#electronmicroscopes

LICONN (light-microscopy connectomics)

An innovative brain mapping method uses light microscopy to provide unprecedented detail.

ISTA and Google Research researchers have developed a light microscope method for brain "mapping." LICONN (light-microscopy-based connectomics) simplifies brain mapping, or connectomics, which is tough. The breakthrough should speed brain and neurological illness research.

For almost a decade, Google Research has used Artificial Intelligence to map brain cell connections using connectomics. The brain's incredible information processing capacity is due to neurone circuitry and chemical characteristics. Researchers must understand the brain's spatial organisation, find cellular features like axons and dendrites, resolve synaptic connections, and assign them to neurones to understand how it works. Intensive cellular labelling and nanoscale volumetric imaging are needed.

The nanoscale resolution needed for dense, synapse-level circuit reconstruction has traditionally been achieved with costly, specialist electron microscopes (EM). EM-based technologies have substantially increased connectomics, but they still have several disadvantages, notably in retrieving tissue molecular information. Light microscopy can see specific molecules, but resolution, contrast, and volumetric imaging limits prevent dense, synapse-level circuit reconstruction.

LICONN overcomes these issues with deep-learning-based segmentation and analysis and a customised tissue expansion method. ISTA researchers developed a method to grow brain tissue while preserving its cells. This expansion increases the distances between cellular features, allowing spinning-disc confocal microscopes to capture nanoscale details like molecules, cells, and their connections that would otherwise require super-resolution or electron microscopy.

Tissue must be placed in a swellable hydrogel for LICONN expansion. Instead of optical super-resolution, LICONN employs hydrogel expansion to increase resolution. High-fidelity iterative hydrogel expansion yields a 16-fold expansion factor. A light microscope objective with a high numerical aperture can resolve 20 nm laterally and 50 nm axially. Hydrogel embedding helps homogenise the refractive index and acquire extended volumes laterally and along the z-axis, which other super-resolution methods may struggle with.

Infusing mice with a fixative solution containing hydrogel monomers gives cellular molecules vinyl residues that co-polymerize with the hydrogel. Brains are then sliced, collected, and treated with multifunctional epoxide compounds like GMA and TGE to stabilise biomolecules and functionalise proteins for hydrogel anchoring. Epoxide therapy improved synaptic properties and cellular ultrastructure compared to other anchoring methods. Mechanically robust triple-hydrogel-sample hybrids are simpler to handle and image due to their 16-fold expansion.

Google Research's open-source image processing and AI capabilities recreated cells and their interactions using light microscopy data. After imaging using a spinning-disc confocal microscope, utilising automated methods like SOFIMA (scalable optical flow-based image montaging and alignment), overlapping subvolumes are fused into seamless larger volumes.

Larger volumes are examined using deep learning-based segmentation methods developed from EM connectomics. Flood-filling networks (FFNs), known for their high segmentation accuracy on connectomic datasets, were trained for autonomous neural structure segmentation. Although early splits are sacrificed, the segmentation process minimises erroneous neurite mergers. An automatic aggregation and rigorous hand proofreading solve them. Semantic segmentation employing a neural network model automatically classifies segments into glia, dendrites, and axons.

LiconN measures spatially resolved molecular information directly and concurrently in addition to structural mapping. Specific protein immunolabelling does this. Researchers used immunolabelling to identify excitatory (SHANK2, bassoon, PSD95, VGLUT1) and inhibitory (gephyrin, bassoon) post- and pre-synapses to include molecular information directly into synapse-level reconstructions. Comparing to EM, this capacity is advantageous. Automated synapse recognition, which often uses deep learning to forecast molecular sites, simplifies connection analysis.

Several approaches validated LICONN's capabilities. Manual tracing of neural structures in LICONN data showed remarkable consistency (low axon mistake rates, high spine identification accuracy) compared to ground truth from sparse fluorescent labelling using eGFP. After proofreading, automated FFN segmentation produced minimal axon or dendritic trunk faults and 95.6% edge accuracy. Comparing immunolabelling- and deep learning-based synapse identification algorithms to human annotations showed high fidelity (F1 score > 0.9). Additionally, LICONN's neuronal connectivity statistics matched prior EM findings.

This method has mapped mouse brain tissue, including the hippocampus and primary somatosensory cortex. Researchers found imaging volumes of around 1 million µm³ at the native tissue size, similar to previous EM datasets. Long-term goals include imaging a mouse brain with LICONN. Iterative block-face imaging and sectioning of the larger hydrogel may enable smooth volume fusion at greater depths by axial scaling. This was demonstrated by axial fusion of 205 µm volumes.

LICONN claims to be a straightforward method for integrated structural and molecular characterisation of cells, spatial scales, and brain regions. It was designed to replicate complex tissue structures like the brain, but it should work for other organs and systems that need high-resolution tissue analysis. LICONN is accessible since it uses ordinary light microscopy apparatus and procedures that aren't more difficult than current expansion techniques. Open-source deep-learning analysis tools use bespoke code and frameworks.