Stable and Unstable Chromosomal Aberration Detection System for Biodosimetry

Biological dosimetry relies on chromosome analysis to predict the radiation levels to which the human body has been exposed. However, biological dosimetry typically requires manual processing by trained experts, a time-consuming and costly task, which highlights the need for automation. Although recent advances in deep learning have addressed various problems, research in the field of chromosomal aberration analysis is limited, with studies focusing only on certain karyotypes. In particular, studies have focused only on Giemsa-stained data, limited to the classification of unstable chromosomal aberrations. Thus, we proposed a Diffusion and Mapping-based Chromosomal Aberration Detector (DM-CAD) for fluorescence in situ hybridization (FISH) data. DM-CAD is an integrated deep-learning framework designed to detect both stable and unstable chromosomal aberrations within a single FISH-based analytical workflow. Based on the assumption that the challenges in chromosome counting and aberration detection are attributable to the limited number and poor quality of chromosome images, we developed preprocessing models and data augmentation techniques to overcome these problems. As a result of DM-CAD, we achieved an accuracy of 0.9850 and an F1 score of 0.9925 in the task of counting chromosomes, and an accuracy of 0.7725 and an F1 score of 0.8716 in the task of detecting chromosomal aberrations, demonstrating significant performance improvements.