Data-driven inference of slice ordering / timing (resting state fMRI)

This notebook implements a PoC that verify whether the slice ordering / timing in use make sense data-wise.

It is based on the idea that if the slice timing / order is the right one, then the local signal coherency should be maximal (following the slice timing correction idea). To measure local coherency, one can use the regional homogeneity (ReHo). Here this was tested with local NN filter of 27 neighbours or, to emphasise out-of-slice contributions a 18 option (mask to measure only correlation with neighours on adjacent slices) where a reasonable correlation should be restored by correct slice timing (at least in functionally homogenous ROI).

This migth be useful when metadata (e.g dicom header tags) are not encoding this info reliably and is not immediately clear from the signals upon visual inspection. It allows for consistent automation of the ordering inference to also check whether a dataset made use of uniform slice ordering in all scans.

All six canonical order are being applied to each rs-fMRI dataset with minimal pre-processing. Once cleaned, the ReHo is extracted from each dataset and its average value should be highest when using the correct slice ordering -> using correct pattern should recover best local signal coherence.

So even if the slice timing used for scanning is unclear or not reported (e.g. Philips or old GE) with this you might be able to infer the most likely pattern alternatives used in a dataset.

Absolutely no warranty on functionality, tested on freely available datasets from OpenfMRI project (http://www.openfmri.org) on ds000133, ds000172, ds000201, ds000220, ds000245, ds000256. The approach is functional in all tested cases.

Note:

1 - it was found to always work to automatically detect whether interleaved rather than sequential was used

2 - it detects mixed cases of patterns within same dataset

3 - inference of direction (ascending / descending) less reliable if data is noisy or with lots of motion ( -5 helps)

4 - inference of correct slice order from short TR scans works for main pattern but direction less clear (TR < 1.5 sec)

... might be not relevant correction anyway ?

5 - improving the stats might be less effective than using context data (ie. what is the scanner/model/slices/...)

PS: you need to have fsl, nipype and so on available locally !