Enhancement of elemental imaging using ResNet34 based UNet with perceptual loss


Veronika Dillingerová^1, Johannes T. van Elteren^2, Vid S. Šelih^2, Martin Šala^2, Dino Metarapi^2

^1Department of Chemistry, Masaryk University in Brno,

Kamenice 753/5, Brno, 625 00, Czech Republic

^2Department of Analytical Chemistry, National Institute of Chemistry,

Hajdrihova 19, Ljubljana, Slovenia


                                         veronika@dilli.sk



Elemental imaging is visualization of the distribution of elements of interest within solid
samples. It is widely used across different fields analysing various materials by many analytical
techniques. With destructive methods, such as LA-ICP-MS or LIBS, where the sample is ablated by
laser beam, spot, by spot, we get pixelated image where the resolution is dependent on the beam
size. Decreasing the beam size infinitely is not possible, which leads us to image enhancement by
computers.

This research is focused on using advanced neural networks for image improvement, with aim on noise
removal and super resolution. Several state of the art architectures were used on different
datasets and compared. The training set consisted from ground truth images – the original pictures,
and their second version created by virtual ablation, which reduces their quality. Modelling was
performed using five consecutive matrix manipulation steps [1]: laser sampling of the sample in the
ablation cell; introduction of Flicker noise (mostly associated with pulse reproducibility);
aerosol washout from the ablation cell and transfer to the detector; measurement by integrating the
counts for a chosen dwell time; introduction of Poisson noise (as a result of counting statistics).

From all tested architectures UNet [2] had the best performance. This deep learning model was build
using the fast.ai library. It consists of UNet, where the encoder was created from a pretrained
ResNet34 and Perceptual loss combined with a Least Absolute Deviations and gramian function.
Enhanced images were compared to the original and ablated version, using 2 metrics: structural
similarity index (SSIM) and peak signal-to-noise ratio (PSNR). The results show a promising way in
improving imaging, which can lead to cheaper and faster analysis.




1.       van Elteren, J.T., Šelih, V.S. and Šala, M., 2019. Insights into the selection of 2D
LA-ICP-MS (multi) elemental mapping conditions. Journal of Analytical Atomic Spectrometry, 34(9),
pp.1919-1931.

2.       Ronneberger, O., Fischer, P. and Brox, T., 2015, October. U-net: Convolutional networks
for biomedical image segmentation. In International Conference on Medical image computing and
computer-assisted intervention (pp. 234-241). Springer, Cham.