Transfer Learning Analysis of Image Processing Workflows for Electron Microscopy Datasets

Neuroscientists are collecting Electron Microscopy (EM) datasets at increasingly faster rates. This modality offers an unprecedented map of brain structure at the resolution of individual neurons and their synaptic connections. Despite sophisticated image processing algorithms such as Flood Filling Networks, these huge datasets often require large amounts of hand-labeled data for algorithm training, followed by significant human proofreading. Many of these challenges are common across neuroscience modalities (and in other domains), but we use EM as a use case because the scale of this data emphasizes the opportunity and impact of rapidly transferring methods to new datasets. We investigate transfer learning for these workflows, exploring transfer to different regions within a dataset, between datasets from different species, and for datasets collected with different image acquisition techniques. For EM data, we investigate the impact of algorithm performance at different workflow stages. Finally, we assess the impact of candidate transfer learning strategies in environments with no training labels. This work provides a library of algorithms, pipelines, and baselines on established datasets. We enable rapid assessment and improvements to processing pipelines, and an opportunity to quickly and effectively analyze new datasets for the neuroscience community.