Precise Alignment of Serial Section Electron Microscopy Images and Analysis of Neural Circuits

Abstract

Neural circuits provide a means to test how the structure of a brain is related to its function, but reconstructing large neural circuits is challenging. A promising path is to image a brain with serial section electron microscopy, use automatic methods to produce a preliminary reconstruction, then manually correct errors in the result via proofreading to recover circuits from the data. The imaging process introduces significant physical deformation that causes the automatic methods to make many errors, limiting the size of the circuits that can be reasonably proofread, so it’s crucial to remove the deformation in a step called alignment. Parameters in existing approaches to alignment can be difficult to tune. By augmenting existing approaches with human-in-the-loop intervention, I was able to produce a very accurate alignment of a five teravoxel mouse cortex dataset. This approach failed to correct deformation caused by cracks and folds, and the amount of human effort involved made it prohibitively expensive to process larger datasets. To overcome these limitations, I developed a new alignment pipeline that uses deep optic flow models to more precisely estimate the deformation. Using this new alignment pipeline, I was able to accurately align much larger datasets, including a 125 teravoxel whole fly brain and a 1.4 petavoxel cubic millimeter of mouse cortex. These alignments enabled low-error automatic reconstructions that are now being proofread with manageable human effort to produce some of the largest neural circuits to date. Improvements in the reconstruction technology, such as this alignment pipeline, are alleviating the bottleneck to analyze more neural circuits.I analyzed a circuit reconstructed from the terascale mouse cortex dataset, finding that the sizes of synapses in a shared connection between layer 2/3 pyramidal neurons have a correlated binary component, and that the network of these pyramidal neurons is well-described by a configuration model. Both results may have implications for how to constrain biological models of memory or neural network organization.