Machine learning for identification and counting of Naturally Occurring Asbestos

Open-pit nickel mining is the main economic activity in New Caledonia. Lateritic Ni-ore deposits formed on weathered ultrabasic rock cover more than a third of the territory. However, among the mineral phases that make up these laterites, some belong to the asbestos family and have the capacity to emit pathogenic fibres. The inhalation of air polluted by such fibres may lead to severe respiratory diseases; asbestos may penetrate deep into the lungs causing at worst malignant mesothelioma.

In order to manage the natural occurrence of these fibres and take the necessary measures for the protection of workers, it is necessary to evaluate and monitor the concentration of asbestos fibres into the environment (e.g., airborne, waterborne). The current monitoring approach adopted by asbestos laboratories relies on counting method using Transmission Electron Microscopy (TEM), according to French regulation (NF X 43-050). Analysts operatively count and measure fibres and elongated mineral particles (EMPi) with on a filter viewed through the microscope device at high magnification. It is worth noting that analytical procedures involving electron microscopies are time-consuming, and show an intrinsic bias related to the subjectivity of operator analysis. These drawbacks explain the need to develop an automatic method for fibre and EMPi detection and quantification.

This paper presents a new method for detecting fibres on filters by using image processing and machine learning methods, discriminating single fibres, particles, juxtaposed objects and fibre bundles, minimizing as much image noise.