Artificial Intelligence and generative data augmentation for automated chromosomal aberration detection in cytogenetic FISH imaging

Following accidental exposure to ionizing radiation, it is necessary to refine the assessment of the dose received to conduct effective sorting of asymptomatic victims. Among the available dosimetric techniques, biological dosimetry based on cytogenetic imaging consists of counting chromosomal aberrations (CA) within circulating lymphocytes. These aberrations can be unstable (ex. dicentrics) or more persistent (ex. translocations). The latter are more suitable to perform retrospective dosimetry several years after exposure and can be identified through color colocalizations in fluorescence in situ hybridization (FISH) imaging.The counting procedure is long and tedious, requiring trained biologists and to our knowledge, no automated solutions are presently available. The present study aims to develop a deep learning-based workflow for detection of stable CA in FISH images. Thanks to recent advances in computer vision, we deployed state-of-the-art object detection models allowing the localization and classification of fluorescent chromosomes. We faced two major challenges: the limited number of annotated data and the rarity of the translocation at the metaphase scale. The idea was then to develop a semi-supervised approach taking advantage of both spatial annotations and unannotated data.Faster-RCNN, an object detection neural network, is relevant here because it can simultaneously locate and classify chromosomes. Additionally, we studied classification at the chromosome level via ResNet models, after segmentation by U-Net models. Both approaches lead to accurate detection of fluorescent chromosomes, but the distinction of aberrations remains subject to improvement.To overcome the restriction in the number of annotated data, we explored the possibility of creating synthetic data using generative diffusion models. More specifically, image-to-image models can transform an input image to match the characteristics of a target image. In this work, we generated artificially translocated chromosomes from the blue channels (DAPI equivalent) of our unannotated images.In conclusion, the present work provides promising results about deep learning model deployment for an automated chromosomal aberration detection in cytogenetic imaging.