View symposium recording
Q&A - unanswered questions from the symposium
Dr. Dirk Poot - RIM for quantitative MRI
Is it possible to address B0 (3T vs 1.5 tesla) influence in your qMRI reconstruction?
D.P.: Obviously ‘yes’ if you take the method broadly enough 😊. The real question probably is: what needs to be modified or will be different in the results? This depends on the tissue property changes due to field change. Typically, T1 and T2 values change (1.5T->3T, T1 increase, T2 shortening). If the training distribution is wide enough this change might hardly affect the quality of the results, but (obviously?) reflect the intrinsic value difference between the field strengths. If for some property the value change between field strenghts is so large that at the new field strength it would fall outside of the training distribution of the old field strength, then retraining with the appropriate distribution should be done.
Your recurrent DTI method, can it be used to harmonize DWIs across scanners, sites and scanners?
D.P.: We demonstrated that, because of the forward model, the RIM is not so sensitive to the precise acquisition protocol. Still, if you can influence, I would recommend to harmonize the acquisition protocol as much as possible as typically any model for diffusion MRI is only approximate, so even though the RIM could properly fit the used (DTI) model to the diffusion data, there could still be biases in the estimated parameters due to unmodelled effects.
Dr. Matthan Caan - Generalizable, robust and fast: applications of deep-learning accelerated MRI
In the relaxometry RIM paper the training loss was defined on the relaxometry parameters, wheras on the DTI RIM paper the training loss was defined on the data predicted by the DTI parameters. Why did you change this strategy? Especially since the T2 relaxation and DTI model are very similar in functional form.
M.C.: These papers were written by a different team. Possibly, this was done to account for certain ambiguity in the orientation of the diffusion tensor, requiring careful parametrization in order to be used in a loss function, although not impossible.
D.P: If I recall properly, the first reason to change the metric was because measuring distances between diffusion tensors proved quite hard. Maybe it can be solved, but after a few tries we basically gave this up and computed distances on the image intensities instead. This has as additional benefit that the partial volume effects that were incorporated into the generation of the training data are also be handled gracefully; i.e. put simply we did not really have ground truth parametric maps at the image resolution to compare to.
Is the loss function weighted differently for the time steps/iterations? Would results become worse if you only define your loss on the final time step?
M.C.: We have experiment with this in different papers. A weighing scheme aids in learning the RIM to act as an iterative inverse problem solver. Yes, a positive weight for the first iterations is needed.
I am assuming that RIMs are indeed a discretized ODE with fixed iterations, like a fixed-step Euler solver. Is memory a problem where we need to scale to 3D or 4D dynamic MRI since we need to store every iteration (isn't it?). I can see a dynamic neural ODE might be a good candidate in case of controlling inference speed and reconstruction precision.
M.C.: Yes, memory is of importance. Thank you for suggesting a dynamic neural ODE. We have worked with computed rather than stored hidden states in the iRIM (Putzky).
About the DTI RIM's ability to deal with variable number of directions, I assume each direction set was optimally distributed on the sphere. Did you test it with direction sets that lead to poor conditioning of the DTI problem?
M.C.: This question is best addressed to the authors of this paper.
D.P.: We did not explicitly investigate that in the paper, mainly because we thought it would be rather artificial; assuming any scheme that is going to be used has at least decent conditioning. Probably the RIM would perform worse as ‘poor conditioning’ almost certainly means that the RIM should use a different relation between gradient and step-to-be-taken.