基于渐进式半监督迁移学习的染色体实例分割方法

A Chromosome Instance Segmentation Method Based on Progressive Semi-Supervised Transfer Learning

摘要: 针对医学图像标注成本高昂的问题，本文提出一种融合迁移学习与半监督学习的染色体实例分割方法。基于混合级联网络(HTC)，引入基于伪标签的半监督机制与迁移学习，通过少量标注数据训练初始模型，并利用其生成大量未标注数据的高置信度伪标签，经迭代训练有效扩充训练集，显著降低对标注数据的依赖。实验表明，该方法在显著减少标注需求的同时，有效提升了目标定位性能。然而，在高精度掩码评估标准下(如Mask AP75)，伪标签生成的掩码边界仍存在不足，完全依赖伪标签将导致分割性能下降。综上所述，本文方法在节约标注成本的同时，保障了模型的核心性能，为染色体核型自动分析提供了可行方案。未来工作将致力于提升伪标签的边界生成质量，进一步减少对标注数据的依赖。

Abstract: To address the challenge of high annotation costs in medical imaging, this paper proposes a chromosome instance segmentation method that integrates transfer learning and semi-supervised learning. Building upon the Hybrid Task Cascade (HTC) network, our approach incorporates a pseudo-label-based semi-supervised mechanism alongside transfer learning. The process begins by training an initial model with a small amount of annotated data. This model is then used to generate high-confidence pseudo-labels for a large volume of unannotated data. Through iterative training, the training set is effectively expanded, significantly reducing reliance on extensively annotated datasets. Experimental results demonstrate that our method substantially diminishes annotation requirements while effectively enhancing object localization performance. However, under high-precision mask evaluation standards such as Mask AP75, the mask boundaries generated by pseudo-labels exhibit shortcomings. Complete dependence on pseudo-labels leads to a decline in segmentation performance. In summary, the proposed method achieves significant savings in annotation costs while maintaining the model’s core performance, offering a practical solution for automated chromosome karyotype analysis. Future work will focus on improving the boundary quality of pseudo-labels to further reduce dependency on annotated data.