Patch based gaussian mixture modelling of medical imaging
- First, we need data to be aligned, and we need the masks of those alignments (perhaps even the interp-matrices?)
Preprocess subvolumes, e.g. adniprep.m
, get md
struct. This involves building the medialDataset
structure via restorationmd
and processing the images via processmd
.
- Break up dataset into "subvolume columns" via grid, e.g. via
md2subvols.m
, and save subvolume columns. This can take a long amount of time and/or memory. - Run
wgmm
on cluster distributed on each subvolumes, e.g. viasgeTrain.sh
. This can be done viamodel0
(isotropic data) ofmodel3
(weighted data).
- run
mccRecon.m
(papago.recon
) on all patches on each subvolume, and store the reconstructions! - (unfinished) re-compose volume.
Loop steps for Training and Testing for various K. Choose K based on best patch reconstruction at each location.
This is usually done on a subset of the image grid.
Loop over (sub)grid:
- create/load subvolume column.
- run
wgmm
viapapago.train
- run
papago.recon
to reconstruct all the patches in this subvolume (perhaps for just a subset of subjects)
Quilt patches.
If you find this library useful, please cite (download bib):
-
Medical Image Imputation from Image Collections
A.V. Dalca, K.L. Bouman, W.T. Freeman, M.R. Sabuncu, N.S. Rost, P. Golland
IEEE TMI: Transactions on Medical Imaging 38.2 (2019): 504-514. eprint arXiv:1808.05732 -
Population Based Image Imputation
A.V. Dalca, K.L. Bouman, W.T. Freeman, M.R. Sabuncu, N.S. Rost, P. Golland
In Proc. IPMI: International Conference on Information Processing and Medical Imaging. LNCS 10265, pp 1-13. 2017.