InterPore2026

A machine learning method to automatically segment solid and multiple fluid phases in time-dependent 3D (4D) images

19 May 2026, 14:05

15m

Oral Presentation (MS15) Machine Learning in Porous Media MS15

Speaker

Zhuangzhuang Ma

Description

Capturing dynamic processes like pore-filling and snap-off using fast synchrotron X-ray micro-tomography enables time-resolved quantitative and qualitative analysis. However, time-resolved imaging often generates noisy, low-contrast images, and the resulting datasets are often large. These factors present challenges for effective and accurate 4D image segmentation. Frame-by-frame segmentation methods treat each time step as an independent 3D image without considering temporal consistency, which often results in flickering and physically implausible interface evolution.

To address this, we present Spatio-Temporal SwinUNETR (ST-SwinUNETR), a deep-learning technique that segments 4D images by modelling space and time jointly. We validate the method on dynamic synchrotron micro-CT datasets and evaluate performance using both image-based and physics-based criteria, including porosity and phase saturation over time. ST-SwinUNETR improves spatial accuracy while enhancing temporal consistency of the predicted segmentations over time.