Abstract
As nuclear sites age, structures and their materials become a key focus of decommissioning efforts. On such sites, asbestos is particularly concerning as it is a known carcinogen if inhaled. A crucial challenge with asbestos remediation is estimating its concentration in air which ensures safe working conditions. When viewed using phase contrast microscopy (PCM), the risk of human error when performing fibre counting increases due to fibres’ peculiar morphology, bundling tendencies, and their size. A focus of this research surrounds developing automated solutions using machine learning techniques to efficiently count asbestos fibres on filters, for air quality monitoring of sites during decommissioning. The training of models to achieve such solutions require images of samples containing asbestos fibres which, problematically, are seldom available in the public domain. This research presents a novel approach to procedurally generate synthetic data which emulates both amphibole and serpentine asbestos fibres when viewed using PCM. Using this synthetic data, a direct-count regression model, with a feature pyramid network (FPN) as a backbone, is trained which can count the number of fibres in an image to within 5% of the true count. Alongside counts, models produce useful feature maps, indicating a similar counting approach to human analysts.
| Original language | English |
|---|---|
| Publication status | Unpublished - 15 Apr 2026 |
| Event | Postdocs in Nuclear Energy 2026 - University of Strathclyde, Glasgow, United Kingdom Duration: 14 Apr 2026 → 15 Apr 2026 https://www.nuclearinst.com/events/events/postdocs-in-nuclear-energy-pine-2026 |
Conference
| Conference | Postdocs in Nuclear Energy 2026 |
|---|---|
| Abbreviated title | PiNE 2026 |
| Country/Territory | United Kingdom |
| City | Glasgow |
| Period | 14/04/26 → 15/04/26 |
| Internet address |
Keywords
- image processing
- deep learning
- computer vision
- fibre counting
- asbestos
- nuclear decommissioning
Fingerprint
Dive into the research topics of 'Automated asbestos fibre counting: a procedurally generated data-driven approach'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver