papago

Patch based gaussian mixture modelling of medical imaging

Pre-processing

• First, we need data to be aligned, and we need the masks of those alignments (perhaps even the interp-matrices?)

Execution

Data Preparation

Preprocess subvolumes, e.g. adniprep.m, get md struct. This involves building the medialDataset structure via restorationmd and processing the images via processmd.

Training On Cluster

1. 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.
2. Run wgmm on cluster distributed on each subvolumes, e.g. via sgeTrain.sh. This can be done via model0 (isotropic data) of model3 (weighted data).

Testing On Cluster

1. run mccRecon.m (papago.recon) on all patches on each subvolume, and store the reconstructions!
2. (unfinished) re-compose volume.

Evaluation for optimal K at each location (On Cluster)

Loop steps for Training and Testing for various K. Choose K based on best patch reconstruction at each location.

Training and Testing On a Single Machine

This is usually done on a subset of the image grid.
Loop over (sub)grid:

1. create/load subvolume column.
2. run wgmm via papago.train
3. run papago.recon to reconstruct all the patches in this subvolume (perhaps for just a subset of subjects)

Quilt patches.

Papers

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.