Deep learning for asbestos counting.

Area of Science:

• Occupational Health and Safety
• Microscopy
• Artificial Intelligence

Background:

• Traditional phase contrast microscopy (PCM) for asbestos counting is time-consuming and expensive due to required sample treatments.
• There is a need for more efficient and cost-effective methods for airborne asbestos detection.

Purpose of the Study:

• To implement and evaluate a deep learning (DL) procedure as an alternative to PCM for asbestos fiber counting.
• To assess the reliability and competence of DL analysis on untreated airborne samples.

Main Methods:

• Deep learning models were trained on images of airborne asbestos samples (chrysotile and crocidolite) collected using standard Mixed Cellulose Ester (MCE) filters.
• A database of 140 real and 13 artificial images was created, with approximately 7500 fibers manually annotated according to NIOSH Method 7400.
• A 20x objective lens and backlight illumination were used for image acquisition.

Main Results:

• The best-trained DL model achieved a precision of 0.84 and an F1-Score of 0.77 at a confidence level of 0.64.
• Post-detection refinement, excluding fibers < 5 µm, further enhanced the precision.
• The DL method demonstrated reliable performance in counting asbestos fibers.

Conclusions:

• Deep learning analysis of directly acquired images presents a viable and efficient alternative to conventional PCM for asbestos fiber counting.
• This AI-driven approach reduces sample preparation time and costs associated with asbestos monitoring.