DD3D-EIT

TwoThree networks are trained: a VAE $\mathbb{VAE}$, a mapper $\Xi$, and a material classifier $\Upsilon$.

The final architecture of the reconstruction network is defined by

$$ \Gamma := \Xi \circ \Psi : \mathbf{u} \mapsto \mathbf{z} \mapsto \hat{\gamma} $$

in parallel with the material classification network

$$ \Upsilon : \mathbf{u} \mapsto m $$

Here, $\mathbf{u}$ represents the EIT data, and $\hat{\gamma}$ is the reconstructed conductivity in a three-dimensional domain by the final reconstruction network architecture.

Hyperparametertuning $\beta$-VAE

Finally, model iteration 21 was selected (also marked with a green dashed line).

Hyperparametertuning Mapper $\Xi$

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

Final reconstruction network architecture results

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 $\gamma$. The lower row represents the predictions of the reconstruction model $\hat{\gamma}$.

Environment

To install the used Python (3.11.2) environment, use

conda env create -f environment.yml