Allow DKI to be used in place of DTI?

[arokem]

Looks like we're allowing DTI and FWDTI here:

mriqc/mriqc/interfaces/diffusion.py

Line 462 in 61b59e0

DTIModel = FreeWaterTensorModel if self.inputs.free_water_model else TensorModel

DKI can also be used to calculate FA, and that might have advantages in some cases (generally more reliable, doesn't require excluding higher b-values ... to a point).

[oesteban]

Would DKI be substantially slower?

Regarding reliability, it depends on your definition:

• If you mean robustness to data variability (including satellite operations such as filtering out b-values and their side effects), then yes, we should go for reliability.
• If you mean DKI gives "better" FA maps, then we actually may want to use DTI because those "bad" voxels can tell you something is not great at those points in the image.

I think we could actually have both. Since DTI is only used with the lower b-values, we are basically not looking into the high values' problems. It could make sense to use DTI for low b-values and also DKI for high b-values.

wdyt?

[arokem]

DKI is not substantially slower than DTI, because it's also solved as a linear set of equations.

I agree that it would be good to know about voxels with aphysical signals (which are a good canary in the coal mine), but I think that DKI would also report about those. The reliability I am referring to is consistency across different sets of b-values. As in this paper (Figure 9).

It would be interesting to produce both and see whether that's not overwhelming. Either way, I think that it would be easy to implement, but would require a slight adaptation of the API, because we'd want self.inputs.free_water_model to allow selection between the three different models.