idea: fieldmap correction based on anatomical + functional images

[yarikoptic]

I believe that given scanning sequence metadata of functional image, and anatomy/geometry and thus knowing where air and where brain tissue, there should be a way to guestimate the fieldmap correction without acquiring dedicated fieldmap sequences. Given a vast array of fieldmaps and anat/func pairs shared couldn't we try/test this idea on providing deep network to produce fieldmaps for studies and/or simply just directly correcting functional data.

[satra]

@yarikoptic - this gets tricky mostly because functional scans as often gradient-echo based and have losses that cannot be recovered. hence the fieldmap sequences often use other acquisitions to minimize those losses. some tools have leveraged motion as a way to estimate fieldmaps from functional series, since each movement causes a specific effect. how this plays out across different types of sequences (multiband, in-plane acceleration, phase encoding, etc.,.) also becomes complicated. the SDCflows project (cc:ing @oesteban) has been trying to optimize fieldmap detection given different types of sequences. whether neural networks can estimate fieldmaps well and quantitatively is a bit of an empirical question, but perhaps one that a generative model could be used for. i would hypothesize that we would be able to generate the relative distortion field, the part that may be tricky is getting the quantitative part of this field valid.

[oesteban]

I agree with @yarikoptic that it doesn't seem very hard to be able to predict the relative fieldmap that corresponds to a given anatomy. I wonder how helpful it would be for this prediction to have the face information (perhaps there's some influence of the shape of nasal sinuses on the fieldmap realization).

I agree with @satra that actual correction is a different problem where signal pile-up and drop-out will require "smarter" models that not just resolve the geometrical warping but also correctly resolve the intensity modulation/collapse problem.

However, for the fieldmap estimation to be useful, I would guess calibration of inhomogeneity would be the main pitfall, which falls between both of the above ideas. In other words, to correct for distortion you really need an estimation of the fieldmap in Hz (i.e., it must be a quantitative map). I guess such calibration will not be easy to determine (even conditioning on scanner field-strength or head size, etc.), and the problem becomes somehow dual to directly correcting data. Using functional or diffusion data of the same-subject could be a way of informing the model to quantify the field estimation, but now it is necessary to condition the model on the effective echo-spacing.