P03-03 Automation of Chromosomal Aberration Analysis Using Deep Neural Networks

The chromosomal aberration assay is a key cytogenetic test in genetic toxicology, widely used to assess structural chromosomal damage induced by chemical, physical, or environmental agents. Despite its importance in regulatory toxicology and risk assessment, the assay remains labor-intensive, requiring manual microscopic evaluation that introduces variability and limits reproducibility. Advancements in artificial intelligence, particularly deep neural networks (DNNs), offer a promising solution for automating aberration detection. We investigated the feasibility of an automated workflow using Metafer (MetaSystems, Germany), integrating multiple DNNs to detect metaphases, capture images, and identify chromosomal aberrations. Automated metaphase detection using DNNs is well established, with additional potential to refine mitotic index calculations involving a dedicated DNN. Our current aberration detection network is in development. Initial training involved 669,075 images of normal and aberrant chromosomes, including dicentric, polycentric, and ring chromosomes, as well as fragments and chromatid gaps. Internal testing on unclassified images yielded an accuracy of 97.84%. Further expansion of the dataset, particularly incorporating chromatid-type aberrations, is planned to enhance performance. Preliminary results indicate that detection sensitivity and precision scale with the volume of training images. No inherent limitations to automation were observed, provided sufficient training data is available. With an expected increase in classified images by September, we anticipate presenting updated findings at the Eurotox 2025 conference.