Radiation biodosimetry uses biological markers, mostly chromosomal aberrations such as dicentric chromosomes (DCs) and chromosome translocations to measure radiation doses that individuals are exposed to. A biologically absorbed dose may be different from the physical dose measured by instruments due to variability between individuals. Similarly, radiation of a certain physical dose can produce different biological exposures due to underlying physiological or environmental conditions. Knowledge of the biological dose is of particular importance for both diagnosis and treatment.
The DC assay is the gold standard of the World Health Organization (WHO) and International Atomic Energy Agency (IAEA) for assessing biological radiation exposure in people. It was the first assay recommended by the IAEA and WHO for radiation dose assessment. DC frequency is relatively stable for approximately 4 weeks after radiation exposure1 and their quantitative correlation with emitted radiation dose is accurate, which make DCs the ideal biomarker. The relationship between radiation dose (referenced in Gray [Gy] units), and DC frequency (referenced as number of DCs per cell) can be expressed as a linear-quadratic function.
The cytogenetic DC assay has been the industry standard for about 55 years2. It has been performed manually, requiring 1 - 2 days to analyze microscope data from a single blood sample. Several hundred to several thousand images are needed to accurately estimate radiation exposure depending on the dose3. At doses exceeding 1 Gy, IAEA recommends a minimum of 100 DCs be detected. Examination of 250 - 500 metaphase images is common practice in biodosimetry cytogenetic laboratories. For samples with exposures <1 Gy, 3,000 - 5,000 images are suggested due to the lower probabilities of DC formation. In either case, it is a labor-intense task.
Cytogenetic biodosimetry laboratories create their own in vitro radiation biodosimetry calibration curves before assessing biological doses in test samples. Blood samples from normal, control individuals are exposed to radiation and lymphocytes are then cultured and prepared for metaphase chromosome analysis. Using these samples, biological doses received are calibrated to the known physical doses emitted by a standard radiation source. After metaphase cell images are recorded, experts examine images, count DCs and calculate DC frequencies for each sample. A calibration curve is built by fitting a linear-quadratic curve to the DC frequencies at all doses. Then, exposures in test sample from individuals can be inferred by matching the DC frequencies to the calibrated doses on the curve or by specifying them in the corresponding linear quadratic formula.
We have automated both the detection of DCs and dose determination to expedite this procedure using software. Automated Dicentric Chromosome Identifier and Dose Estimator (ADCI) uses machine learning-based image processing techniques to detect and discriminate dicentric chromosomes (DCs) from monocentric chromosomes (MCs) and other objects and automates radiation dose estimation. The software aims to significantly reduce or eliminate the necessity for manual verification of DC counts and to accelerate dose estimation through automation. It has been developed with the involvement of reference biodosimetry laboratories at Health Canada (HC) and Canadian Nuclear Laboratories (CNL). Their feedback will ensure that performance will continue to meet IAEA criteria for this assay.
The software performs the following functions: 1) filtering DCs and selecting optimal metaphase cell images for analysis, 2) chromosome recognition, DC detection, and DC frequency determination, and 3) estimating radiation dose from dose-calibrated, cytogenetic radiation data. This software processes groups of metaphase images from the same individual (termed a sample), counts the number of DCs in each using image processing techniques, and returns the estimated radiation dose received by each sample in units of Grays (Gy).
The software has been designed to handle a range of chromosome structures, counts, and densities. However, the algorithm performs optimally in metaphase images containing a near complete complement of well-separated, linear chromosomes4. Images containing highly overlapped sets of chromosomes, multiple cells, incomplete metaphase cells, sister chromatid separation, nuclei, non-chromosomal objects, and other defects can reduce the accuracy of the algorithm. Dedicated image selection models and other object segmentation thresholds can filter out the majority of sub-optimal images and false positive DCs.
Dicentric chromosome detection is performed when an image is processed. The algorithm attempts to determine which objects in an image are chromosomes and then locates the two regions most likely to be centromeres on each chromosome. Then, a series of different Support Vector Machine (SVM) learning models distinguish chromosomes as either DCs or normal, monocentric chromosomes. The SVM models differ in sensitivity and specificity of DC detection (see Step 3.1.4 below), which can affect the DC frequencies that are determined in a sample.
ADCI processes sets of Giemsa- (or DAPI-) stained metaphase digital images (in TIFF or JPG format) for one or more samples. The software analyzes DCs in both calibration samples and test samples. The physical doses (in Gy) of calibration samples are known and are used in the generation of a calibration curve. The physical and biological doses of individuals with unknown exposures are inferred by the software from the machine-generated calibration curve. Although laboratories use comparable techniques, the calibration curves from different laboratories often vary3. Both calibration curve and test samples from the same laboratory should be processed for accurate dose estimation in test samples.
This software offers speed, accuracy and scalability which addresses the productivity required to handle an event in which many individuals must simultaneously be tested. It was developed from 2008 - 20174,5,6,7,8,9,10,11,12,13. Using recent computer hardware, this desktop PC software can process and estimate radiation dose in a patient sample of 500 metaphase genome equivalents in 10 - 20 min 4. The code is based on a set of proprietary image segmentation and machine learning algorithms for chromosome analysis. Expert analysis of each chromosome exposed to 3 Gy radiation gave comparable accuracies to ADCI. In a set of 6 samples of unknown exposures (previously used in an international proficiency exercise), the software estimated doses within 0.5 Gy of the values obtained by manual review of the same data by HC and CNL, fulfilling the IAEA's requirements for triage biodosimetry. Furthermore, inter-laboratory standardization and ultimately reproducibility of dose estimates benefit from having a common, automated DC scoring algorithm. Nevertheless, the software permits customization of image filtering and selection criteria, enabling differences in chromosome preparation methods and radiation calibration sources to be taken into account.
This software is a graphical user interface (GUI)-based system which analyzes sets of chromosome images containing Giemsa (or DAPI)-stained metaphase cells for abnormalities that result from exposure to ionizing radiation. The image sets are digitally photographed with a light (or epifluorescent) microscope system and each set corresponds to a different sample. The software utilizes image processing techniques to detect and discriminate DCs from MCs and other objects. Empirically-derived segmentation filters then automatically eliminate false positive DCs without affecting true DCs. Finally, the software automatically filters out undesirable images based on various image properties found poor quality metaphase images with precomputed (or user-specified) image selection models. These images include those containing excessive or insufficient numbers of "noisy" objects, multiple overlapping chromosomes, images lacking metaphase chromosomes, excessive numbers of sister chromatids4. The automatically curated image data are used to generate dose calibration curve from samples of known radiation dose and are used to estimate exposures of test samples exposed to unknown doses.
Output of the software can be viewed and saved as: 1) text-based output viewed in the console, 2) plots which can be saved as images, and 3) reports in HTML format. Many aspects of the software are customizable to suit the specific needs of different laboratories. Individual laboratories usually provide both calibration and test samples prepared and collected based on the cytogenetic protocol validated in that laboratory. This maintains uniformity of sample preparation and allows calibration curves generated from calibration samples to be meaningfully applied to test samples derived using the same protocol. Calibration curves may also be created from either curve coefficients or DC frequencies at defined doses. The most accurate dose estimates are obtained by filtering out lower quality images and false positive DCs (FPs). Selection of optimal image subsets within each sample is accomplished using 'Image selection models' that eliminate subpar images which tend to introduce FPs. A series of pre-validated models are included with the software, however additional models with customized thresholds and filters can be created and saved, by the user.
Once the software successfully loads, the main graphical user interface (GUI) is presented (see Figure 1). From this interface, samples, each consisting of a folder of metaphase cell image files, may be selected and processed to identify DCs, calibration curves may be created and compared, and radiation exposure dose of samples may be determined.
Figure 1: The Major Sectors of the Graphical User Interface Include: a list of samples (1), a list of calibration curves (2), the process queue (3), which monitors the status of DC detection in each set of images of each sample, a plot display (4), which summarizes statistical or other quantitative properties of a set of images in samples or calibration curves, and a console (5) which contains descriptive text as outputs of each operation performed by the program. Please click here to view a larger version of this figure.