.png){kind=icon}
.png){kind=icon}
Return
Scribble-supervised active learning for microscopy instance segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2024.129169.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate manual annotation for large-scale microscopy instance segmentation is time-consuming and expensive. Recently, weakly supervised learning has been proposed to reduce the annotation cost by exploiting weak annotation, such as scribble. However, existing weakly supervised learning methods work on randomly selected annotation and cannot fully exploit hard samples, resulting in unsatisfying performance. In this paper, we propose a two-stage, scribble-supervised active learning method for microscopy instance segmentation to automatically exploit hard samples and reduce the annotation cost. First, we propose across pseudo-label strategy to warm up the segmentation network leveraging a small number of scribbles, which generates the initial instances. Second, we propose an instance-based region selection strategy via active learning to exploit hard samples and further improve the segmentation performance. Specifically, we iteratively select hard regions and correct segmentation errors in these regions with a few scribbles to update the network parameters. Extensive experiments on 2D and 3D microscopy datasets demonstrate the superiority of our proposed method over existing weakly supervised learning and active learning methods.
Keywords:
Microscopy instance segmentation
Active learning
Weakly supervised learning