Deep learning-assisted chrysotile asbestos screening in soil: A hierarchical statistical approach to error propagation

Manual polarized light microscopy (PLM) is widely used for asbestos screening in soil, yet the technique remains labor-intensive and susceptible to inter-analyst variability and mineralogical interferences. We developed a chrysotile-focused, deep learning-assisted PLM workflow that links field-of-view (FOV) detections to sample-level screening decisions under regulatory-style sampling designs. The method combines a two-stage instance-segmentation pipeline with rotation-assisted paired imaging and recovery (0°/90°) to mitigate optical extinction and retain auditable optical evidence. FOV-level errors are propagated through slide- and sample-level consensus rules using an occupancy parameter q, defined as the probability that a given FOV contains detectable fibers, together with a dataset-calibrated proxy to approximate mass-fraction interpretation. Under a baseline design (3 slides per sample; 20 FOVs per slide), and assuming a 10% sample-level prevalence, strict consensus supported confirmatory performance near moderate occupancy (q≈0.20; ∼0.05% w/w under the present proxy) under the observed workflow-specific false-positive rate. Sustaining confirmatory performance at lower occupancy required increasing the number of scanned FOVs. Below q ≈ 0.07 (approximately 0.017% w/w under the present proxy calibration), confirmatory performance could not be jointly maintained under the present workflow assumptions and estimated background-error structure. These results are translated into an operational triage framework: samples satisfying the strict consensus criterion support a confirmatory-positive call; samples with only one consensus-positive slide trigger targeted additional scanning and/or complementary analysis; and consensus-negative samples support a negative screen within the protocol's confirmable range are reported as "not detected under this protocol".