Robust radial deformation model to simulate overall displacement of tissue in the brain caused by tumor growth or treatment. Check out this interactive example using the radial deformation model with grid search to approximate tissue displacement in longitudinal MRI of glioblastoma.
Latest version: 2
Fig. 1: Description and release schedule. Outlined tumor masks are shown for two time points (vertical). cancer-sim v1 (A) and v2 (A, B) are robust radial growth models that, when tuned, produces realistic-looking second time-point MRI examinations of either pushing tumor growth (v1, v2) or shrinking tumor from successful treatment (v2). This is accomplished by deforming first time-pont MRI using a displacement field produced by cancer-sim and linear interpolation. Version three (not yet available) builds upon the ideas of previous versions and produces more realistic displacement fields by using second time-point MRI and gradient-based optimization similar to non-rigid registration.
The lesionmask and brainmask input nifti files need to be stored in LPI voxel order, see 1 and 2.
- To create a single displacement field based on a brain and lesion mask, follow printed instructions from
python3 cancer-displacement.py --help - To find best fit model parameters on longitudinal data, follow instructions from the grid search repository
- Maximum tissue displacement [mm]: The largest tissue displacement produced, which is the scaled magnitude of vectors normal to the ellipsis in Fig. 1 A and B.
- Infiltration [0-1]: The extent of brain coverage, or smoothess of the displacement field in and outside of tumoral regions.
- Irregularity <0,1]: The granularity of Perlin noise added to displacements to simulate irregularity of tumoral displacements.
Version 1 and 2 can be used to deform MRIs in tumor regions mimicking pathology or treatment changes, and thereby create synthetic second time-point MRIs with associated ground truth displacement fields. This data can be used to measure how well a non-rigid registration method produces the simulated ground truth displacement field.
Fig. 2: Comparing the displacement field from the radial growth model (v1) with the estimated field from ANTs SyN (on the post-contrast T1-weighted MRI pair) reveals inconsistencies in displacement estimation in regions with poor textural features (such as necrosis).
Having a pair of structural MRIs and lesion mask, describe the structural change according to the three parameters, by using a grid search extension to fit the cancer-sim (v1 or v2) model.