Author: Adayabalam Balajee, Elijah Berberette, Maria Escalona, Dray Gentry, Chester R. Ramsey, Terri Ryan 👨🔬
Affiliation: ORAU, Thompson Proton Center, University of Tennessee 🌍
Purpose:
Dicentric chromosomes, characterized by two centromeres on a single chromosome, are key biomarkers in biological dosimetry for quantifying ionizing radiation exposure. However, manual detection by laboratory experts is labor-intensive and time-consuming. This project leverages artificial intelligence (AI) to automate dicentric chromosome detection, using convolutional neural networks (CNNs) and image processing to enhance speed and consistency in biodosimetric analysis.
Methods:
High-resolution cytogenetic images were prepared from irradiated samples using standard techniques, yielding a dataset of 772 annotated dicentric chromosomes and an equal number of non-dicentric counterparts. A two-step AI pipeline was implemented: Meta’s Segment Anything Model segmented chromosomes from metaphase images, and a CNN classified each chromosome as dicentric or non-dicentric. The model was trained and validated using a 90/10 split, optimizing performance over 600 epochs on an A100 GPU with a learning rate of 1E-4.
Results:
The CNN achieved a best validation accuracy of 95.45%. Additional experiments demonstrated a 2.79% improvement in accuracy when the training dataset increased from 252 to 1,390 samples, highlighting the model’s ability to capitalize on additional data. These findings show the potential for further performance gains with larger datasets and optimized algorithms.
Conclusion:
This AI-driven pipeline automates dicentric chromosome detection, significantly reducing manual effort and improving reliability. The results confirm the importance of robust training data and advanced AI methods in enhancing classification accuracy. Future efforts will focus on refining segmentation models, integrating state-of-the-art classifiers, and enhancing explainability to maximize practical utility in biodosimetry. By modernizing this critical process, the project sets a new standard for radiation exposure assessment and emergency preparedness.