Development of an automated asbestos counting ...

4 downloads 269 Views 731KB Size Report
Abstract An emerging alternative to the commonly used analytical methods for asbestos analysis is fluores- cence microscopy (FM), which relies on highly ...
Environ Monit Assess (2015) 187:4166 DOI 10.1007/s10661-014-4166-y

Development of an automated asbestos counting software based on fluorescence microscopy Maxym Alexandrov & Etsuko Ichida & Tomoki Nishimura & Kousuke Aoki & Takenori Ishida & Ryuichi Hirota & Takeshi Ikeda & Tetsuo Kawasaki & Akio Kuroda

Received: 28 February 2014 / Accepted: 17 November 2014 # Springer International Publishing Switzerland 2014

Abstract An emerging alternative to the commonly used analytical methods for asbestos analysis is fluorescence microscopy (FM), which relies on highly specific asbestos-binding probes to distinguish asbestos from interfering non-asbestos fibers. However, all types of microscopic asbestos analysis require laborious examination of large number of fields of view and are prone to subjective errors and large variability between asbestos counts by different analysts and laboratories. A possible solution to these problems is automated counting of asbestos fibers by image analysis software, which would lower the cost and increase the reliability of asbestos testing. This study seeks to develop a fiber recognition and counting software for FM-based asbestos analysis. We discuss the main features of the developed software and the results of its testing. Software testing showed good correlation between automated and manual counts for the samples with medium and high fiber concentrations. At low fiber concentrations, the automated counts were less accurate, leading us to implement correction M. Alexandrov, E. Ichida, and T. Nishimura have equal contributions to this work. M. Alexandrov : T. Nishimura : T. Ishida : R. Hirota : T. Ikeda : A. Kuroda (*) Department of Molecular Biotechnology, Graduate School of Advanced Sciences of Matter, Hiroshima University, Higashi Hiroshima, Hiroshima 739-8530, Japan e-mail: [email protected] E. Ichida : K. Aoki : T. Kawasaki Advanced Technology Research and Development Institute, INTEC Inc., Yokohama, Kanagawa 221-8520, Japan

mode for automated counts. While the full automation of asbestos analysis would require further improvements in accuracy of fiber identification, the developed software could already assist professional asbestos analysts and record detailed fiber dimensions for the use in epidemiological research. Keywords Automated asbestos counting software . Fluorescence microscopy . Asbestos fibers

Introduction Asbestos is the name given to six fibrous silicate minerals that have been widely used in various construction materials because of their chemical and thermal stability (Mossman et al. 1990; National Institute of Occupational Safety and Health (NIOSH) 2011). Asbestos minerals are made up of microscopic bundles of silicate fibers that can easily become airborne. Although the use of asbestos is now prohibited in most developed countries, asbestos contamination remains a common problem, contributing to the increasing incidence of asbestoslinked pleural mesothelioma and lung cancer (Davis et al. 1978; World Health Organization (WHO) 2000). A number of analytical methods are available for airborne asbestos detection and identification. The most commonly used method for air samples relies on phase contrast microscopy (PCM) to identify and count all fibers that are longer than 5 μm, thinner than 3 μm, and have aspect ratios larger than 3:1 (National Institute of Occupational Safety and Health (NIOSH) 1994).

4166, Page 2 of 10

While simple and cheap, PCM has a number of limitations. It has lower sensitivity for thin chrysotile fibers and does not provide an easy way to distinguish asbestos fibers from other natural or man-made fibers of similar dimensions (National Institute of Occupational Safety and Health (NIOSH) 1994, 2011). A common alternative to PCM for detecting and analyzing airborne asbestos is electron microscopy (EM), which could be used to detect extremely thin fibers and identify different types of asbestos. However, its application is rather limited due to the high cost and time-consuming sample preparation and analysis (Taylor et al. 1984). Like EM, polarized light microscopy (PLM) can differentiate asbestos from non-asbestos fibers, but is generally not applied to the analysis of airborne asbestos due to low sensitivity (below that of PCM). Recognizing the limitations of these methods, the National Institute for Occupational Safety and Health has recently identified development of improved analytical methods for asbestos fibers as a strategic research goal (National Institute of Occupational Safety and Health (NIOSH) 2011). An emerging alternative to the commonly used analytical methods is fluorescence microscopy (FM), which employs fluorescent staining by specific asbestos-binding probes to both visualize asbestos and distinguish it from the unstained non-asbestos fibers (Kuroda et al. 2008; Ishida et al. 2010, 2012, 2013). Compared to the analytical methods for other industrial pollutants, testing for airborne asbestos fibers suffers from two major drawbacks. First, all the analytical methods require careful microscopic examination of large number of fields of view and are therefore extremely labor intensive. Second, the counting procedure is relatively complicated and is prone to large subjective errors. The differences in analyst skill and the degree of adherence to the counting rules combine to produce large variability between asbestos counts by different analysts and laboratories (National Institute of Occupational Safety and Health (NIOSH) 1994). A possible solution to these problems is to provide a reliable automated alternative to manual fiber counting. An automated asbestos counting system based on image analysis software would drastically reduce human involvement in asbestos testing, leading to lower cost and increased reliability of asbestos testing. The involvement of human operator could potentially be limited to relatively simple tasks of sample preparation, focusing the microscope and selection of (random) fields of view for image acquisition. Since these tasks do not require any special

Environ Monit Assess (2015) 187:4166

skills beyond basic microscope operation, the testing could be conducted on site by the companies/ organizations dealing with possible asbestos contamination, rather than by specialized asbestos testing laboratories. One of the earliest and the most intensive efforts to develop an automated fiber counting system was carried out in 1970s by Manchester University in collaboration with the Health and Safety Executive in the UK (Kenny 1984). Their PCM-based system was reportedly used as a reference by some laboratories in the UK reference sample programs, but was eventually discontinued because it was not sufficiently consistent for all types of samples (Baron 2001). Subsequent attempts to automate fiber counting originate from Asian countries and are much more recent (Inoue et al. 1998, 1999; Kawabata et al. 2009; Ishizu et al. 2010). Unfortunately, these attempts have not gone beyond development of prototype hardware and/or software and their preliminary testing. One of the issues that the PCM-based counting system had to deal with was interference from air bubbles and halos around larger fibers and particles, which could be misidentified as fibers (Ishizu et al. 2010). Yet, the main problem with the PCM-based automated asbestos analysis is the inability to differentiate between asbestos and non-asbestos fibers. Therefore, any automated fiber counting system based on this analytical method would be severely limited by the capabilities of PCM. Compared to PCM, fluorescence microscopy provides a highly advanced capability to differentiate asbestos from non-asbestos fibers (Ishida et al. 2010, 2013). While this capability is below the level provided by electron microscopy, it does not require any advanced skills on the part of the operator. When using polarized light or electron microscopy, each fiber needs to be individually analyzed by a highly skilled operator to determine if it is asbestos or not. With FM, no special skills or even human involvement are necessary to distinguish specifically stained asbestos from unstained and therefore invisible non-asbestos fibers. Fluorescence microscopy is therefore the most suitable platform for automated asbestos fiber detection and counting. In 2013, Cho and coworkers reported the first largely FMbased automated fiber counting system that relies on the fluorescent asbestos probe developed by Kuroda et al. (2008) in combination with dual-mode high-throughput microscopy (DM-HTM) device that is capable of automated image acquisition (Cho et al. 2013). The initial

Environ Monit Assess (2015) 187:4166

tests indicated a good correlation between manual and automatic counts for artificial dust-free samples of chrysotile asbestos. However, the counting software used by Cho and coworkers (ImageJ freeware) does not implement the commonly used rules for counting splayed and crossed fibers, and there are no test results for the actual asbestos samples. Therefore, in order to fully realize the potential of FM-based asbestos analysis, we developed an image analysis software that is specifically tailored for automated fiber recognition and counting. This paper describes the main features of the developed software and presents the results of its testing.

Materials and methods Sample preparation The accuracy of the automated fiber counting system was tested using 103 air sample filters provided by the Japan’s Ministry of the Environment (43 samples) and two independent asbestos testing laboratories, Kankyo Research (Hachioji, Japan; 30 samples) and Kyuden Sangyo (Fukuoka, Japan; 30 samples). Since all the samples were provided by the outside parties for the express purpose of testing the developed system, only limited information has been made available regarding sample sources and preparation methods. Approximately half (50) of the tested samples were prepared using relatively pure asbestos and some non-asbestos fibers. Such samples generally had very low density of nonfibrous dust particles and, for the purpose of this study, were grouped as “low dust samples.” Low dust samples were prepared by filtering air from an air-tight chamber containing artificially dispersed fibers through a nitrocellulose filter. Almost half of low-dust samples were prepared using chrysotile: only chrysotile (11 samples), chrysotile and amosite (7 samples), chrysotile, amosite, and mineral wool (2 samples); and chrysotile, tremolite, and mineral wool (3 samples). Samples without chrysotile were prepared using the following fibrous minerals: only amosite (11 samples); amosite and mineral wool (4 samples); amosite and crocidolite (3 samples); only crocidolite (4 samples); tremolite and mineral wool (3 samples); and wollastonite and mineral wool (2 samples). The remaining (53) samples were collected at various demolition sites in Japan or prepared by milling asbestos-containing construction materials, dispersing

Page 3 of 10, 4166

the resulting mixture of asbestos fibers and nonasbestos particles in an air-tight chamber, and filtering air from the chamber through a nitrocellulose filter. Since the artificially prepared samples in this group generally have high dust loads and are virtually undistinguishable from the demolition site samples, all of the remaining samples are considered “high dust samples.” Fluorescence staining and image acquisition Samples were stained using Asbester Air 2 kit by Siliconbio Inc. (Hiroshima, Japan) according to the manufacturer’s instructions. Sample staining, acquisition of FM images for the automated counting, and manual counting were conducted by the staff of Hiroshima University (Hiroshima, Japan; 40 samples), Kankyo Research (33 samples) and Kyuden Sangyo (30 samples). All the FM images were acquired using Primo Star iLED microscopes equipped with AxioCam ICm1 CCD (Carl Zeiss Microscopy, Oberkochen, Germany). Automated counting was conducted using Asbester Counter software (INTEC Inc. 2012) by INTEC Inc. (Kanagawa, Japan) installed on the notebook computers running Windows 7 operating system.1 All the laboratory staff involved in image acquisition and manual counting have extensive experience using PCM for asbestos analysis and therefore required only short training in the application of FM method and basic operation of the fluorescence microscope. Manual counting was conducted according to the fiber counting rules of NIOSH Method 7400 (“Asbestos and Other Fibers by PCM”). The only significant modification of NIOSH Method 7400 fiber counting rules was that the counting (both manual and automated) was conducted across the whole image, rather than the circular Walton– Beckett graticule area used in PCM testing. Preprocessing of the images The developed software uses a combination of background correction and contrast adjustment to improve visibility (and automated detection) of the thinner fibers. First, the software divides each image into 40 rectangles 1

The automated counting software and the fluorescent probe can be provided for the purpose of testing. The software can be provided free of charge with the temporary license for the duration of testing.

4166, Page 4 of 10

and measures the average background brightness in each rectangle. These data are used to correct the variations in the background brightness and to adjust the contrast across different image areas. The method used for contrast adjustment is histogram equalization, as it can dramatically improve contrast in the image areas with low local contrast, where the brightness of fibers is close to that of background. Histogram equalization works by spreading out the most frequent pixel brightness values and is far less effective when the pixel brightness values already show large variation across the image area. This can happen due to the presence of brightly fluorescent dust particles. To achieve a meaningful improvement in contrast and fiber visibility, image preprocessing includes removal (subtraction) of fluorescent dust particles from the microscope images. However, some of the fibers adjacent to dust particles could, at this point, be mistakenly identified as elongated particles. To avoid removing such particles along with the adjacent fibers, the particle subtraction algorithm is set to target particles with aspect ratio (length divided by width) below 1.5. This value is well below the minimum aspect ratio of countable fibers, which is commonly set at 3. With such strict settings for particle aspect ratio, the algorithm can remove the bulk of interfering dust for effective contrast adjustment, while retaining more elongated particles for subsequent analysis.

Fiber identification and reconstruction of complex fiber structures

Environ Monit Assess (2015) 187:4166

individual fibers and fiber fragments and determine their relative orientation. Measurements of the angles between individual fibers and fiber fragments are used to reconstruct more complex fiber structures that may be present in the FM image, such as splayed (branching) fibers, crossing fibers, and longer fibers that are split into several fragments by overlapping dust particles. The final identification of countable fibers is conducted based on the reconstructed fiber dimensions and the rules for counting complex fiber structures. Fiber classification and counting When highlighting the identified fibers, the software classifies them into the “red” and “yellow” groups according to fiber dimensions. The fibers are included in the yellow group if they satisfy any of the following conditions: 1. Length between 5 and 10 μm 2. Width between 2 and 3 μm 3. Aspect ratio between 3 and 5. The fibers in the red group are longer than 10 μm, thinner than 2 μm, and have an aspect ratio above 5. The automated fibers counts are calculated by adding the red and yellow group counts. When reporting the test results, the software converts the automated fiber counts for each sample into fiber concentration expressed in fibers per liter (f/L). Correction mode

The main task of the image analysis software is to distinguish countable fibers from particles and other fibers that do not fit the dimensions of countable fibers. According to the generally accepted fiber counting rules, countable fibers are longer than 5 μm, narrower than 3 μm, and have length/width ratio (aspect ratio) ≥3:1. The developed software calculates fiber length (L) and width (W) from fiber perimeter (P) and surface area (A) by solving the following equations: P ¼ 2ð L þ W Þ A ¼ LW The perimeter (and subsequently the area) of the fibers are determined using a general contour tracing algorithm. When multiple fibers or fibers and particles overlap, thinning algorithm is used to separate

To facilitate the use of asbestos counting software by asbestos analysis professionals, we implemented additional tools for correction of automated counts. When the software is used in correction mode, a qualified operator can check fiber highlighting and measurements for the automated counts and reject the misidentified yellow group fibers. As the correction is limited to yellow group fibers, it is somewhat less accurate but faster than manual fiber counting. The software records operator’s corrections and converts the corrected fiber counts for each sample into fiber concentration expressed in fibers per liter (f/L). Correction of automated counts was conducted for 27 high dust load samples with automated fiber counts below 25 f/L.

Environ Monit Assess (2015) 187:4166

Results and discussion Automated detection of asbestos fibers Automated detection of asbestos fibers requires sufficient contrast between the fibers and background, which is not always attained across each fluorescence image. We found that FM images required image processing to improve the contrast and visibility of the thinnest fibers near the detection limit. The detection limit does not necessarily mean low absolute values of fiber brightness, but rather small difference between the brightness of the thinnest fibers and that of the background. To compound the problem, the brightness of the background itself generally varies between and across fluorescence microscopy images. This is usually due to uneven distribution of the excitation light intensity across the field of view, local distribution of brightly fluorescent fibers and particles, and differences in the residual staining of the membrane filter. While human eye can easily adapt to large variations in background brightness, automated detection requires image preprocessing to reduce and eliminate this variation. The preprocessing steps include correction of fluorescent background, subtraction of fluorescent dust particles, and contrast adjustment. The resolution of the images taken for asbestos analysis is 1280×960 pixels, and 1 pixel equates to approximately 250 nm. Light diffraction determines the resolution limit of approximately 200–250 nm for most kinds of optical microscopy using visible light. This means that even 30 nm fibers would have apparent width of approximately 250 nm (1 pixel). Since the size of the area represented by each pixel is also 250 nm, the thinnest fibers would be captured in either one or two adjacent pixels. When the fibers (or their segments) are represented in each of two adjacent pixels, these pixels would have lower brightness values than the pixels that fully represent the width of the same fiber. In the developed software, histogram equalization was used for contrast adjustment because it helped improve the contrast between dim fibers and the background, facilitating correct detection of thinnest fibers. Following contrast adjustment, even half of the original brightness value would generally be sufficient to correctly detect all sections of thin fibers.

Page 5 of 10, 4166

dimensions is the most important step in the automated fiber detection. The easiest method of automated fiber identification is to fit all the fluorescent objects on the FM image inside the smallest possible rectangle and assume that the dimensions of each object equal the length and width of its rectangle. If the rectangle dimensions satisfy the definition of countable fibers, the object would be counted as asbestos fiber. The obvious problem with the above method is that it cannot discriminate between fibers and elongated particles (or several overlapping particles). Moreover, such a measurement of fiber dimensions is only accurate for straight fibers. For other kinds of fibers, in particular chrysotile asbestos, this method would mistake fiber curvature for fiber width, producing large systemic biases. Therefore, we chose to calculate fiber length and width from fiber perimeter and surface area. The main advantage of using fiber perimeter and surface area is that these values are largely independent of the fiber curvature. This method therefore gives reasonably accurate results for wavy or curvy chrysotile fibers. For the chrysotile fibers in Fig. 1, the calculated width was 2 μm, and the lengths of fibers 1, 2, and 3 were 29, 39, and 28 μm respectively. The calculated values generally agreed with the (more approximate) estimates by human observers. Reconstruction of complex fiber structures Asbestos samples may contain some complex fiber structures, such as fiber bundles, agglomerates, splayed and crossed fibers. Furthermore, longer fibers are

Measurement of fiber dimensions Since the countable fibers are defined based on their length, width, and aspect ratio, measurement of fiber

Fig. 1 Recognition of curvy fibers. The identified fibers are numbered and highlighted by red ovals. Bar, 25 μm

4166, Page 6 of 10

sometimes partially obscured by overlaying dust particles and could be incorrectly counted as two or more separate fibers. Asbestos counting involves application of a relatively complex set of counting rules, which stipulate which visible structures should be counted as one, two, or more fibers, and which structures should not be counted at all. Following tentative fiber identification, it is necessary to check whether the identified fibers are indeed separate and whether they belong to any complex fiber structures. For this task, the software determines relative position and orientation of tentatively identified fibers in order to reconstruct more complex fiber structures, reconnect longer fibers that are split into several fragments by overlapping dust particles, and distinguish between the splayed fibers and crossed fibers. According to the asbestos counting rules, splayed fibers must be counted as a single fiber, while crossed fibers must be separately counted. As shown in Fig. 2, the software initially identifies individual fibers that form the splayed fiber, but counts them as a single fiber as the angle between the constituent fibers is