Table 2 Unified benchmarking of all methods using the accelerated approach developed as part of this work

From: Multi-parameter molecular MRI quantification using physics-informed self-supervised learning

Stage

Method

Autodiff-based (AD) model fitting (VBMFa)

Dictionary-based pattern matching

NN-based mapping (supervised learning)

Self-supervised NN + AD (NBMFa)

Preparation + 1st subject mapping

3 min

25 sb (generation) + 93 s (matching)

25 s (generation) + 58 s (training)

3 min (fit & train)

Nth subject mapping (N > 1)

3 min

93 s (matching)

1 s (NN inference)

1 s (inference)

Consistency with raw per-subject data

✓(up to convergence)

✓(up to grid resolution)

×(empirically, in our test)

✓(up to prior)

Implicit NN-based smoothing prior

× 

× 

✓

✓

  1. It utilizes a GPU-based JAX formulation of the Bloch McConnell ODEs numerical/analytical solutions.
  2. The presented times refer to a whole brain reconstruction on a single GPU-equipped desktop machine. The comparison was based on a semisolid MT MRF protocol using a dictionary consisting of 400K entries and a whole-brain quantification task (194K voxels), see additional details in Supplementary Note 2.
  3. Additional analysis and detailed discussion of the compared qualitative features are available in Supplementary Notes 1–3.
  4. NBMF neural Bloch McConnell fitting, VBMF voxelwise Bloch McConnell fitting.
  5. aMethods introduced in this work are highlighted in bold.
  6. bExhibiting linear scaling: it took 80 min to generate the 79M entries of the CEST and semisolid MT cartesian grid dictionary used in ref. 12.
Back to article page