Large-scale EM Analysis of the Drosophila Antennal Lobe with Automatically Computed Synapse Point Clouds

TL;DR

This paper presents a large-scale analysis of the Drosophila antennal lobe using automated synapse detection to create synapse point clouds, aiding in brain region segmentation and cross-modal registration.

Contribution

It introduces a methodology for large-scale synapse detection and analysis in EM datasets, improving region selection and cross-modal registration in connectomics.

Findings

Extensive synapse detection enables large-volume analysis.

Synapse point clouds assist in identifying brain compartments.

Methodology supports cross-modal registration between light and EM images.

Abstract

The promise of extracting connectomes and performing useful analysis on large electron microscopy (EM) datasets has been an elusive dream for many years. Tracing in even the smallest portions of neuropil requires copious human annotation, the rate-limiting step for generating a connectome. While a combination of improved imaging and automatic segmentation will lead to the analysis of increasingly large volumes, machines still fail to reach the quality of human tracers. Unfortunately, small errors in image segmentation can lead to catastrophic distortions of the connectome. In this paper, to analyze very large datasets, we explore different mechanisms that are less sensitive to errors in automation. Namely, we advocate and deploy extensive synapse detection on the entire antennal lobe (AL) neuropil in the brain of the fruit fly Drosophila, a region much larger than any densely annotated to date. The resulting synapse point cloud produced is invaluable for determining compartment boundaries in the AL and choosing specific regions for subsequent analysis. We introduce our methodology in this paper for region selection and show both manual and automatic synapse annotation results. Finally, we note the correspondence between image datasets obtained using the synaptic marker, antibody nc82, and our datasets enabling registration between light and EM image modalities.