Unsupervised Segmentation and Statistical Shape Analysis for PanVision

PanVision is a new microscopy technique that uses light microscopy to provide images of cells with a resolution comparable to electron microscopy (EM). It builds on expansion microscopy and can be used both to make the whole proteome of a cell visible but also in combination with different antibody staining to identify specific structures within the cell. The image acquisition is fast, and compared to EM, the preparation of samples is simpler, and an ordinary confocal microscope can be used instead of specialised equipment. This will enable researchers to investigate cellular ultrastructures and the distribution of proteins within the cell more easily. To assist in this process, I will build on recent advances in computer vision to develop unsupervised segmentation methods that identify these structures automatically, and I am going to apply statistical analyses to quantify the shapes and distribution of objects within the resulting segmentations.

My work will be structured into the following major aims:

Aim 1: Unsupervised segmentation. While unsupervised segmentation methods are currently a very active field of research in computer vision, these new approaches are mainly developed for natural images. Therefore, the first step of this project is to adapt state-of-the-art unsupervised segmentation methods for images produced by PanVision. To do so, I am going to investigate which components of current approaches are most susceptible to the challenges coming with microscopy data like a high noise to signal ratio and low contrast, and I am going to research how to adapt the loss functions and architectures to make them more robust.

Aim 2: Incorporating prior knowledge. Moreover, most new unsupervised segmentation methods are developed for images of a wide range of different contexts and therefore do not take specific prior knowledge into account. However, in this application, we have a rough idea of the resulting cluster sizes and know that we can expect a high degree of self-similarity between the cellular ultrastructures. Therefore, I am going to research ways to incorporate this prior shape information into the segmentation algorithms.

Aim 3: Statistical shape analysis. Finally, I will use the resulting segmentation masks both in 2D and 3D to analyse the shape [9, 10] and distribution of different structures within the cell-like mitochondria. Moreover, it was possible to prove that our methods work for HeLa cells. I aim to apply the developed techniques to tissues in different biological contexts.

In this project, I will closely collaborate with the group led by Joerg Bewersdorf (Yale University, New Haven, USA), who is developing the PanVision protocol. There is already sufficient data available for Aim 1 and 2, and additional images are in process as well. Additional collaborations with the St Johnston Lab (University of Cambridge) to apply this imaging protocol to intact tissues are planned.

This work will enhance unsupervised segmentation algorithms to make them more reliable in the context of microscopy data. Additionally, I aim to enable new biological discoveries and enhance the understanding of fundamental biological processes by providing imaging pipelines that allow the automatic segmentation and shape analysis of cellular ultrastructures.

This project falls within the EPSRC Healthcare Technologies/medical imaging research area.