Active learning based workflow for automated 3D-EM segmentation
Large 3D electron microscopy images require labor-intensive segmentation for further quantitative analysis. Recent deep learning segmentation methods automate this computer vision task, but require large amounts of labeled training data. We present BioSegment 2.0, a turnkey platform for experts to automatically process their imaging data and fine-tune segmentation models. It provides a user-friendly annotation experience, integration with familiar microscopy annotation software and a job queue for remote GPU acceleration. Various active learning sampling strategies are incorporated, with maximum entropy selection being the default. For mitochondrial segmentation, these strategies can improve segmentation quality by 10 to 15% in terms of intersection-over-union score compared to random sampling. Additionally, a segmentation of similar quality can be achieved using 25% of the total annotation budget required for random sampling.