A sharper MRI from a faster scan

Setup MRIs are expensive and slow - can we speed them up?

Pros Beats today's standard fast method outright; close competition to the best AI method (E2E-VarNet) at several times the speed.

Usage Take a look at these beautiful brains - you can see them!

What you (may not want to) see:

FELA reconstructs a clear brain image from a faster scan. It beats today's standard reconstruction outright, and it stays close to the best AI method while running several times faster and on ordinary hardware. MRI scanners use a frequency native coordinate space called k-space. Our FNO models excel here, and are native in this kind of problem domain - where we can produce high quality MRI scans on 6x fewer k-space samples. That means a faster, cheaper, less coolant and energy intensive scan!

Vs the quick pass standard method

FELA wins outright. From a 6×-faster scan with fewer measurements, FELA produces a sharper image than the conventional reconstruction gets from a slower 4× scan. A clear gain over what hospitals use now.

Vs the best AI method (E2E-VarNet)

Where it runs

Light enough to run on the scanner's own reconstruction computer, on a plain CPU, with no dedicated GPU and no cloud - and no potential for HIPAA problems! The reconstruction can happen right on the machine, so the scan never leaves the hospital. The leading AI method (E2E-VarNet) is heavier and needs a dedicated GPU, so it is typically run on separate or cloud compute.

Why it matters and who buys

Scanner time is the bottleneck in radiology: the machines are expensive, waitlists are long, and lying still in the bore is hard for patients. A scan that finishes faster but still reads clearly lets a hospital serve more patients per day at a lower cost per scan. FELA also runs on ordinary hardware, including a plain CPU, so it fits existing workflows without a dedicated accelerator, getting results in the hands of doctors and patients faster!

Full card on Hugging Face ↗

Interactive demo

→ Pick a slice, then compare the four reconstruction methods side by side

Sharper is better. The reference is the fully sampled scan; the conventional column is the standard non-AI reconstruction; FELA and E2E-VarNet are the two AI methods. These are real reconstructions of the same 4×-accelerated slice. (These are precomputed!)

Live compute

Headline · measured

FELA at a 6× faster scan is sharper than the conventional method at a slower 4× scan.

Fewer measurements means a shorter scan. A shorter scan that still reads clearly means more patients per day, lower cost per scan, and a more comfortable visit.

The numbers, on the same data · measured

Scan	Conventional	FELA	E2E-VarNet

Speed and hardware · measured

Recorded once on a B200 GPU (and on CPU for FELA), not measured live in this page.