Development of a rapid fiber-detection system using artificial intelligence in phase-contrast microscope images of actual atmospheric samples

Yukiko Iida¹Takashi Yamamoto^2*Kazuharu Iwasaki³Ken-Ichi Yuki¹Kentaro Kiri³Hayato Yamashiro³Toshiyuki Toyoguchi¹Atsushi Terazono²

• ¹Environmental Control Center Co., Ltd., Tokyo, Japan
• ²National Institute for Environmental Studies (NIES), Tsukuba, Ibaraki, Japan
• ³Japan NUS Co. Ltd., Tokyo, Japan

In this study, we attempted to detect fibers in phase contrast microscope images of actual atmospheric samples using an automatic fiber detection system based on artificial intelligence (AI) models and image processing. In order to detect and correct the release of asbestos fibers due to improper demolition and removal operations of asbestos-containing building materials as early as possible, it is essential to develop a method that can rapidly and accurately measure airborne asbestos fibers. Current rapid measurement method is the combination short-term atmospheric sampling with counting using a phase contrast microscope. However, visual fiber counting takes a reasonable amount of time and is not sufficiently rapid. Additionally, since the counting process relies on visual fiber counting, analytical accuracy can be decreased due to factors such as analyst fatigue. Ambient air samples or air samples collected near demolition sites were observed and acquired using a phase contrast microscope. From the acquired microscopic images and the fiber counting results by the expert analysts, we created a set of 98 training datasets. The Segformer, one of the semantic segmentation models that had achieved good accuracy in previous studies, was adopted as an AI model for automatic fiber detection system. Of the 98 training datasets, 77 datasets were used for training the model, and 21 datasets were used to evaluate the accuracy of the automatic fiber detection system. The achieved detection accuracy by the AI model was 0.90 for recall, 0.68 for precision, and 0.77 for F1 score. Fiber counting accuracy using an automatic fiber detection system based on AI models and image processing was 0.78 for recall, 0.67 for precision, and 0.72 for F1 score. The time required to detect fibers was about one second per image using a graphics processing unit. The counting accuracy by this automatic fiber detection system based on AI model is comparable to that of manual counting by a skilled analyst, yet the time required for fiber counting is 12 to 50 times faster, significantly reducing the time required for analysis.