Non-destructive imaging is vital for analyzing particle size, shape, and distribution in mining, materials science, and geology. Yet, segmenting particles in micro-CT data remains difficult due to variations in morphology and contact areas, limiting traditional methods like watershed. Creating annotated training data for machine learning is also labor-intensive, error-prone, and hard to scale. In this talk, I present a self-supervised learning approach that expands the training dataset without manual annotations. Particles are added if they can be reliably matched in at least one resampled scan. After three iterations, the method segments 98% of the particle volume, identifying over 50,000 quartz particles. Beyond training, this method also supports unsupervised evaluation, demonstrated by comparisons with existing instance segmentation techniques.

Dr Philipp Loesel has been a Postdoctoral Research Fellow in the department of Materials Physics at the Research School of Physics since April 2023. Loesel completed his PhD in Mathematics and Computer Science at Heidelberg University, Germany in April 2022 - “GPU-based Methods for Biomedical Image Segmentation”. Prior to this Loesel was a Research Associate at the Interdisciplinary Center for Scientific Computing (IWR), Heidelberg University, Germany from Nov 2014 – Apr 2023. He also lectured at the Mathematics Institute of Pharmacy and Molecular Biotechnology, Heidelberg University from 2013-2015, after completing studies in Mathematics and Political Science, (from Heidelberg University) in 2013.