Join us to find out more about research taking place in AI for Science across the Accelerate Science community.
Details of future talks are available on Talks@Cam
Lunch provided, please register to attend via this form so we can confirm catering arrangements.
Using machine learning approaches to automate the diagnosis of small intestinal biopsies
Professor Elizabeth Soilleux, Department of Pathology, University of Cambridge
Histopathology is a clinical discipline in which pathologists look down microscopes at biopsies, to make a diagnosis. Nowadays many departments scan the glass microscope slides, so that pathologists can view them on a screen. This opens up the possibility of automating histopathological diagnosis, particularly as there are international shortages of pathologists, leading to backlogs and delays. We chose the duodenum (small intestine) as a starting point due to its low medicolegal risk.
We apply a series of processing steps, including artefact removal, division into small tiles and colour normalisation, before applying a multiple instance learning approach to biopsy classification, leading to accuracy >97% against a carefully curated ground truth. In order to facilitate adoption of this technology by pathologists and their clinical colleagues, we are now working on making our categorisation processes more explainable. We have spun out a company, Lyzeum Ltd, to progress our software to market.
Machine learning models guide viral discovery in museum bat collections
Maya M. Juman, PhD Student, Department of Veterinary Medicine
Natural history museum collections are valuable but underutilized resources for viral discovery, offering opportunities to test hypotheses about pathogen occurrence across space, time, and taxonomic groups. We developed trait-based machine learning models of bat host suitability to guide viral screening of 1821 tissues in a museum collection. Our coronavirus and paramyxovirus predictive models performed with 79% and 92% predictive accuracy, respectively, and we used these models to generate ranked lists of suspect “novel” host species for screening. For the first time, we recovered these viruses from archived museum tissues, confirming three novel coronavirus host species and three novel paramyxovirus host species (3% and 33% prediction success rate, respectively). These sequences included a SARS -like coronavirus from an Angolan bat collected in June 2019, suggesting that viruses with epidemic potential may be more widespread in sub-Saharan Africa than previously believed. This case study lays out a framework for using predictive machine learning models to unlock pathogen data hidden in historical specimens.
These seminars are open to members of the University of Cambridge. For further details, please email accelerate-science@cst.cam.ac.uk.