MaksChroms: An open-source CNN-based tool for dose prediction based on dicentric chromosome assay

One of the most widely used biological dosimetry methods for estimating ionizing radiation dose is the dicentric chromosome assay. Accurate dose estimates are essential for effective medical response in radiological and nuclear emergencies and have applications in molecular epidemiology. Manual scoring of chromosomal aberrations is time-consuming and subject to scorer variability. We introduce MaksChroms, a novel software for automated analysis of dicentric chromosomes in microscopic images, using an approach that separately detects chromosome candidates and centromeres. Image sets of metaphases from the MULTIBIODOSE project (14,269 in total), including 5,444 expert-labeled images, were used for training and testing. Samples were from blood exposed to gamma radiation (0.25–5 Gy) and unexposed controls. Two models were developed: model A, based on R-CNN and EfficientNet, performs instance segmentation and classification of chromosomes; model B, based on U-Net, performs semantic segmentation to locate centromeres. Outputs are combined in a final model. The root mean square error of dose reconstruction for this model was 0.42 Gy. Differences in scoring were also observed between human experts. MaksChroms supports manual correction of detected chromosomes and classifications, and enables dose calculation using built-in or user-defined calibration curves, providing an efficient automated or semi-automated tool for dicentric chromosome assay. • MaksChroms automates dicentric chromosome assay using deep learning models (R-CNN, EfficientNet, U-Net). • The tool was used for dose prediction across a 0–5 Gy gamma radiation range. • Combines instance and semantic segmentation models to detect chromosomes and localize centromeres. • MaksChroms is open-source and supports both automated scoring and manual correction for biological dosimetry.