-
4.1.0-dev
Changed - Reflect padding is now edge-padding througout - No longer uses OneCycle learning rate scheduling - Trains for a fixed number of iterations (longer than 4.0) - Uses wavelet loss alongside L1 - Contains support for iSMORE (but not recommended atm) - Mathematically correct handling of padding at inference time Added - Includes custom implementation of ESPRESO for robustness, speed, and simplicity
-
v4.0.4 protected
This bugfix release addresses patch extraction for small-FOV volumes by correctly padding images prior to patch extraction.
-
v4.0.3 protected
This bugfix release pins a new `degrade` package version. The updated version fixes a bug which affected images that had in-plane resolution != 1x1mm.
-
v4.0.2 protected
This bugfix changes the tolerance when determining whether an image is "isotropic" in-plane. The tolerance now is 1e-2; this means that an in-plane resolution of 0.745 and 0.756 will be considered the same (as will 0.74 and 0.749).
-
v4.0.1 protected
This is a major overhaul from SMORE3 -> SMORE4. The associated publication is the SASHIMI 2023 paper. Below is a summary of changes: - Now a single model to do SR - Replaced EDSR with WDSR - Now estimate the residual rather than just the SR image - Algorithm to ensure the correct slice number even with pixel shuffle in WDSR - Fully training from scratch: no more pre-trained weights - Support for SLR pulse design
-
-
-
-
v2.1.1 protected Release: v2.1.1
iSMORE with slice selection profile kernel argument support and bug fixes.
-
-
v3.1.0 protected Release: v3.1.0
Feature release for custom kernel, weighted sampling and separate CLI options
-
-
v3.0.1 protected Release: v3.0.1
Hotfix release to update code and README based on runtime debugging