Validation of a deep neural network workflow for fully automated dicentric chromosome analysis in biological dosimetry

Abstract Artificial intelligence has recently found its way into biological dosimetry. This study validated the initial version of a deep neural network (DNN) workflow implemented by MetaSystems to detect dicentric and polycentric chromosomes in microscopy images. This was performed through parameter optimisation, calibration curve generation and independent testing. Optimisation of thresholds for the class score parameter defining the aberration classes was performed and brought DNN scoring closer to manual scoring ( 60 Co dataset) for low doses by reducing the number of false-positive dicentrics, while balancing the influence of false-negative detections, especially for higher doses. Although slightly higher, the DNN achieved an acceptable yield of dicentrics for the control. A comparison of the calibration curves showed that the DNN curve was between manual and semi-automatic curves. DNN curves based on two independent test datasets (X-ray) from different laboratories showed very good agreement, with dose estimations matching the reference doses for all 12/13 test samples precisely, even for low doses and for triage scoring. This DNN-workflow enables the establishment of calibration curves for different radiation qualities and successfully performs dose estimations. It could replace semi-automated scoring and enhance the capacity of biological dosimetry laboratories in the event of large-scale radiological or nuclear incidents.