TwoThree networks are trained: a VAE
The final architecture of the reconstruction network is defined by
in parallel with the material classification network
Here,
Finally, model iteration 21 was selected (also marked with a green dashed line).
VAE | Iteration | Predictable (%) | Median volume error (%) | Median position error (%) |
---|---|---|---|---|
21 | 1 | 92.19 | 0.16 | 4.14 |
21 | 2 | 90.94 | 0.16 | 4.31 |
21 | 3 | 92.09 | 0.14 | 4.39 |
21 | 4 | 92.77 | 0.15 | 4.12 |
21 | 5 | 54.65 | -0.10 | 8.73 |
21 | 6 | 92.54 | 0.09 | 4.72 |
21 | 7 | 92.58 | 0.14 | 4.58 |
21 | 8 | 96.01 | 0.13 | 4.55 |
21 | 9 | 90.83 | 0.12 | 4.44 |
Five randomly selected EIT measurements were taken from the test data. The test data was not used throughout the training phases. The presented graph provides a proof of concept and shows the feasibility of reconstructing different objects within a phantom tank using a data-driven reconstruction approach.
The top row is the true conductivity distribution
To install the used Python (3.11.2) environment, use
conda env create -f environment.yml