Semantic segmentation of microscopic neuroanatomical data by combining topological priors with encoder-decoder deep networks

Abstract

Understanding of neuronal circuitry at cellular resolution within the brain has relied on tract tracing methods which involve careful observation and interpretation by experienced neuroscientists. With recent developments in imaging and digitization, this approach is no longer feasible with the large scale (terabyte to petabyte range) images. Machine learning based techniques, using deep networks, provide an efficient alternative to the problem. However, these methods rely on very large volumes of annotated images for training and have error rates that are too high for scientific data analysis, and thus requires a significant volume of human-in-the-loop proofreading. Here we introduce a hybrid architecture combining prior structure in the form of topological data analysis methods, based on discrete Morse theory, with the best-in-class deep-net architectures for the neuronal connectivity analysis. We show significant performance gains using our hybrid architecture on detection of topological structure (e.g. connectivity of neuronal processes and local intensity maxima on axons corresponding to synaptic swellings) with precision/recall close to 90% compared with human observers. We have adapted our architecture to a high performance pipeline capable of semantic segmentation of light microscopic whole-brain image data into a hierarchy of neuronal compartments. We expect that the hybrid architecture incorporating discrete Morse techniques into deep nets will generalize to other data domains.

Detection and skeletonization of single neurons and tracer injections using topological methods

“…In Step 1, the input image is converted to a density field defined on a 3D grid ρ : K → R. For fMOST data, raw image intensity was treated as the density value subjected to Morse skeletonization. For the STP data, a process-detection step 56 was first applied to segment labelled axon fragments in the high-resolution 2D images. A 3D volume was created to summarize the density of axon fragments within each voxel at lower resolution.…”

“…STP tracer injection dataset. When applying the DM-Skeleton pipeline on the tracer injection dataset, in Step 1, the original image data (see Section Data Collection) was passed through a hybrid deep CNN with topological priors 56 to detect axon fragments. The output of this automated detection process was then manually proofread and corrected.…”

“…The STP dataset was first processed with a topologically motivated convolutional neural network 56 for the detection of the tracers. The network, termed as DM++, takes in whole STP sections and divides them into 512 × 512 pixel tiles.…”

“…For whole-brain tracer injection STP data, a significant portion of the raw images is background (see the example in Figure 1). Hence we first apply the learning-based process-detection module 56 to remove the background and segment the foreground. We further apply a Gaussian filter to smooth the values across the domain.…”

“…The DM pipeline for summarizing multiple-neuron tracer injection datasets has multiple stages. First, the raw image stack is preprocessed to detect or highlight neuronal processes 56 . Then a variant of the Discrete-Morse module 19,21 is employed to produce a graph skeleton containing all potential neuronal trajectories.…”