TL;DR
A University of Southampton team has built an open-source AI platform, CenSegNet, that separates two centrosome defects in breast cancer which pathology has long counted as a single abnormality. Working with the city’s hospital across 127 patients, the study found one of the two tracked with aggressive disease and shorter survival. The tool is not ready for routine clinical use.
One marker turns out to be two
Centrosomes are the structures that organise cell division and hold a cell’s shape. When they go wrong, genetic errors accumulate — which is why their abnormality has been a recognised signature of cancer for more than a century. Studying them in actual patient tissue, at any useful scale, has been the problem.
The distinction CenSegNet draws is between cells that have acquired too many centrosomes and cells whose centrosomes have grown unusually large. Until now those were treated as the same defect. Analysing over 330,000 individual cells, the team found the two can occur in separate regions of one tumour and may be doing different jobs entirely.
The clinical signal sits with size rather than number. Tumours carrying more enlarged centrosomes were likelier to show markers of aggression: a higher grade, spread into neighbouring lymph nodes, and particular genetic changes. Patients whose tumour cores held fewer of them lived longer overall.
Why the release matters as much as the finding
“CenSegNet allows us to analyse these defects at single-cell resolution across entire tumours and uncover patterns that were previously impossible to see,” said Dr Salah Elias, who led the work. He put the practical destination as new biomarkers and, in time, treatment matched to the individual tumour.
The team has released it free as open-source software, and has already shown it working on tissue from the kidney, colon and appendix.
Looking forward
Nothing here reaches a patient soon — the researchers are clear it is not hospital-ready — and the honest framing is that this is a research instrument that may eventually inform how risk is assessed.
The licensing decision is the part worth noting. Most clinical AI arriving in the NHS comes through procurement, where the model is the product and the hospital rents access to it. Southampton has given the platform away and kept the biology as the contribution. For an NHS repeatedly warned about vendor dependence in diagnostic tooling, a capable pathology model that any trust or research group can simply run is a different kind of asset — and a rarer one than the steady flow of commercial announcements suggests.