Introduction
The microscopy prototyping team works within the EM STP to develop new methods to help us deal with the ever-increasing amount of electron microscopy image data that is produced at the Crick.
We aim to identify and alleviate some of the bottlenecks in existing workflows by creating new hardware and software systems. The team consists of members with backgrounds in physics, biology, microscopy and computer science.
Take a three-dimensional tour of our microscopy prototyping lab.
Take a tour of the @EM_STP @TheCrick - the microscopy prototyping lab - Spherical Image - RICOH THETA
Hardware development
The development of correlative imaging methods has allowed us to combine different kinds of information from different microscope types to gain additional insight into scientific questions. Our hardware development projects focus on correlative light and electron microscopy (CLEM) workflows.
New in-resin fluorescence (IRF) protocols allow fluorescent probes to survive the chemically harsh electron microscopy sample preparation, which means that we can perform fluorescence microscopy after resin-embedding. This significantly simplifies the process of combining light and electron microscopy data.
Ultramicrotome light microscope (ultraLM)
Since fluorescence microscopy gives us functional information about a sample, we can use it to help us when serial sectioning on an ultramicrotome (an ultramicrotome is the device that lets us slice the sample into extremely thin slices, down to around 70nm. Think about it like a very specialised bacon slicer!). By only collecting sections that exhibit fluorescence, we can greatly reduce the amount of electron microscopy we need to do by post-selecting which sections to image.
Our device mounts onto the ultramicrotome and takes a fluorescence snapshot of the blockface at each cutting stroke, allowing the later targeting of individual regions of interest in the electron microscope.
ultraLM
more
Miniature light microscope (miniLM)
Another technological advance in recent years has been the ability to perform serial block face volume electron microscopy in an integrated system, with the ultramicrotome incorporated within the microscope.
In our Gatan 3View system, the gap between the sample and detectors is just 3mm, so to achieve fluorescence imaging of the blockface, we designed a system with a miniaturised objective lens (just 2.8mm in diameter) where the image is relayed to the outside world via a fibre bundle.
MiniLM
even more
Infrared branding
Sometimes it's not possible to use a sample preparation that's compatible with in-resin fluorescence. In these situations, we can create a proxy for the functional information by using a high-power infrared laser to 'brand' marks into the tissue. This branding mark is visible in the electron microscope, allowing us to quickly relocate a region of interest that we found during fluorescence imaging. Our team built a custom scanning microscope (based on a multiphoton microscope) that enables us to both find regions of interest by fluorescence and to mark those regions with branded landmarks.
Software development
Etch a Cell
A vital part of any microscopy workflow is image analysis. As data sets are getting larger and larger, the resources required to process and analyse the data become a major limiting factor. We are investigating different approaches to help overcome this bottleneck, particularly as much of the analysis performed in electron microscopy requires significant amounts of user interaction.
To extract meaning from our images, we perform a task known as 'segmentation', where we delineate regions of the images into, for example, the nucleus, the mitochondria, the endoplasmic reticulum and many other sub-cellular structures. Once the images are labelled in this way, we are able to quantitatively analyse them, for example to understand whether a drug treatment affects the ability of a diseased cell to divide and replicate.
One method that has been showing a lot of promise in many image analysis tasks at the moment is deep learning, in particular convolutional neural networks which have emerged as the best method for a wide range of classification and segmentation tasks.
To provide enough training data for this type of technique, we have set up a 'citizen science' project called Etch a Cell, where we ask people to trace lines onto structures of interest from their web browser. When you first log in you are given a tutorial and within a few minutes you can be contributing to our research!