InterPore2026

Fast-to-Long Acquisition Projection Learning for Denoising X-ray Microtomography

22 May 2026, 14:30

15m

Oral Presentation (MS15) Machine Learning in Porous Media MS15

Speaker

Luan Vieira (Universidade Federal do Rio de Janeiro)

Description

The need to increase experimental throughput and support time-resolved imaging of dynamic laboratory experiments motivates the reduction of acquisition time in X-ray microtomography. However, faster acquisitions inevitably lead to lower signal-to-noise ratios, since fewer photons contribute to each projection, resulting in reconstructions with increased noise levels and degraded structural definition. This limitation can be mitigated using deep learning methods trained on paired acquisitions of the same sample obtained under different exposure times.

In the acquisition protocol adopted here, scans of 2 minutes and 35 seconds (fast) and 60 minutes (long) were performed sequentially on each rock plug without removing or repositioning the sample in the scanner, ensuring spatial alignment between acquisitions. In this setting, the fast scan provides a noisy representation of the sample, while the long scan serves as the target image, forming well-defined input–target pairs for supervised learning in which fast acquisitions encode acquisition-related noise and artifacts and long acquisitions define the desired reconstruction quality. Microtomography data for both exposure times were acquired using a VTomex M system (Baker Hughes). The fast acquisition employed timing = 50, average = 1, and skip = 0, whereas the long reference acquisition applied timing = 100, average = 40, and skip = 1. In both time configurations, the number of two-dimensional projections was kept constant to enable paired datasets. Because the number of projections is a key factor for reconstructed volume quality, higher values are desirable. To achieve a fixed total of 801 projections under the fast setting, the acquisition parameters were adjusted. The fast acquisition employed timing = 50, average = 1, and skip = 0, whereas the long reference acquisition applied timing = 100, average = 40, and skip = 1. The X-ray source operated with energies between 140–150 keV and tube currents in the range of 220–250 μA.

Based on these paired datasets, a supervised machine learning approach was applied to a set of 12 Brazilian carbonate plug samples. The model was trained to map the two-dimensional projections from the fast acquisition to the corresponding projections from the long acquisition. Operating directly in the projection domain is advantageous since it avoid compounding artifacts introduced in the reconstruction step, as our goal is to reduce acquisition noise. To assess generalization, a leave-one-out validation strategy was adopted. In each iteration, projections slices from 11 samples were used for training, while no slice from the remaining sample was included in the training set. The held-out sample, unseen during training, was reserved exclusively for evaluation. This process was repeated until all samples had been used once as test cases.

Model performance was evaluated on the reconstructed volume obtained from the network-generated projections. Reconstruction used the same acquisition parameters as the fast scans. The leave-one-out validation strategy captured variability across samples, reflecting the heterogeneity of carbonate rocks. Quantitative signal-to-noise metrics showed consistent improvements over fast acquisitions, with reconstructed volumes closer to the long-exposure reference and exhibiting a more concentrated grayscale range, although some smoothing and blurring were observed.