Deep neural networks for image-based cell analysis and profiling

Microscopy and image-based analysis are important tools in a biologist's arsenal. Its use is ubiquitous throughout areas such as cell biology and in biomedical research and is an important technique for bioengineers in the development and analysis of novel biomaterials. For example, comprehensive and accurate characterisation of relationships between substrate topographies and cell growth in biomaterial engineering requires complex data analysis strategies on large collections of microscopy data. Advances in the technology and a growing collection of software tools are enabling researchers to collect vast quantities of complex multidimensional data. Naturally, automated computational analysis of such large datasets has attracted considerable attention in recent years (see Caicedo et al. (2017) for a review).
The current approaches to analysis however are slow, requiring significant manual intervention from the user and resulting in a bottleneck in the overall analysis procedure thus hindering truly high-throughput microscopy. A researcher may have to set and fine tune anywhere between 50 and 100 parameters for a single analysis. For a typical high throughput experiment, many hours can be spent optimising the pipeline to achieve reasonable results. This inevitably leads to biases as the human element is subjective and often based on ad-hoc decisions which can influence the conclusions drawn from the analysis.
It is therefore desirable to extend the current computational procedures to not only improve the speed at which the data can be processed but to improve the accuracy and reliability of the analysis, allowing truly high-throughput microscopy-based cell analysis.
Research statement and methodology
I propose that adopting a state-of-the-art machine learning approach would allow building of robust, interpretable, high performance models for image-based cell analysis, overcoming the issues with classical, hand-tuned algorithms and pipelines, streamlining the analysis procedure and improving the quality of the scientific results.
To achieve this, I expect to work towards the following objectives over a three year period:
1. Adapting and extending current end-to-end deep neural networks in a Bayesian neural network structure. This will provide improved and robust segmentation and single cell recognition. Furthermore, the Bayesian treatment will allow a systematic quantification of any uncertainty related to the results.
2. Designing, training and evaluating machine learning models that map from images to interpretable single cell profiles suitable for human interpretation and dissemination.
3. Designing, training and evaluating a Bayesian neural network that maps from images to functional properties of the cells as measured, for example, by gene expressions.
4. Investigating the potential issues with interpretability of neural networks in the biological domain and developing new visualisation tools to mitigate this. This can be achieved by ensuring that the networks learn disentangled representations which are inherently easy to interpret, thus possibly circumventing the need for objective.
The new techniques will be evaluated and benchmarked against state-of-the-art methods (such as CellProfiler) on custom datasets or publicly available datasets, which are also suitable for initially training the (mostly supervised) machine learning models.