Protecting humans from the inhalation of potentially hazardous fibres requires controlling their exposure by microscopic characterization of filter-sampled aerosols. Therefore, it is a crucial task to reliably recognize, characterize morphologically and count critical fibres in electron microscopic images. Since visual inspection of large greyscale image areas is a tedious and time-consuming process, an automated AI-assisted fibre instance recognition approach was developed. It combines pixelwise semantic image segmentation by a U-Net-like model with algorithms to search, trace and refine fibre-shaped segments. This way, a two-step instance segmentation is achieved that provides fibre morphology, criticality and number information even for fibres in intersecting or overlapping configurations. The model was optimized by supervised training on a dataset of more than 1,000 human-annotated (20 Mpx) scanning electron micrographs. The quality of these training data as well as the reliability of fibre instance recognition of both our automated approach and of human evaluators were assessed by comparing their performance for 6 carbon nanotube materials that exhibited different degrees of morphological complexity. The results show good agreement with human evaluators in terms of the number of fibres found, length and width determination. With limitations for very complex morphological configurations where also humans face undecidable ambiguities or extremely time-consuming fibre tracing tasks, our fibre tracing algorithm proved capable to disentangle intersections and to count fibres in unambiguously countable clusters. Our approach achieves a large reduction in human workload. The resulting fibre instance data opens the possibility to develop purely AI-based instance segmentation solutions.

This article is published in the Journal "Powder Technology" (2025).