LICONN (light-microscopy-based connectomics)
A Novel Approach to Brain Mapping Employs Light Microscopy to Provide Unprecedented Detail
Researchers from the Institute of Science and Technology Austria (ISTA) and Google Research have revealed a novel approach of “mapping” the brain with light microscopes. The goal of this method, known as LICONN (light-microscopy-based connectomics), is to greatly simplify the difficult process of brain mapping, or connectomics. New findings concerning the brain and neurological disorders are expected to be accelerated by the discovery.
Google Research has been working on connectomics for more than ten years, employing Artificial Intelligence to accurately map the connections between each brain cell. The physical wiring pattern between neurones and their molecular properties determine the brain’s amazing capacity for processing information. Researchers must decipher the spatial structure of the brain, precisely locate cellular elements such as axons and dendrites, resolve synaptic connections, and allocate them to particular neurones in order to comprehend how the brain functions. This calls for intensive cellular tagging and volumetric imaging at nanoscale resolution.
In the past, expensive, specialist electron microscopes (EM) have been the main tool used to achieve the nanoscale resolution required for dense, synapse-level circuit reconstruction. Even though EM-based technologies have greatly advanced connectomics, they still have a lot of drawbacks, especially when it comes to how simple it is to retrieve molecular information about the tissue under study. Specific molecules can be seen well using light microscopy, but dense, synapse-level circuit reconstruction has proven impossible because of resolution, contrast, and volumetric imaging restrictions.
By combining thorough deep-learning-based segmentation and analysis with a specially designed tissue expansion approach, LICONN solves these difficulties. A specific technique was created by ISTA researchers to increase brain tissue while maintaining its cellular composition. The distances between cellular features are effectively increased by this expansion, making it possible to capture nanoscale details like molecules, cells, and their connections using common, standard light microscopes like spinning-disc confocal microscopes that would otherwise require super-resolution techniques or electron microscopy.
In order to implement the LICONN expansion technique, tissue must be embedded in a swellable hydrogel. In contrast to optical super-resolution, LICONN uses this hydrogel expansion to improve resolution. The technique achieves an expansion factor of about 16 times by using a high-fidelity iterative hydrogel expansion approach. This translates to an effective resolution of about 20 nm laterally and about 50 nm axially when using a light microscope objective with a high numerical aperture. In addition to helping to homogenise the refractive index, the hydrogel embedding makes it easier to acquire extended volumes laterally and along the z-axis something that other super-resolution techniques may find challenging.
In order to prepare the sample, mice are infused with a fixative solution containing hydrogel monomers, which equips cellular molecules with vinyl residues that co-polymerize with the hydrogel. After that, brains are cut, collected, and treated with multifunctional epoxide chemicals such as glycidyl methacrylate (GMA) and glycerol triglycidyl ether (TGE) to stabilise biomolecules and functionalise proteins more widely for hydrogel anchoring. Compared to alternative anchoring techniques, this epoxide treatment highlighted synaptic characteristics and enhanced cellular ultrastructure. The roughly 16-fold expansion of the mechanically strong triple-hydrogel-sample hybrids makes handling and prolonged imaging easier.
In order to recreate the cells and their relationships from the light microscope data, Google Research provided its suite of open-source image processing and artificial intelligence technologies. The raw data is processed and stitched together using automated techniques such as SOFIMA (scalable optical flow-based image montaging and alignment) to fuse overlapping subvolumes into seamless bigger volumes after imaging, which can be done with a standard spinning-disc confocal microscope.
Larger volumes are subsequently analysed using deep learning-based segmentation techniques that were modified from EM connectomics. For automatic neural structure segmentation, flood-filling networks (FFNs), which are renowned for reaching cutting-edge segmentation accuracy on connectomic datasets, were specifically trained and put to use. Even at the expense of early splits, the segmentation pipeline is made to minimise improper merges between neurites. These are then fixed by automated agglomeration and thorough manual proofreading. Segments are automatically categorised into discrete subclasses, such as glia, dendrites, and axons, using semantic segmentation using a neural network model.
In addition to structural mapping, spatially resolved molecular information can be measured directly and simultaneously with LICONN. Immunolabelling for particular proteins accomplishes this. In order to incorporate molecular information directly into synapse-level reconstructions, researchers employed immunolabelling to identify excitatory (SHANK2, bassoon, PSD95, VGLUT1) and inhibitory (gephyrin, bassoon) post-synapses and pre-synapses. Compared to EM, this capacity is a significant benefit. The examination of connections is further streamlined by automated synapse recognition techniques, which occasionally involve deep learning prediction of molecule positions.
Several methods were used to validate LICONN’s capabilities. When compared to ground truth produced from sparse fluorescent labelling using eGFP, manual tracing of neuronal structures in LICONN data demonstrated great consistency (low error rates for axons, high accuracy for spine identification). Following proofreading, the automatic FFN segmentation produced minor errors affecting axons or dendritic trunks and high edge correctness (95.6%). Both immunolabelling-based and deep learning-based synapse detection methods demonstrated good fidelity (F1 score > 0.9) when verified against human annotations. Additionally, LICONN’s statistical findings on neural connection agreed with earlier EM observations.
Mouse brain tissue, including areas like the hippocampus and primary somatosensory cortex, has been mapped using this technique. Similar to earlier EM datasets, researchers showed imaging volumes of about 1 million µm³ at the native tissue scale. A long-term objective is to scale LICONN to image larger volumes, including an entire mouse brain. Iterative block-face imaging and sectioning of the enlarged hydrogel is a potential approach for axial scaling that enables smooth volume fusion at deeper depths. Axial fusion of volumes spanning 205 µm was used to illustrate this.
According to its description, LICONN is a simple technique that offers integrated structural and molecular characterization across cell types, spatial scales, and brain regions. Although it was created to recreate intricate tissue structures like the brain, it should be widely applicable to other organs and systems that require high-resolution tissue analysis. One important aspect of LICONN is its accessibility, which is fuelled by the use of standard light microscopy hardware and protocols that aren’t essentially more complicated than current expansion techniques. Utilizing custom code and previously implemented frameworks, the deep-learning analysis tools are open source.
In conclusion, LICONN provides a major breakthrough in connectomics by allowing direct integration of molecular information and dependable, synapse-level reconstruction of brain circuits using light microscopy. LICONN has the potential to transform neuroscience research by facilitating routine connectomic studies in a greater number of labs and speeding up discoveries about the brain and its disorders by lowering the barrier to entry for high-resolution brain mapping.
News source via Google
