Speaker
Description
Industrial computed tomography allows for non-destructive inspection of complex components,
making it an interesting domain of applications for machine learning methods.
In this context, a large bottleneck for training robust models is data acquisition.
Already time-intensive task of obtaining data using real-world CT scanners
is further exacerbated by the requirement of creating accurate labels.
We solve this problem by implementing a pipeline for the generation of massive
synthetic datasets, and using only some examples of real-world data.
Our data pipeline uses parametrized CAD geometry to generate thousands of
randomized objects from which we derive both ground-truth labels and
attenuation data further fed into tomographic simulations.
The simulation code is GPU-accelerated allowing us to substantially
reduce the time needed to generate the datasets.
To reproduce effects visible in real data, we implement a polychromatic beam,
and simulate artefacts such as non-uniformity of the beam and detector matrix response.
We also apply custom beam hardening corrections for our reconstructions.
We use the synthetic data in semi-supervised training of a W-net model for the specific
task of segmentation of ionization chambers. We show, that semi-supervised approach
allows for effective training with only partial labeling (as low as 25%-50% of the data).
We then validate performance of segmentation models trained using simulations on the real-world data.
