Deep Learning-Driven Biodosimetry for High-Throughput Radiation Triage via Chromosomal Aberration Detection

SUMMARY Radiological and nuclear incidents can expose large and diverse populations to harmful levels of ionizing radiation, requiring rapid and accurate dose estimation to guide effective triage. Traditional biodosimetry methods, such as the Dicentric Chromosome Assay (DCA), focus solely on dicentrics, are labor-intensive, and take 72-96 hours to complete, making them limited for large-scale triage during emergencies. This project proposes to develop an Artificial Intelligence (AI)-powered cytogenetic biodosimetry platform leveraging Deep Learning, specifically the YOLOv9 neural network recently released in early 2024. YOLO, which stands for You Only Look Once, is a Convolutional Neural Network (CNN) designed for fast object detection. The latest version, YOLOv9, introduces enhanced mean Average Precision (mAP) of up to 55.6%, reduced inference time, and greater computational efficiency, thanks to innovations like the Generalized Efficient Layer Aggregation Network (GELAN) and Programmable Gradient Information (PGI). These features make YOLOv9 ideal for detecting a wide range of chromosomal aberrations—dicentrics, rings, and multi-centromeric chromosomes (tricentrics, tetracentrics, pentacentrics) across dose ranges from 0-5 Gray (Gy) for photons and 0-2 Gy for neutrons, considering both high and low Linear Energy Transfer (LET) radiation. To complement the chromosomal aberration identification using YOLOv9, the project will integrate classification algorithms, such as decision trees, random forests, k-means, and DBSCAN, to categorize individuals into low, moderate, or high exposure levels, accounting for individual variability in radiosensitivity, age, sex, and DNA repair capacity. This approach aims to challenge the traditional 2 Gy triage threshold by enabling more personalized and accurate interventions. A curated dataset of high-resolution cytogenetic images annotated for various aberration types and dose-response relationships will be developed to train and validate the AI models. The overall goal is to improve the speed and accuracy of dose estimation during mass exposure events, enhancing public health response. Project outcomes include a validated AI-powered biodosimetry platform and a curated dataset shared with the broader community to support AI validation and strengthen emergency preparedness.