Quantitative and Real-Time Image Analysis for Advanced Light Microscopy.

In microscopy based research it is common to acquire and analyze microscopy images of cells sampled from populations which have undergone differential treatments. A non-exhaustive number of samples (i.e. images of cells) are often acquired due to experimental limitations that include: restricted acquisition periods, provision of expensive equipment or pressures to move onto the next experiment. Random sampling is a methodology that is used to sample representatively. There are some good reasons however for not sampling randomly. For example, cells maybe expressing a reporter protein and the copy number (quantity) of this molecule maybe too high or too low for subsequent analysis and so preference is shown toward those examples where the expression levels are within a certain range. Or, alternatively, the cells in an area maybe compromised by an artifact, be unhealthy, or too dense or sparse for a particular analysis. There are many factors that will influence the choices of an imaging researcher that are fully justifiable, but it is a problem as to how this information is documented and shared with other scientists. Removing the scientist from the pipeline of acquisition and analysis is counter-productive, a better solution is to provide tools that allow them to document and describe any experimental subjectivity as well as boosting reproducibility through automation. The objectives of this first part of the project (1-4) revolve around solving these issues of documenting, describing and automating image acquisition so that experimental design can be better communicated between users and laboratories.
Furthermore, we would like to develop tools and approaches, which allow better visualization and real-time feedback for advanced approaches so as to better inform imaging scientists as they perform their research and make experimental decisions. The second part of the project relates to this goal, objectives 5-6.

In terms of microscopy analysis, great inroads have been made in terms of impartial and systematic analysis but little has been done to ensure that cells under the microscope are selected impartially. For this project DW will develop algorithms that can statistically quantify and describe cellular appearances utilizing the latest machine learning, computer vision (CV) and signal processing techniques and technologies.
DW has in ongoing work investigated the use of CV algorithms in microscopy and has shown very promising results can be achieved by utilizing object detection convolutional neural networks for cellular detection. DW will extend on his use of neural networks for localizing cells (objectives 1-2) and will also develop methods to statistically describe cellular appearance using networks derived from auto-encoders (objective 3), a type of compression network. Neural networks are implemented in several ways, a popular method is to use Tensorflow and DW is an expert in this language. To effectively train neural networks, powerful GPUs are required. Fortunately the WIMM has various computational facilities and DW has access to two powerful GPU equipped servers.
For developing CV and real-time analysis approaches to work with camera and detector hardware, DW will work toward developing algorithms that can be embedded in miniaturized electronics (objective 4). The Nvidia Jetson TX2 Developer kit is a resource which allows you to create and develop neural networks and distribute them onto small hardware boards. DW will develop and test this hardware technique with microscopy based hardware and algorithms.
For objectives 5 and 6 GPU code will be systematically produced in CUDA and will be made as compatible and distributable as possible. Oxford University has several microscopy software development projects (e.g. in Micron) and it is intended that the software and libraries produced by this project will be made compatible with these other projects.