Asbestos Detection with Fluorescence Microscopy Images and Deep Learning

Our plain-English summary of Asbestos Detection with Fluorescence Microscopy Images and Deep Learning, by C. Cai, T. Nishimura, J. Hwang, et al (Sensors, 2021).

The study

• Authors: C. Cai, T. Nishimura, J. Hwang, X.M. Hu, A. Kuroda
• Journal: Sensors
• Published: 2021;21(13)
• Cited by: 6 articles, as counted by NIH iCite
• Read it: PubMed · Publisher · Free full text

Abstract

Fluorescent probes can be used to detect various types of asbestos (serpentine and amphibole groups); however, the fiber counting using our previously developed software was not accurate for samples with low fiber concentration. Machine learning-based techniques (e.g., deep learning) for image analysis, particularly Convolutional Neural Networks (CNN), have been widely applied to many areas. The objectives of this study were to (1) create a database of a wide-range asbestos concentration (0-50 fibers/liter) fluorescence microscopy (FM) images in the laboratory; and (2) determine the applicability of the state-of-the-art object detection CNN model, YOLOv4, to accurately detect asbestos. We captured the fluorescence microscopy images containing asbestos and labeled the individual asbestos in the images. We trained the YOLOv4 model with the labeled images using one GTX 1660 Ti Graphics Processing Unit (GPU). Our results demonstrated the exceptional capacity of the YOLOv4 model to learn the fluorescent asbestos morphologies. The mean average precision at a threshold of 0.5 ([email protected]) was 96.1% ± 0.4%, using the National Institute for Occupational Safety and Health (NIOSH) fiber counting Method 7400 as a reference method. Compared to our previous counting software (Intec/HU), the YOLOv4 achieved higher accuracy (0.997 vs. 0.979), particularly much higher precision (0.898 vs. 0.418), recall (0.898 vs. 0.780) and F-1 score (0.898 vs. 0.544). In addition, the YOLOv4 performed much better for low fiber concentration samples (<15 fibers/liter) compared to Intec/HU. Therefore, the FM method coupled with YOLOv4 is remarkable in detecting asbestos fibers and differentiating them from other non-asbestos particles.

In plain English

This laboratory study tested whether a deep-learning image program, YOLOv4, could find and count asbestos fibers in fluorescence microscopy images of air filters stained with asbestos-binding fluorescent probes. The authors used 60 air filter samples: 40 from airborne dust at demolition sites in Japan and the others from relatively pure asbestos minerals. Fibers were marked by hand following NIOSH fiber counting Method 7400. The program was trained on 176 images from 13 samples and tested on 47 images from seven other samples. On the test images, at a detection threshold of 0.5, YOLOv4 scored higher than the authors' earlier software, Intec/HU, on accuracy (0.997 vs. 0.979), precision (0.898 vs. 0.418), recall (0.898 vs. 0.780) and F1 score (0.898 vs. 0.544). It missed 6 hand-counted fibers, against 13 for Intec/HU. It wrongly counted 6 non-asbestos particles as asbestos, against 64. At low fiber concentrations (under 15 fibers per liter), YOLOv4 counts at a threshold of 0.7 matched manual counts more closely (r = 0.921) than Intec/HU counts did (r = 0.768). The authors concluded that YOLOv4 can help people recognize asbestos fibers in these images, but that more work is needed.

We wrote this plain-English summary from the full article with the help of AI, and a second AI model checked each sentence against the article. It is not the authors' wording; the abstract and the article are their own account.

What the study did and found

• Design: Laboratory method-development and evaluation study (published as a 'Communication'): training and testing a deep-learning object-detection model (YOLOv4) on fluorescence microscopy images, compared with the authors' earlier counting software and with manual counts
• Who was studied: 60 air filter samples: 40 from airborne dust at various demolition sites in Japan, the rest from relatively pure asbestos minerals dispersed in an air-tight chamber. 176 images from 13 samples were used for training and 47 images from seven other samples for testing. 40 filter samples (20 demolition-site, 20 pure-mineral) were used for concentration counting.
• Exposure: Airborne asbestos fibers on filters, stained with the Asbester Air 2 kit. The fiber types in the samples are not specified; the probes are described as detecting serpentine and amphibole asbestos. Fibers counted were more than 5 micrometres long and less than 3 micrometres in diameter, per NIOSH Method 7400, over a range of 0 to 50 fibers per liter.
• Main results:
  • YOLOv4 (trained three times): mean average precision 96.1% plus or minus 0.4% at a threshold of 0.5
  • YOLOv4 at threshold 0.5: precision 0.943 plus or minus 0.006, recall 0.910 plus or minus 0.010, F1 score 0.927 plus or minus 0.006
  • Comparison on test images, YOLOv4 at threshold 0.5 vs. Intec/HU software: accuracy 0.997 vs. 0.979; precision 0.898 vs. 0.418; recall 0.898 vs. 0.780; F1 score 0.898 vs. 0.544
  • Of 59 manually counted asbestos fibers (Table 2), Intec/HU missed 13 and YOLOv4 at threshold 0.5 missed 6. Of 3,650 manually classified non-asbestos objects, Intec/HU counted 64 as asbestos and YOLOv4 counted 6.
  • Correlation with manual counts across 40 filter samples: Intec/HU r = 0.768 at low fiber concentration (under 15 fibers per liter) and r = 0.964 over the wide range. YOLOv4 at threshold 0.7, including pure-mineral samples, r = 0.921 at low concentration and r = 0.979 over the wide range.
  • Example errors: YOLOv4 counted an aggregated particle as a fiber and missed a fiber with a blurred boundary near other particles. It also double-counted a fiber overlapping a non-fiber particle and counted two crossed fibers as one.
• The authors' conclusion: The authors concluded that YOLOv4 is a tool that can assist humans in recognizing asbestos fibers in fluorescence microscopy images, performing better than Intec/HU, especially at low fiber concentrations, but that additional work is needed
• Limitations the authors state:
  • More training images are needed for the model to recognize asbestos in uncommon situations
  • The model does not yet measure fiber dimensions (length, width, aspect ratio)
  • The fluorescence microscopy method is not a NIOSH-approved method

Topics

The National Library of Medicine indexes this study under subjects our site covers: Chrysotile.

Medical terms in this study, explained in our Lexicon: Serpentine Group, Amphibole Group, Fiber Counting, NIOSH.

Related studies

• Deep-learning based classification distinguishes sarcomatoid malignant mesotheliomas from benign spindle cell mesothelial proliferations — J.R. Naso, A.B. Levine, et al. (2021)
• Deep Learning for the Automatic Quantification of Pleural Plaques in Asbestos-Exposed Subjects — I. Benlala, B.D. De Senneville, et al. (2022)
• Evaluation of sensitivity of fluorescence-based asbestos detection by correlative microscopy — T. Ishida, M. Alexandrov, et al. (2012)
• Rapid on-site detection of airborne asbestos fibers and potentially hazardous nanomaterials using fluorescence microscopy-based biosensing — A. Kuroda, M. Alexandrov, et al. (2016)
• Construction of Asbestos Slate Deep-Learning Training-Data Model Based on Drone Images — S.C. Baek, K.H. Lee, et al. (2023)
• Differential Counting of Asbestos Using Phase Contrast and Fluorescence Microscopy — T. Nishimura, M. Alexandrov, et al. (2016)

Browse more research in Sensors, or return to medical research.