Deep learning for asbestos counting

Abstract

The PCM (phase contrast microscopy) method for asbestos counting needs special sample treatments, hence it is time consuming and rather expensive. As an alternative, we implemented a deep learning procedure on images directly acquired from the untreated airborne samples using standard Mixed Cellulose Ester (MCE) filters. Several samples with a mix of chrysotile and crocidolite with different concentration loads have been prepared. Using a 20x objective lens coupled with a backlight illumination system a number of 140 images were collected from these samples, which along with additional 13 highly fibre loaded artificial images constituted the database. About 7500 fibres were manually recognised and annotated following the National Institute for Occupational Safety and Health (NIOSH) fibre counting Method 7400 as input for the training and validation of the model. The best trained model provides a total precision of 0.84 with F1-Score of 0.77 at a confidence of 0.64. A further post-detection refinement to ignore detected fibres < 5 µm in length improves the final precision. This method can be considered as a reliable and competent alternative to conventional PCM.

Keywords: Airborne fibers; Asbestos; Deep learning; Object detection; YOLOv5.