Rapid fiber-detection technique by artificial intelligence in phase-contrast microscope images of simulated atmospheric samples

Our plain-English summary of Rapid fiber-detection technique by artificial intelligence in phase-contrast microscope images of simulated atmospheric samples, by T. Yamamoto, K. Iwasaki, Y. Iida, et al (Annals of Work Exposures and Health, 2024).

The study

• Authors: T. Yamamoto, K. Iwasaki, Y. Iida, K.I. Yuki, F. Nakaji, H. Yamashiro, T. Toyoguchi, A. Terazono
• Journal: Annals of Work Exposures and Health
• Published: 2024;68(4):420-426
• Cited by: 1 article, as counted by NIH iCite
• Read it: PubMed · Publisher · Free full text

Abstract

Since the manufacture, import, and use of asbestos products have been completely abolished in Japan, the main cause of asbestos emissions into the atmosphere is the demolition and removal of buildings built with asbestos-containing materials. To detect and correct asbestos emissions from inappropriate demolition and removal operations at an early stage, a rapid method to measure atmospheric asbestos fibers is required. The current rapid measurement method is a combination of short-term atmospheric sampling and phase-contrast microscopy counting. However, visual counting takes a considerable amount of time and is not sufficiently fast. Using artificial intelligence (AI) to analyze microscope images to detect fibers may greatly reduce the time required for counting. Therefore, in this study, we investigated the use of AI image analysis for detecting fibers in phase-contrast microscope images. A series of simulated atmospheric samples prepared from standard samples of amosite and chrysotile were observed using a phase-contrast microscope. Images were captured, and training datasets were created from the counting results of expert analysts. We adopted 2 types of AI models-an instance segmentation model, namely the mask region-based convolutional neural network (Mask R-CNN), and a semantic segmentation model, namely the multi-level aggregation network (MA-Net)-that were trained to detect asbestos fibers. The accuracy of fiber detection achieved with the Mask R-CNN model was 57% for recall and 46% for precision, whereas the accuracy achieved with the MA-Net model was 95% for recall and 91% for precision. Therefore, satisfactory results were obtained with the MA-Net model. The time required for fiber detection was less than 1 s per image in both AI models, which was faster than the time required for counting by an expert analyst.

© The Author(s) 2024. Published by Oxford University Press on behalf of the British Occupational Hygiene Society.

In plain English

This laboratory study from Japan tested whether two artificial intelligence (AI) models could find asbestos fibers in phase-contrast microscope images faster than counting by eye. It did not study people or disease. The authors made simulated air samples by filtering standard samples of chrysotile and amosite onto membrane filters and photographing fields of view under the microscope. An expert analyst counted fibers at least 5 micrometers long, less than 3 micrometers wide, and at least three times longer than wide. Of 40 image sets, 30 were used to train the models and 10 to test them. Measured against the analyst's 55 fibers in the test images, the Mask R-CNN model was reported at 57% recall and 46% precision, and the MA-Net model at 95% recall and 91% precision. The article's own fiber counts for Mask R-CNN do not match its printed percentages. Both models took less than 1 second per image. The authors concluded that MA-Net gave satisfactory results. They expected samples from real demolition sites to contain many particles and non-asbestos fibers, but they tested only simulated samples in this study.

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 methods study comparing two AI image-analysis models (Mask R-CNN and MA-Net) with counts by an expert analyst
• Who was studied: No human subjects. The material was ten microscope slides of simulated atmospheric samples, five chrysotile (Chr-1 to Chr-5) and five amosite (Amo-1 to Amo-5). Of 40 image and training-data sets, 30 were used for training and 10 for evaluation.
• Exposure: No human exposure was studied. The samples were made from a chrysotile standard sample (Union for International Cancer Control) and an amosite standard sample (Japan Association for Working Environment Measurement), filtered onto mixed cellulose ester membrane filters. Fibers were counted if they were 5 micrometers or longer, less than 3 micrometers wide, and had an aspect ratio of 3 or more.
• Main results:
  • Mask R-CNN model, against the expert analyst's 55 fibers in the evaluation images: 34 fibers detected as fibers to be counted, 21 overlooked, 7 overcounted. It tended to overcount in amosite images and to overlook fibers, especially wavy fibers, in chrysotile images.
  • MA-Net model, against the same 55 fibers: 52 detected as fibers to be counted, 3 overlooked, 5 overcounted. It detected the wavy chrysotile fibers that Mask R-CNN missed.
  • Reported accuracy: Mask R-CNN recall 57% and precision 46%; MA-Net recall 95% and precision 91%.
  • Both models took less than 1 second per image, or about 2 minutes for 100 fields of view, compared with 25 to 100 minutes for counting 100 fields under the current Japanese method. This time excludes manual focusing and image capture.
  • The average image showed 5.5 fibers per field of view. The authors calculated that this corresponds to an airborne concentration of 31.2 fibers/L, assuming 2,400 L of air collected on a 47 mm filter.
• The authors' conclusion: The authors concluded that satisfactory results were obtained with the MA-Net model and that counting time was reduced to 1/12th to 1/50th of an expert analyst's time. They wrote that the technique has potential for early detection and correction of asbestos emissions from inappropriate demolition work.
• Limitations the authors state:
  • Only simulated samples made from asbestos standard samples were tested. The authors assume that real demolition-site samples contain many particles and non-asbestos fibers, and they intend to study such samples in future.
  • The Mask R-CNN model was probably not trained enough because only a small number of training datasets were available.
  • Both models overcounted fibers below the counting criteria because they extracted fiber outlines larger than they were. Contour extraction needs improvement.
  • The time savings do not include manual focus adjustment and image capture. The authors say these need to be automated.

Topics

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

Medical terms in this study, explained in our Lexicon: Asbestos-Containing Material (ACM), Amosite, Chrysotile.

Related studies

• Development of rapid and highly accurate method to measure concentration of fibers in atmosphere using artificial intelligence and scanning electron microscopy — Y. Iida, K. Watanabe, et al. (2021)
• Artificial intelligence-based diagnosis of asbestosis: analysis of a database with applicants for asbestosis state aid — K.B.W. Groot Lipman, C.J. de Gooijer, et al. (2023)
• An inter-laboratory study to determine the effectiveness of procedures for discriminating amphibole asbestos fibers from amphibole cleavage fragments in fiber counting by phase-contrast microscopy — M. Harper, E.G. Lee, et al. (2012)
• Evaluation of the dark-medium objective lens in counting asbestos fibers by phase-contrast microscopy — E.G. Lee, J.H. Nelson, et al. (2015)
• Differential Counting of Asbestos Using Phase Contrast and Fluorescence Microscopy — T. Nishimura, M. Alexandrov, et al. (2016)
• Assessment of occupational exposure to asbestos fibers: Contribution of analytical transmission electron microscopy analysis and comparison with phase-contrast microscopy — C. Eypert-Blaison, A. Romero-Hariot, et al. (2018)

Browse more research in Annals of Work Exposures and Health, or return to medical research.