How MRI reconstruction methods compare
Methods from the studies behind Fig. 2 of the paper, joined when a study measured them side by side. Distance from U-Net is not quality: the closer a method, the more firmly the comparisons tie it to U-Net (circles: ± standard error, the same in every window). Click one for details; the notes under the graph explain the controls.
Height: difference from U-Net. Drag to turn.
- θ is the difference from U-Net (fastMRI), in dB for PSNR or logit for SSIM, fitted only on differences within a study, so methods never compared directly still share one scale.
- A window (2025 covers 2024–2026) is fitted on its studies alone and shows every method they measured, older baselines included: a 2019 method appears in 2026 when a 2026 study compared against it.
- Proposed only hides the methods no study proposed. In a window it keeps only those that came out in its years.
- Measured in, on a method's card, lists the studies its θ comes from; a comparator's own paper is usually not one of them.
- Since marks a proposed method's first year in these studies. Comparators get none, as their own paper can be much older than the studies that compared them.
- 3D lifts each method to its θ, so height is the difference from U-Net: drag down to tilt into it, drag to turn, let go near the top view to return to 2D.
- Other, classical and miscellaneous methods, starts hidden; click a family under the graph to show or hide it. In ≥2 studies keeps the methods measured that often.