Instance Segmentation from Volumetric Biomedical Images Without Voxel-Wise Labeling

Meng Dong, Dong Liu, Zhiwei Xiong, Xuejin Chen, Yueyi Zhang, Zheng-Jun Zha, Guoqiang Bi, Feng Wu

Abstract

Volumetric instance segmentation plays a significant role in biomedical morphological analyses. The improvement of segmentation accuracy has been accelerated by the progress of deep learning-based methods. However, such methods usually rely heavily on plenty of precise annotation, which is time-consuming and may need some expert knowledge to label manually. Although there are several studies focusing on weakly supervised methods in order to save the labeling cost, previous approaches still more or less require voxel-wise annotation. In this paper, we propose a weakly supervised instance segmentation method that needs no voxel-wise labeling. Our approach takes advantage of two advanced techniques: one is the popular proposal-based framework (Faster R-CNN in this paper) for instance detection, and the other is the peak response mapping (PRM) for finding visual cues of instances. Then a new thresholding method combines detected boxes and visual cues to generate final instance segmentation results. We conduct experiments on two biomedical datasets, one of which is a large-scale mouse brain dataset at single-neuron resolution collected by ourselves. Results on both datasets validate the effectiveness of our proposed method.
BibTeX
@inproceedings{dong2019instance,
  author = {Meng Dong and Dong Liu and Zhiwei Xiong and Xuejin Chen and Yueyi Zhang and Zheng-Jun Zha and Guoqiang Bi and Feng Wu},
  title = {{Instance Segmentation from Volumetric Biomedical Images Without Voxel-Wise Labeling}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2019},
  doi = {10.1007/978-3-030-32245-8\_10},
}

Loading related content...

Figure