Speaker
Description
Quantitative analysis of bioimaging data often depends on the accurate segmentation of cells and nuclei. This is both especially important and especially difficult for the analysis of highly multiplexed imaging data, which can contain many input channels. Current deep learning-based approaches for cell segmentation in multiplexed images require simplifying the input to a small and fixed number of channels, discarding relevant information in the process. Here, we first describe a novel deep learning strategy for nucleus and cell segmentation with fixed input channels and show that it outperforms the most widely used current methods across a range of public datasets, both in terms of F1 score and processing time. We then introduce a novel deep learning architecture for generating informative three-channel representations of multiplexed images, irrespective of the number or ordering of imaged biomarkers. Using these two novel techniques in combination, we set a new benchmark for the segmentation of cells and nuclei on public multiplexed imaging datasets. To maximize the usefulness of our methods, we provide open-source implementations for both Python and QuPath.
| Authors | Thibaut Goldsborough*, Peter Bankhead, Hakan Bilen, Andrew Filby, Alan O`Callaghan, Fiona Inglis, Pau Carrillo Barberà |
|---|---|
| Keywords | Cell Segmentation, Nucleus Segmentation, Multiplexed Imaging, Fluorescence imaging, Deep Learning, Open Source |
