Chromosome image segmentation based on class imbalance skeleton semi-supervised model

Abstract

Karyotyping, encompassing chromosome identification, segmentation, and classification, is pivotal in diagnosing cancer-related and other significant global health issues. Major challenges in karyotyping are the reliance on large volumes of annotated data, and the handling of overlapping and touching chromosomes, which often result in the neglect of smaller and less numerous chromosomes. To overcome the limitations regarding these two issues, we propose an advanced semi-supervised chromosome segmentation method—CI3SM (Class Imbalance Skeleton Semi-Supervised Model). CI3SM enhances model consistency and accuracy on unlabeled data by leveraging cross-pseudo supervision and a dual-path skeleton data augmentation strategy. To tackle issues of low segmentation accuracy for small chromosomes and slow category training speed, CI3SM incorporates a CIMM (Class Imbalance Mitigation Module) and a CISCM (Class Intelligent Speed Control Module). Experimental results demonstrate that CI3SM consistently surpasses several state-of-the-art methods in chromosome segmentation tasks, achieving substantial improvements across key performance metrics. Notably, CI3SM realizes a 0.586% enhancement in the Dice coefficient, a 0.897% reduction in Average Surface Distance (ASD), and increases of 0.818%, 0.692%, and 0.654% in the Jaccard index, recall, and F1 score, respectively. These results underscore CI3SM’s capability to deliver superior segmentation outcomes by more effectively harnessing unlabeled data and implementing refined strategies tailored to address the complexities associated with small chromosome segmentation.