Skip to main content

Reconstruction strategies

All strategies implement the same ReconStrategy trait, take a calibrated k-space tensor [ch, kz, ky, kx], and return image- domain volumes. 3D acquisitions with ky-only undersampling decouple along kz via a 1-D IFFT and then run independent per-slice unfolds, so parallel imaging is supported for both 2D multi-slice and 3D Cartesian data.

ifft-rss

Centred IFFT plus root-sum-of-squares coil combination. Fastest strategy; produces magnitude images only. Recommended baseline.

grappa

k-space GRAPPA (Griswold 2002). ACS lines are extracted automatically; a convolution kernel is fit via ridge regression and applied to fill the missing ky lines.

FlagDefaultDescription
--grappa-kernel-ky <k>4Kernel size along ky
--grappa-kernel-kx <k>5Kernel size along kx
--grappa-ridge <lam>1e-3Tikhonov ridge for kernel fit

sense

Image-domain SENSE unfold (Pruessmann 1999) with ridge stabilisation. Sensitivity maps are estimated via either Walsh (eigenvector method, Walsh 2000) or ESPIRiT (auto-calibrating, Uecker 2014).

FlagDefaultDescription
--sense-maps walsh|espiritwalshCoil sensitivity estimation
--sense-walsh-window <w>7Walsh smoothing window
--sense-walsh-iters <n>3Walsh power-method iterations
--espirit-kernel <k>6ESPIRiT kernel size
--espirit-threshold <f>0.02ESPIRiT singular value threshold
--espirit-iters <n>50ESPIRiT power iterations
--sense-ridge <lam>1e-4Tikhonov ridge for unfold
--sense-gfactoroffCompute g-factor map
--write-gfactoroffAlso emit g-factor PNGs

cs

L1-wavelet compressed sensing via FISTA (Lustig 2007; Beck & Teboulle 2009). Slower than GRAPPA/SENSE but better at high acceleration with random ky undersampling.

FlagDefaultDescription
--cs-iters <n>80FISTA iterations
--cs-lambda <lam>1e-3L1 regularisation strength
--cs-wavelet-levels <n>4Wavelet decomposition levels