Weakly Supervised 3d Image Segmentation in Fluorescence Microscopy Using Maximum Intensity Projections

In recent years, biomedical research has greatly benefited from the developments of deep learning techniques for image and data analysis. Widespread adoption is, however, hampered by the annotation effort required to train deep learning algorithms. To alleviate this laborious task for 3D image data, we propose a solution based purely on 2D maximum intensity projections. By utilising "flip" symmetry of the data and a specifically designed loss function based on super-Gaussians, we demonstrate how a 3D semantic segmentation task can be solved using only a single annotated maximum intensity projection image for each 3D data sample, compared to training using full 3D ground truth annotations. The effectiveness of our approach is demonstrated using simulated images of virus particles and experimental data of radiation induced foci imaged using confocal microscopy.