Abstract
After DNA replication, two chromatids with identical genetic information are formed in organisms; these are called sister chromatids. Sister chromatid exchange (SCE) is a recombination event between genetically equivalent sequences. Since no genetic changes occur before or after SCE, it does not have detrimental effects on cells. However, SCE is induced by treatments with DNA-damaging agents or by the knockout of genes encoding recombination suppressors, such as BLM. Therefore, SCE has been widely performed in the fields of basic medicine and biology. Typically, SCE analysis is performed manually by experts using a microscope. This process is time-consuming and involves subjective judgement by experimenters. To address these limitation, we propose a system that detects each chromosome from whole chromosome images and automatically measures the number of SCEs. The proposed method employs two deep-learning-based models. The first is the Mask Region-based Convolutional Neural Network (Mask R-CNN) for detecting single chromosomes. Mask R-CNN identifies the boundaries of multiple objects and performs instance segmentation, enabling the separation and detection of touching or overlapping objects of the same-class. The second is the Vision Transformer (ViT), an adaptation of the Transformer model -originally developed for natural language processing-for computer vision tasks. The ViT is used to classify SCEs. Finally, the number of SCEs is measured using image processing and clustering algorithms. This combined method enables fully automatic SCE analysis and successfully measures the number of SCEs with an accuracy of 84.10%.