Few-Shot Learning for Annotation-Efficient Nucleus Instance Segmentation

• DOI: 10.48550/arxiv.2402.16280
• Publication Date: 2024-02-25
• Authors: Ming Yu; Zihao Wu; Jie Yang; Danyi Li; Yuan Gao; Changxin Gao; Gui-Song Xia; Yuanqing Li; Liang Li; Jin-Gang Yu

Nucleus instance segmentation from histopathology images suffers the extremely laborious and expert-dependent annotation of nucleus instances. As a promising solution to this task, annotation-efficient deep learning paradigms have recently attracted much research interest, such as weakly-/semi-supervised learning, generative adversarial etc. In paper, we propose formulate perspective few-shot (FSL). Our work was motivated by that, with prosperity computational pathology, an increasing number fully-annotated datasets are publicly accessible, hope leverage these external assist on target dataset which only has very limited annotation. To achieve goal, adopt meta-learning based FSL paradigm, however be tailored in two substantial aspects before adapting our task. First, since novel classes may inconsistent those dataset, extend basic definition (FSIS) generalized (GFSIS). Second, cope intrinsic challenges segmentation, including touching between adjacent cells, cellular heterogeneity, etc., further introduce structural guidance mechanism into GFSIS network, finally leading unified Structurally-Guided Generalized Few-Shot Instance Segmentation (SGFSIS) framework. Extensive experiments couple accessible demonstrate SGFSIS can outperform other baselines, semi-supervised simple transfer comparable performance fully supervised less than 5% annotations.