High-precision automated reconstruction of neurons with flood-filling networks

Abstract

Reconstruction of neural circuits from volume electron microscopy data requires the tracing of cells in their entirety, including all their neurites. Automated approaches have been developed for tracing, but their error rates are too high to generate reliable circuit diagrams without extensive human proofreading. We present flood-filling networks, a method for automated segmentation that, similar to most previous efforts, uses convolutional neural networks, but contains in addition a recurrent pathway that allows the iterative optimization and extension of individual neuronal processes. We used flood-filling networks to trace neurons in a dataset obtained by serial block-face electron microscopy of a zebra finch brain. Using our method, we achieved a mean error-free neurite path length of 1.1 mm, and we observed only four mergers in a test set with a path length of 97 mm. The performance of flood-filling networks was an order of magnitude better than that of previous approaches applied to this dataset, although with substantially increased computational costs.

Automated Reconstruction of a Serial-Section EM Drosophila Brain with Flood-Filling Networks and Local Realignment

“…Integrated with an established FFN segmentation pipeline [6], LR and ISS allowed dense segmentation of the entire adult fly brain dataset (Fig. 1), with a very low rate of merge errors (processes erroneously connected to one another), and with segment lengths sufficient to assist many common tracing workflows and analyses.…”

“…1), with a very low rate of merge errors (processes erroneously connected to one another), and with segment lengths sufficient to assist many common tracing workflows and analyses. LR with recheck reduced merge errors by an order of magnitude, while ISS particularly reduced split errors (processes erroneously disconnected from one another), thus tripling the "expected run length" of resulting segments [6]. The forty trillion voxel segmentation will be publicly released to support further efforts in Drosophila circuit understanding and connectomic algorithm development.…”

Workflow-Driven Distributed Machine Learning in CHASE-CI: A Cognitive Hardware and Software Ecosystem Community Infrastructure

“…The combination of these resources innovated the CONNECT workflow in several ways, including: 1) Rapid data transfer between a FIONA and Ceph cloudbased object storage, which is distributed across the Pacific Research Platform (PRP): This is performed harnessing Unidata's Thematic Real-time Environmental Distributed Data Services (THREDDS) [19] server maintained on a node within the PRP allowing Kubernetes to transfer data into the Nautilus system. 2) Applying a new object segmentation algorithm: Instead of using MATLAB functions that use a single CPU to do the object segmentation, a new algorithm, Flood-Filling Network (FFN) [20], was used. The FFN was applied to NASA data using 50 NVIDIA 1080ti GPUs based on Tensorflow.…”

“…The model selected to do rapid segmentation was the FFN model and was adapted to do segmentation of NASA data. CNN is able to separate objects within a 3D volume of spatial data or images by using a deep stack of 3D convolutions [20]. The network is trained to take an input object mask within the networks field of view to infer the boundaries of the objects.…”

Processing Volumetric Data for Correlative Analysis: An Anecdote from a Core Facility

“…Prominent labs devoted to EM study often have teams of annotators, coding staffs, and computer clusters to create and apply advanced data processing tools. These teams have made great strides in the field scientifically, even involving high profile companies such as Google [2]. However, many of these advancements remain inaccessible to labs which have few staff and limited resources.…”