeled.
My concern is that overlapping voxels between ROIs will affect the connectivity between these two regions given the perfectly correlated time series. Perhaps a hard threshold after ICA is required so that these voxels get assigned to either of the overlapping regions? What are your thoughts on this?
Warm Regards
Pravesh