Point-Based Weakly Supervised 2.5D Cell Segmentation

Volumetric microscopic images show cells in their natural state and solve various problems inherent to 2D projections. The development of competent Deep Learning methods to segment cells in 3D images is, however, held back by the extremely time-consuming and error-prone process of manual ground truth creation. To reduce the burden of manual annotation in 3D, we propose a weakly supervised 2.5D cell segmentation approach that learns to accurately predict 3D segmentation masks from weak, slice-wise point labels. We show that even a single point per cell as ground truth label is sufficient to train a network on par with a fully supervised model that outperforms a top contender of the ISBI Cell Tracking Challenge, and with performance close to that of a fully 3D approach while requiring only a fraction of the resources. The slice-wise, point-based annotation scheme, not only reduces the time required to annotate 3D cell datasets by an estimated factor of 6, but also simplifies the complex and error-prone process of manually segmenting cells using 3D software.