Skip to content

README.md

Latest commit

 

History

History
30 lines (27 loc) · 2.38 KB

README.md

File metadata and controls

30 lines (27 loc) · 2.38 KB

Synthetic Patient Simulation Studies

This directory contains scripts required to run simulation studies on a "synthetic" patient. The geometry is taken to be that of a real patient that has been appropriately pre-processed with the pipeline laid out in this code base. In particular, patient 101 from the UPENN-GBM dataset. The steps of the study are as follows,

  1. Generate observational data from a known underlying model and corrupt measurements with Gaussian noise.
  • Data generation is handled by the script generate_synthetic_observations.py.
  1. Solve the deterministic inverse problem for the MAP point reconstruction and generate the Laplace approximation to the posterior.
  • The inverse problem is run by the run_synth_bip.py script.
  1. Draw samples from the prior and posterior, either by sampling the Laplace approximation or via Markov Chain Monte Carlo (MCMC).
  • To draw samples from the prior and the Laplace approximation, please use draw_la_samples.py.
  • To draw samples from the posterior with MCMC, please use draw_mcmc_samples.py.
  1. Compute the pushforward of the samples.
  • To compute the pushfowards, please use run_fwd_prop.py.
  • To concatenate the results from the forward propagation run, use concat_fwd_prop.py.
  1. Compute the quantities of interest for comparison.
  • Various quantities of interest are computed with the compute_qoi.py script.

The study was executed on TACC's frontera supercomputer and the associated jobfiles may be found in the jobs subdirectory. Assuming that the geometry was processed with FreeSurfer as previously described, the pp_VBG.slurm jobfile shows how to run the necessary registration pipeline. The implementation of the above pipeline would require execution of the following scripts:

generate_synth_data.slurm (step 1)
run_synth_rd_bip.slurm (step 2)
run_synth_prop.slurm (steps 3+4)
compute_qoi.slurm (step 5)

Note(s):

  • We utilize an observation operator to convert from state measurements (on the mesh) to the discrete data-space (voxel images)
  • We use a coarser mesh to evolve the dynamics in the inverse problem so as to avoid committing an "inverse crime".
  • To test the BIP for a tumor in a box example, consider run_box_bip.py.
  • To utilize a continuous observation operator, use run_cont_synth_bip.py.
  • To compute the mass-matrix orthogonal eigenvectors (instead of the prior-preconditioned), use mass_orth_misfit.py.