Researchers at the University of Southampton have built an AI tool that identifies previously undetectable patterns inside breast cancer cells, a development that could help doctors spot high-risk patients and design more targeted treatments. The team analyzed more than 330,000 cells from 127 patients using the platform, called CenSegNet, and published findings that distinguish between two types of centrosome abnormalities that had previously been treated as a single condition.
Two distinct abnormalities, one disease
Centrosomes are structures that help cells divide properly and maintain their architecture. When they malfunction, cells accumulate genetic errors - a common feature of cancer. The Southampton team found that some cancer cells develop too many centrosomes, while others develop centrosomes that are unusually large. These two changes can appear in different parts of the same tumor and may contribute to cancer progression in different ways.
Using CenSegNet, the researchers examined tumor samples at single-cell resolution. The tool processes images with what the team describes as "unprecedented speed and precision," making it possible to study entire tumor samples rather than small fragments.
What the patterns reveal
The study found that tumors with higher numbers of enlarged centrosomes were more likely to show features linked to aggressive disease: higher tumor grade, spread to nearby lymph nodes, and certain genetic changes. Patients with fewer enlarged centrosomes in the center of their tumors tended to have better overall survival.
Dr. Salah Elias, who led the research, said: "For more than a century, centrosome abnormalities have been recognized as a hallmark of cancer, but studying them in patient tissues has been extremely challenging. CenSegNet allows us to analyze these defects at single-cell resolution across entire tumors and uncover patterns that were previously impossible to see."
He added: "This opens the door to developing new biomarkers and, ultimately, more personalized treatment strategies."
Beyond breast cancer
The team has also tested CenSegNet on tissue samples from the kidney, colon, and appendix, suggesting the tool could apply to other cancer types. The software is available free of charge as open-source code, allowing researchers worldwide to use and modify it.
The technology is not yet ready for routine hospital use, but the researchers say it could eventually help clinicians determine which cancers are more likely to grow, spread, or resist treatment. It could also support drug development by identifying tumors with specific weaknesses that new therapies might target.
For developers and IT professionals, the project demonstrates how AI can be applied to complex biological data at scale. The open-source release of CenSegNet means the underlying architecture is available for study and adaptation, offering a concrete example of how machine learning tools are being built for specialized scientific workflows. Those working on similar problems in AI for IT & Development may find the model's approach to image analysis and pattern recognition relevant to their own projects.
Why this matters for IT and development professionals
This research is a working example of AI applied to a high-stakes domain: medical diagnosis. The practical lessons extend beyond oncology. CenSegNet shows how a purpose-built model can process thousands of images, classify subtle variations, and produce results that human experts could not achieve manually. For developers building similar tools, the key takeaway is the importance of domain-specific training data and the value of releasing models as open source to accelerate validation across different contexts. The same approach - building a focused model, testing it on multiple data sets, and publishing the code - applies to any specialized AI application, from quality control in manufacturing to fraud detection in finance. Those interested in broader applications of AI in AI for Science & Research will find this study a useful reference point for how computational tools are reshaping fields that have traditionally relied on manual analysis.
Your membership also unlocks:
