Weakly Supervised Cell Segmentation From Volumetric Brain Images

The processing and analyses of high-resolution microscopic brain images can con-tribute to the research of brain,including learning about the brain structures from cell level and understanding the relationships between brain functions and structures.The knowledge about brain mechanism can provide guidance and inspiration for the devel-opment of artificial intelligence.For the morphological analyses of brain images from cell level,image segmentation can be a powerful tool.Image segmentation is an im-portant research field in computer vision area,and it can serve for higher tasks such as image analyses,statistics et al.The popular deep learning based segmentation methods have excellent performance on both natural images and biomedical images,but most of them rely heavily on large amounts of high-costly manual labeling.The problem is more serious especially for biomedical images like our brain data,where the images are volumetric and need special softwares to view.The labeling may also require expert knowledge,and it is difficult to obtain pixel-wise manual annotation for segmentation tasks.Therefore,we propose to complete morphological cell segmentation with weakly supervised learning strategies,which can achieve dataset preparation and segmentation tasks with much lower labeling cost.Targeting on the weakly supervised segmentation task,we build volumetric mouse brain and macaque brain image datasets for cell segmentation.Both datasets have weakly annotated training set and precisely annotated testing set,and the weak labels are easy to acquire.Meanwhile,we propose three weakly supervised segmentation ap-proaches,and all these approaches do not rely on any voxel-wise labeling.The main work of this article can be summarized as below:1.We build a mouse brain image dataset for soma segmentation,where the training set has inscribed globules to label the position and rough size of soma.Based on this dataset,we firstly propose a weakly supervised soma semantic segmenta-tion method.The algorithm uses globules to divide voxels into foreground,back-ground and uncertainty,and the segmentation model learns to segment soma with these weak pixel-wise labels and our proposed sample balancing strategy.Then we propose a weakly supervised soma instance segmentation approach,where we train a soma detector with bounding boxes transformed from the globules.During testing phase,the Peak Response Mapping(PRM)module is introduced to the de-tector,and the instance segmentation can be accomplished with detected boxes,PRM results and our designed thresholding method.We conduct experiments on different datasets to compare our methods with existing advanced works,and consequently both the quantitative and visualized results demonstrate their effec-tiveness and generalization.2.We build a macaque brain image dataset for cell segmentation,where the training set has mixed points annotation and rough mask annotation to label the position and shape of cells.Based on this dataset,we propose a novel weakly supervised cell instance segmentation approach.The approach uses our designed appear-ance loss and instance loss to learn a peak-shape probability map,then the map is used to segment instances by peak detection and watershed algorithm.Our experiments demonstrate that our method achieves excellent segmentation preci-sion on both our macaque brain dataset and a public nuclei dataset.Our method outperforms the existing methods with the same labeling cost quantitatively and visually.