A deep learning-based automatic chromosome segmentation method for metaphase cell images

Chromosomes, cellular structures carrying genetic information, are fundamental to cellular genetics research, with abnormalities potentially causing serious genetic disorders. This makes chromosome screening an essential component in prenatal care for monitoring fetal development and detecting potential genetic abnormalities early in pregnancy. However, manual chromosome analysis is highly inefficient and increasingly burdensome for medical laboratories as prenatal examinations become more common in modern healthcare practices worldwide. Current karyotype analysis systems face significant challenges in three primary tasks: detection, segmentation, and orientation correction of chromosomes, particularly when chromosomes overlap in microscopy images, compromising analytical accuracy and resulting in incomplete or inaccurate karyotype images.

To address these critical challenges, we developed MCSegNet, an automated deep learning model specifically designed for precise chromosome segmentation. Our innovative three-stage architecture effectively combines Swin Transformer as the backbone network for feature extraction, Hybrid Task Cascade for sophisticated detection and segmentation tasks, and test time augmentation for enhanced model robustness and generalization capabilities.

Utilizing a comprehensive dataset of 30,055 chromosome images for training and validation, our model achieved exceptional performance metrics: 98.9 % accuracy and precision, 99.7 % recall, and 99.3 % Dice coefficient, significantly surpassing existing methods reported in the literature.

Through extensive validation using real clinical data, we confirmed the model's practicality and robustness in clinical applications, promising to enhance both the accuracy and efficiency of prenatal diagnosis for expectant mothers while substantially reducing the workload of medical professionals in laboratory settings.