Tckdfc overlap rate

Dear MrTrix experts,

I was experimenting with the track-weighted dynamic functional connectivity method (twdFC) and I’m trying to figure out how the dynamic functional connectivity information is sampled from the BOLD data. I work with HCP data (1200 time points) using a rectangular time window of ~40 seconds length (55 time points). Tractograms (1M streamlines) were coregistered to MNI space and tw-dFC was carried out at original spatial resolution of BOLD data (2mm). Here is the example command:

tckdfc -dynamic rectangle 55 -template MNI152_T1_2mm_brain.nii.gz -vox 2 -backtrack my_input_tractogram_1M.tck my_input_BOLD_data.nii.gz my_output_tckdfc_data.nii.gz

I see that my_output_tckdfc_data.nii.gz dimensions are 91x109x91x1200 i.e. it has the same number of time points as the input BOLD data. I got from the reference paper (Calamante et al., 2017, Brain Structure and Function) that each time point in the output volume represents tw-dFC signal sampled from a given time window centered on the original BOLD data time points, and that the time window was incremented by one BOLD volume (i.e., 1 TR) for each step, so I assume we are talking about partially overlapping time windows. My questions are:

  1. How do I get the same number of time points between input and output data? I mean, should I assume that the first and last time windows are somewhat smaller than the other ones, as they are centered on the first (or last) volumes respectively, and they have no preceding or following time points from which FC is calculated?
  2. What’s the resulting overlap rate between time windows? Am I right at assuming it should be around 0.98 (i.e., 54/55)? I didn’t found such information on the reference paper.
  3. If I were interested in selecting only non-overlapping time windows, would it be appropriate to sub-sample the output tckdfc (e.g. by selecting only time windows nr. 27, 82, 137 and so on)? Is this an appropriate sub-sampling scheme?
  4. Is there something I figured out wrong?
    Thanks in advance for your attention!

Hi Gianp,

  1. The output image should have the same number of time points as the input image. For time points near the start and end of the acquisition, the sliding window will partially extend beyond the available image data, and these time points obviously do not contribute to the calculation, so the correlations within those time windows will be computed from fewer time points.

  2. The overlap between adjacent time windows can be calculated based on the fact that there is one output volume per input volume and the shape of the time window. So yes, for a rectangular window it would be (N-1)/N where N is the number of time points in the window. For non-rectangular time windows the “overlap” is less trivially defined.

  3. If you wanted non-overlapping time windows, then yes, you would need to select a subset of output volumes that achieves that. It does however demonstrate the somewhat arbitrary nature of selection of the offset by which you make that selection; e.g. starting from volume 27 might be optimal at the start of the sequence given the first sliding window will be fully filled, but may result in a poorly-filled sliding window at the end of the sequence.


1 Like

Thank you very much!!