Automating Analysis of Chromosomal Aberrations Using a Neural Network Algorithm

Abstract Potential of using a neural network algorithm to automate the analysis of chromosomal damage is investigated. The research is based on the YOLOv8 model, which through anchor-free architecture and multi-scale detection enables identification of target objects of varying sizes. The dataset consists of metaphase plates images obtained from blood cells of male Macaca Mulatta monkeys. The training set consists of 721 images containing 30,831 chromosomal objects (including damaged chromosomal objects) categorized into 9 classes. Model performance was evaluated using standard metrics (mAP@0.5, Precision, Recall) and real-world task testing through the construction of the dose-effect curve. The distribution of chromosomal aberrations depending on radiation dose and the model’s predictions with the Poisson distribution were compared. The model’s test results were compared with experimental data obtained by the researcher.