RESUMO
The precise identification of chromosomal abnormalities based on microscopy images in clinics heavily relies on manual analysis by karyotype experts, which is expensive and labor-intensive. Therefore, there is an urgent need for automatic classification technologies. However, current deep learning-based methods require a large amount of annotated data to achieve satisfactory performance. Additionally, there are domain shifts between datasets obtained from different instruments and clinical agencies, which hinders the utilization of existing public data resources. To overcome these challenges, we propose a novel ensemble learning framework called ChromEDA. This method incorporates soft pseudo-label learning, adversarial learning, and angle classification learning strategies to mitigate the domain shift. To evaluate its performance, we conducted cross-domain classification experiments using both public and private datasets. The results demonstrate that ChromEDA effectively addresses the domain shift issue and outperforms existing methods in cross-domain chromosome classification tasks. RESEARCH HIGHLIGHTS: The research transferred an unsupervised cross-domain algorithm into the chromosome microscopic image processing area, which can support fast and efficient deployment in equipment such as microscopies. To our knowledge, it is the first domain adaption application in chromosome image analysis. The research designed an ensemble learning framework-based chromosome classification model to alleviate the domain shift between different domains. It is a combination of soft pseudo-label learning, adversarial learning, and self-supervised angle recognition, which can significantly reduce the annotation requirements for chromosomal image processing. The results find that for chromosomes sharing the same centromere position category, banding information may contribute more to within-group classification. Thus, banding pattern information is encouraged to be considered for more precise classification of chromosomes.
