Chromosome localisation and segmentation in fluorescence microscopy images

Segmenting chromosomes in fluorescence microscopy images is essential step for the precise definition of chromosomes allowing the robust detection of chromosome aberrations. Classical segmentation methods like thresholding and region growing methods have some limitations, especially when dealing with complex chromosome structures and the problem of overlapping and touching chromosomes. Recent advances in machine learning, particularly deep learning, offer promising solutions to these challenges. In this study, we propose an innovative pipeline that combines YOLO for initial chromosome detection and U-Net for precise segmentation, even in cases of overlapping or touching chromosomes. Our experiments, conducted on diverse datasets, including those containing challenging scenarios, showcase the efficacy of our approach. The U-Net architecture delivers high accuracy, recall, specificity, F1-score, and IoU metrics. The integration of YOLO and U-Net yields promising outcomes, achieving a precision of 90.46%, a recall of 97.44%, and an F1-score of 93.72% on a test dataset.