The conventional dicentric chromosome assay is the most reliable approach for measuring the level of absorbed radiation in victims of radiation accidents among the various biological dose-assessment methods. Therefore, the entire process, from acquiring metaphase cells to scoring and identifying chromosomes, has been well-harmonized by IAEA 201117 and ISO 201418. However, the dicentric chromosome assay requires trained experts, is time-consuming, especially the scoring process, and has limited utility for processing a large number of samples. Various laboratories have collaborated to standardize the biological dosimetry method to overcome these challenges, validating its accuracy and performance31–34 and minimizing the scoring time34–36. We investigated a method for automatically constructing dose curves and calculating the absorbed radiation dose using DLADES, an automated dicentric chromosome scoring system based on a deep learning technique. Blood samples from two healthy donors in their 30s were exposed to cobalt-60 gamma rays at a 0.8 Gy/min dose rate over six dose ranges from 0 to 4 Gy. Metaphase cells were induced following the IAEA 2011 protocol17, and images were captured using the Metafer4 System. The captured images were exported to JPEG format, and the number of monocentric and dicentric chromosomes for each dose was counted using DLADES. The number of dicentric chromosomes increased dose-dependently. However, the results differed from those obtained using conventional dicentric chromosome assays. The dicentric chromosome follows a Poisson distribution in a conventional dicentric chromosome assay using low-LET radiation. As shown in Table 1, most doses were overdispersed. Romm et al. attributed this outcome to the image quality26. In the conventional assay, the reviewer selects complete cells with 46 chromosomes, resulting in images of similar quality and condition.
However, the automatic system does not classify entire cells when analyzing the image. Therefore, the variable appearance of chromosomal types in captured images is a limitation of automatic scoring systems similar to manual scoring. We resolved this problem using the DLADES filtering function to sort the “Accepted” images. DLADES accurately counted and identified dicentric and monocentric chromosomes (97% and 90%, respectively)28. Therefore, we attempted to sort similar-quality images by “Accepted” classification of DLADES, yielding an average number of chromosomes in S1 and S2 images of 41.12 and 42.10, respectively, with a standard deviation of 9.75 for both; thus, “Accepted” data were clustered around the mean compared to the total dataset. By excluding the 1.5 × IQR outlier, images with similar conditions were obtained, causing the mean and standard deviation to increase slightly from 42.99 to 43.35 and from 7.07 to 8.19, respectively.
By sorting similar condition images, our goal was to determine whether the dicentric chromosomes followed a Poisson distribution. We confirmed that most doses followed a Poisson distribution, except for 3 Gy of S1. In the case of S1 3 Gy, we tried to exclude up to 4 × outliers; however, this increased data loss and the remaining data did not follow the Poisson distribution. Poor sample conditions may explain the lower average number of chromosomes and a higher standard deviation in 3 Gy of S1 than in other samples. However, “Filtered” data represented 28.21% of the total data, which is lower than the average loss rate of 34.33%. Thus, our future research will focus on the confounding factors affecting the data.
We successfully constructed a dose-response curve using the general linear model recommended by AIEA 2012 because our data, except for 3 Gy in S1, followed the Poisson distribution. The absorbed radiation dose should be quickly estimated to treat victims of a radiation accident. Therefore, the estimated dose should be categorized into four ranges (1-2, 2-4, 4-6, and > 6 Gy) for rapid estimation, and information for medical treatment should be provided18. We irradiated the blood with 0.5, 1, 2, 3, and 4 Gy doses and calculated the estimated dose to determine whether the dose-response curve constructed using the automatic system could accurately reflect clinical triage. In all the samples, the actual dose was within the 95% confidence intervals of the estimated dose, and the difference between the estimated and actual doses was minimal. Therefore, the medical treatments were successfully categorized. However, this study did not include the construction of a dose-response curve or estimating doses higher than 4 Gy. The limited metaphase cell yield at high doses challenged obtaining a sufficient number of samples. However, a method must be established for samples with exposures above 4 Gy to comply with the clinical triage system.
Furthermore, estimating doses at exposures below 1 Gy necessitates greater precision. According to Korea’s Enforcement Decree of the Nuclear Safety Act (Article 2 and attached Table 1)37, the dose limits for radiation workers are 50 mSv per year and 100 mSv in five consecutive years. When the dose limit is exceeded, the person is classified as a “person with an abnormal reading,” and safety measures should be taken. Therefore, to prepare for the risk of radiation exposure accidents, a method to estimate doses below 0.1 Gy is required in the radiation work environment. The dicentric chromosome assay identifies a low threshold dose of 0.1 Gy38 and requires scoring more than 5000 cells for statistically significant results17, which is difficult to achieve using conventional scoring methods. Therefore, an automatic scoring system can compute legal norms in the event of radiation workplace accidents. However, this study did not evaluate whether it is possible to estimate the exposure at doses greater than 4 Gy and less than 0.5 Gy; further research is warranted to confirm these aspects.
In summary, we constructed an effective dose-response curve using an automated system and calculated the estimated dose for the unknown sample to validate the outcome. In this study, we confirmed that accurate and rapid dose assessments could be facilitated by obtaining results using an automatic system without human review.