[acerjio]

For MRI access to the raw data alone is probably not that helpful, with the exception of some of the neural network based upsampling approaches. The real power from working with k-space comes from developing custom pulse sequences.

The Cartesian (rectilinear) approaches to sampling k-space are generally very inefficient acquisition time-wise, but very tractable for a cheap computer to reconstruct via inverse FFTs. Non-Cartesian readouts (such as radial, spiral, etc) can radically reduce acquisition time (read: time for patients to lie in the scanner) but radically increase the computational complexity of the reconstruction. Estimates range from 1000x to 10000x times as complex if you use all the fancy stuff to maximize image quality and reduce acquisition time as much as possible.

There are numerous (solvable) technical challenges to implementing these strategies, but the biggest is financial. If one MRI scanner can do the work of 3 or 4 MRI scanners by moving patients through faster, that's a lot of lost revenue for manufacturers, so there's not much of an incentive to change the current paradigm.

[theglocksaint]

Non-Cartesian sampling of k-space is routinely used today in clinical practice and the reconstruction time or resources isn't a limiting factor. It is the reduced SNR among other effects that is limiting in these approaches.

If one MRI scanner can be more efficient, there is enough competition in the market for it to be wildly successful. The major vendors have enough technical parity for efficiency to be a key difference maker. Hospitals and health systems are the customer and they need throughput.