Identifying weak signals in inhomogeneous neuronal images for large-scale tracing of neurites

Abstract

Reconstructing neuronal morphology across different regions or even the whole brain is important in many areas of neuroscience research. Large-scale tracing of neurites constitutes the core of this type of reconstruction and has many challenges. One key challenge is how to identify a weak signal from an inhomogeneous background. Here, we addressed this problem by constructing an identification model. In this model, empirical observations made from neuronal images are summarized into rules, which are used to design feature vectors that display the differences between the foreground and background, and a support vector machine is used to learn these feature vectors. We embedded this identification model into a tool that we previously developed, SparseTracer, and termed this integration SparseTracer-Learned Feature Vector (ST-LFV). ST-LFV can trace neurites with extremely weak signals (signal-to-background-noise ratio <1.1) against an inhomogeneous background. By testing 12 sub-blocks extracted from a whole imaging dataset, ST-LFV can achieve an average recall rate of 0.99 and precision rate of 0.97, which is superior to that of SparseTracer (which has an average recall rate of 0.93 and average precision rate of 0.86), indicating that this method is well suited to weak signal identification. We applied ST-LFV to trace neurites from large-scale images (approximately 105 GB). During the tracing process, obtaining results equivalent to the ground truth required only one round of manual editing for ST-LFV compared to 20 rounds of manual editing for SparseTracer. This improvement in the level of automatic reconstruction indicates that ST-LFV has the potential to rapidly reconstruct sparsely distributed neurons at the scale of an entire brain.