tomoalign is Python module for multi-GPU reconstruction of tomographic data with non-rigid projection alignment. For details, see Nikitin, Viktor, et al. "Distributed optimization for nonrigid nano-tomography." IEEE Transactions on Computational Imaging 7 (2021): 272-287. (https://ieeexplore.ieee.org/document/9364757)
The module allows for compensating sample drift/deformation during tomographic data acquistion. The proposed method is based on solving a distributed optimization problem invloving tomographic and deformation estimation subproblems. Tomographic subproblems is solved by using the Conjugate gradient method, while the deformation is estimated with using the Farneback algorithm. Both subproblems are coordinated inside an ADMM (Altenrative Directions Method of Multipliers) scheme. Note: the method works better for interlaced scanning protocols (4-8 sample rotations).
Example of reconstruction:
create conda environment and install dependencies:
conda create -n tomoalign -c conda-forge cupy swig scikit-build dxchange tomopy cmake matplotlib conda activate tomoalign conda install -c conda-forge opencv
Note: CUDA drivers need to be installed before installation
install:
python setup.py install
check adjoint tests:
cd tests
Test deformation:
python test_deform.py
sample output:
registration time: 5.561098337173462 apply flow time: 0.01982426643371582 data0-data1=5928.0068359375 data0-data1_unwrap=2606.488037109375 norm flow = 10330.0009765625 <data,D*Ddata>=<Ddata,Ddata>: 1.108183e+08 ? 1.164992e+08
Test tomography:
python test_tomo.py
sample output:
norm data = 21722.1015625 norm object = 4057758.5 <u,R*Ru>=<Ru,Ru>: 4.718072e+08+0.000000e+00j ? 4.718676e+08
- experimental data reconstruction
battery reconstruction:
cd experimental/battery
processing.py - preprocessing
test_center.py - find rotation center
cg.py - convetional reconstruction by CG
admm.py - admm-based reconstruction with optical flow alignment
Viktor Nikitin (vnikitin@anl.gov)