FocAnnot: Patch-Wise Active Learning for Intensive Cell Image Segmentation

Bo Lin, Shuiguang Deng, Jianwei Yin, Jindi Zhang, Ying Li, Honghao Gao

Abstract: In the era of deep learning, data annotation becomes an essential but costly work, especially for the biomedical image segmentation task. To tackle this problem, active learning (AL) aims to select and annotate a part of available images for modeling while retaining accurate segmentation. Existing AL methods usually treat an image as a whole during the selection. However, for an intensive cell image that includes similar cell objects, annotating all similar objects would bring duplication of efforts and have little benefit to the segmentation model. In this study, we present a patch-wise active learning method, namely FocAnnot (focal annotation), to avoid such worthless annotation. The main idea is to group different regions of images to discriminate duplicate content, then evaluate novel image patches by a proposed cluster-instance double ranking algorithm. Instead of the whole image, experts only need to annotate specific regions within an image. This reduces the annotation workload. Experiments on the real-world dataset demonstrate that FocAnnot can save about 15% annotation cost to obtain an accurate segmentation model or provide a 2% performance improvement at the same cost.