Unsupervised Annotation of Complex 3D BioMedical Data.

We aim to develop new methods for the unsupervised annotation of complex 3D biomedical data. This will allow for the efficient labelling of 3D models of cellular data for use by either human biomedical researchers or future machine learning models. Deep learning methods are increasingly being used to analyse complex data however these methods require large quantities of labelled training data to be effective. Such training data is often difficult and expensive to acquire. The recent availability of large volumes of 3D data poses challenges in its annotation as this is a significantly more difficult task than the 2D equivalent. This presents a barrier to research. The solution we propose is using unsupervised annotation. While labelling data from scratch is prohibitively expensive, requiring a significant time investment from a team of experts, combining and manipulating labels obtained via unsupervised methods provides a more realistic approach. Through this method we can reduce the time and monetary investment required to produce labelled training data for further research and provide significant utility for future studies.
Furthermore, Virtual Reality (VR) tools have been developed in recent years which offer an immersive view of 3D cellular models allowing doctors and researchers intuitive, detailed views of cells they wish to analyse. Of particular interest is the is the MiCellAnnGELo cell painting tool which offers manual annotation of 3D biomedical data while working with 4D timeseries data. The immersive nature of VR is well suited to this task as it can allow experts a close-up interactive view of cells to streamline the annotation process. MiCellAnnGELo makes use of the Unity gaming engine, a platform designed for the navigation of 3D environments, to allow for cell models to be easily manipulated by the end user. This provides an easy-to-use platform for the annotation of biomedical data that can alleviate much of the prohibitive nature of 3D annotation.
We will develop new functionality to facilitate the dynamic labelling of cell components. As fully automating the annotation process may jeopardise accuracy, we aim to incorporate suggestions into the tool to identify likely cell structures and components within a model. In doing so we can streamline the annotation process further and increase the utility of these visualisation tools, allowing researchers to better analyse complex structures within cells while providing a visual aid to facilitate a better understanding of the cellular data.
We will make use of standard frameworks and models to achieve our goals. For the development of our neural networks, we intend to use the PyTorch framework due to its ease of use and flexibility which will allow us to quickly prototype and iterate on our research. Should problems arise with PyTorch, the TensorFlow and Keras frameworks are available as alternatives.
For our models, we plan to base our work upon Graph convolutional Networks for surface data and 3D Convolutional Neural Networks for volume data. Cluster based Generative Adversarial Networks will be used to simulate additional training data. Currently clustering has been primarily applied to 2D datasets where the number of expected clusters is already known. An important component of our project is therefor to develop strategies for combining networks which are either based on different numbers of clusters or using clustering methods that do not depend on a fixed number. An important contribution will then be to implement the clustering of data at different resolutions using subsampling that can preserve the topology of surface meshes, and superpixel approaches for volume data. We must ensure that our model can train on and annotate data of different resolutions and is not limited to a single size of input. By adapting and combining these models we achieve a strong foundation for our research and obtain a basis for the segmentation and annotation of our own 3Ddata