Presented at MICCAI 2016, Athens, Greece.
The provided code recreates the phantom experiment that is presented in our MICCAI 2016 paper
Simon Pezold, Antal Horváth, Ketut Fundana, Charidimos Tsagkas, Michaela Andělová, Katrin Weier, Michael Amann, and Philippe C. Cattin: Automatic, Robust, and Globally Optimal Segmentation of Tubular Structures.
The camera-ready version of the paper can be found as pdfs/pezold2016.pdf.
In the experiment, we compare isotropic total variation (TV) with anisotropic
total variation regularization (ATV) for tubular structure segmentation. For a
fair comparison, we make a grid search over the two parameters of the
nonterminal cost/capacity, which we modeled as an edge detector term (see
paper for equation). Running the full experiment thus might take some hours.
Note that since running the experiment for the actual publication, we widened
the search range for the parameters a bit, which slightly increases the Dice
coefficients for TV (TV dice at level 1.5 before widening search range: 0.8837,
after: 0.8895, ATV dice at level 1.5: 0.9332). The original search range can
still be found as a comment in phantom_experiment.py.
The experiment can be launched via
python src/phantom_experiment.py
using Python 2.X. For speeding up the result presentation, the following lines
in phantom_experiment.py should be commented out:
self.nt_weightings = np.logspace(-2.5, 0.5, 24 + 1)
self.nt_decays = np.logspace(-1, 3, 16 + 1)
self.noise_levels = np.linspace(0.1, 1.5, 15)They should be replaced by:
self.nt_weightings = [10 ** -1.625, 10 ** -1.25]
self.nt_decays = [10 ** 1, 10 ** 1.5]
self.noise_levels = [1.5]which will result in calculating the best combinations for the highest noise level only.
Currently, the provided code runs on Python 2.X only. Apart from numpy and
scipy, it requires the following packages:
mayavipyopencl
Note that pyopencl needs OpenCL-capable hardware and drivers. I only
successfully ran the code on Nvidia GPUs so far. Compiling the .cl files for
use with an Intel driver caused me problems with includes that were not found.
Any help or advice is greatly appreciated.
src/helpers/helix_phantom.py: create the noise-free phantom image volumessrc/processing/qfrangi_gpu.py: calculate Frangi's vesselness feature on the GPUsrc/processing/adjust_directions.py: adjust the vesselness main directions so that neighboring vectors point in approximately the same rather than the opposite direction, as described in Section 2.3 of our paper (mostly on the GPU).src/processing/gradient_vector_flow.py: calculate gradient vector flow on the GPUsrc/processing/continuous_cut_anisotropic_gpu.py: actual segmentation code, running on the GPU