How feasible to reconstruct MRI images from fMRI images

I am aiming to design a deep learning model which can reconstruct MRI from fMRI of the human brain. I have gone through a series of literature about MRI and fMRI but have not found any literature where they insight the importance of having a model that can retrieve MRI from fMRI or vice versa.

I am looking for related experts to have their thoughts in this regard. My concerns to experts are:
→ How often do medical cases require both MRI and fMRI?
→ How likely is it that MRI records can be retrieved from fMRI?

1 Like

I am assuming by MRI you mean structural MRI?

Not often. fMRI in hospitals are mostly acquired only in research settings and not part of standards of care. An exception is being made for disorders of consciousness, where sometimes an fMRI could be part of diagnosis.

Not very likely. Anatomical images are made at higher resolutions than fmri, and the weighting in fmri does not provide a good tissue contrast. Both of these will create a noise ceiling that will hamper the ability to make a good reconstruction. But anatomical images are almost always collected in fMRI studies. What is your motivation for wanting to reconstruct T1 images?


1 Like

Dear Steven,

Thank you for your prompt response to my queries.

The motivation of reconstructing structural MRI from fMRI using ML model is to-

  • save resources used for MRI scan
  • avoid extra burden of patients going through two scanning processess
  • get rid of having different artifacts and noises between structural MRI and fMRI for the same subject (patient)

Thank you,

Hi @Harikala_Kandel ,

I’d agree with @Steven that it’s rare for current standard of care to require both an fMRI and an MRI scan. I’ve seen more work on this idea of estimating new scan types in quantitative MRI, like in this paper:

Multiparametric mapping in the brain from conventional contrast-weighted images using deep learning - PubMed



1 Like

I agree with @Steven too. A good quality MRI (assuming you mean structural/anatomical MRI) is much easier and faster than acquiring functional MRI data (well not technically faster per brain image because fMRI readouts has to be extremely fast to keep making whole images every few seconds).

One key difference is that fMRI requires you to sample the k-space extremely fast so that e.g. you can create a whole brain volume every few seconds. This comes at the cost of acquiring your signal while it is decaying over a longer time window, getting affect by many non-linear spatial artifacts (spatial distortions, signal dropouts…). Whereas in anatomical imaging, you collect pieces of the k-space over a longer period of time (minutes) to construct a single 3D whole brain volume. Which is in a simplified way less effected by the spatial artifacts.

For further details, you can refer to: Handbook of MRI Pulse Sequences | ScienceDirect

  • Chapter 14.1 Gradient Echo (related to MRI in your description), and
  • Chapter 16.1 Echo Planar Imaging (related to fMRI in your description).

So, in a way, @Harikala_Kandel maybe you can consider looking at the differences between different readout methods. If you can make the EPI readouts free of their spatial artifacts while somehow increasing their resolution, this might be useful to replace some anatomical MRI pulse sequences with EPI readouts, making them really fast. However, I would advise come caution because the MRI pulse sequence development field is doing these types of research for so many years and I am pretty sure there are plenty of papers on similar issues and techniques.