Deep learning methods for the automated segmentation and phenotyping of cells in biological tissues

Abstract

This thesis aims to advance the field of tissue phenotyping by developing and refining computational tools to accurately analyse and quantify microscopy images of biological tissues. Tissue phenotyping involves characterizing cellular composition, architecture, and function. This is essential for understanding the biological processes underpinning multiple diseases, including cancer. With the increasing complexity of microscopy data, it is crucial to rely on quantitative methods that can detect, identify, and characterize individual cells and their interactions.

The initial focus of this research is to improve the accuracy of cell and nucleus segmentation methods, as this is often a critical step in subsequent analyses. We introduce InstanSeg, a novel, embedding-based instance segmentation algorithm optimized for accuracy, efficiency and portability. We demonstrate state-of-the-art accuracy on six publicly available datasets as well as a substantial reduction in processing time, allowing for the analysis of very large images. We also enable InstanSeg to be deployed within popular open-source software tools, ensuring that it can be widely used.

We then extend InstanSeg for the simultaneous segmentation of nuclei and cells in multiplexed fluorescence images. To this end, we develop a neural network architecture for generating three-channel representations of multiplexed images irrespective of the number or ordering of imaged biomarkers. We pair this architecture with InstanSeg and demonstrate state-of-the-art segmentation of multiplexed images on two publicly available datasets.

Furthermore, we demonstrate that InstanSeg can be used as an effective precursor to cell classification by developing an end-to-end cell classification workflow to detect and classify immune cells in kidney biopsies. Finally, we extend our cell classification workflow to multiplexed images allowing for the accurate phenotyping of cells based on biomarker positivity.

Overall, this thesis advances the field of bioimage analysis by providing new computational tools for the segmentation and phenotyping of cells in complex histology images. Our open-source implementations contribute to the improvement and democratisation of state-of-the-art bioimage analysis tools available to biologists.