Weakly Supervised Nucleus Segmentation Using Point Annotations via Edge Residue Assisted Network

Abstract

Cervical cell nucleus segmentation can facilitate computer-assisted cancer diagnostics. Due to the obtaining manual annotations difficulty, weak supervision is more excellent strategy with only point annotations for this task than fully supervised one. We propose a novel weakly supervised learning model by sparse point annotations. The training phase has two major stages for the fully convolutional networks (FCN) training. In the first stage, coarse mask labels generation part obtains initial coarse nucleus regions using point annotations with a self-supervised learning manner. For refining the output nucleus masks, we retrain ERN with an additional constraint by our proposed edge residue map at the second stage. The two parts are trained jointly to improve the performance of the whole framework. As experimental results demonstrated, our model is able to resolve the confusion between foreground and background of cervical cell nucleus image with weakly-supervised point annotations. Moreover, our method can achieves competitive performance compared with fully supervised segmentation network based on pixel-wise annotations.