giving pathologists the super resolution vision to scan millions of invisible biomarkers
In Brisbane, a team of scientists has built a tool that teaches machines to see what human eyes have never been able to see in the quiet language of tissue and cells. STimage, developed at QIMR Berghofer, applies artificial intelligence to the ordinary slides pathologists have used for generations — revealing the molecular signatures of cancer that staining and microscopy alone cannot surface. It is a quiet but consequential step in the long human effort to catch disease before it announces itself too loudly, and to extend the reach of specialist knowledge to those who live far from its centers.
- Cancers are hiding in tissue samples that pathologists examine every day, invisible not from lack of skill but from the limits of what human eyes and conventional staining can reveal.
- STimage disrupts that boundary by reading molecular patterns in standard H&E slides, accurately detecting breast, skin, and kidney cancers as well as liver immune disease — and flagging which patients may respond poorly to treatment.
- Unlike opaque AI systems, STimage shows its reasoning, identifying the specific cellular features behind each prediction and keeping the pathologist firmly in the role of final arbiter.
- Because it runs on slides already in use and delivers rapid results, the tool could bring specialist-level molecular diagnostics to regional and remote clinics that currently lack that access.
- Clinical trials in pathology labs are the next step, with researchers aiming for integration into standard practice within two years as they expand the tool to cover more cancer types and earlier stages.
A research team at QIMR Berghofer in Brisbane has developed an AI tool called STimage that can detect cancers hidden within standard pathology slides — the same hematoxylin and eosin preparations pathologists have examined under microscopes for over a century. While H&E staining reveals physical changes in cells, the molecular activity that signals disease remains invisible to conventional examination. STimage bridges that gap by applying spatial analysis to these slides, surfacing genetic markers that human eyes cannot see.
Led by Associate Professor Quan Nguyen, the project demonstrated in Nature Communications that the tool could accurately predict breast, skin, and kidney cancers, as well as a liver immune disease. It also showed potential in identifying which patients were likely to respond to existing drugs and which faced higher risk of poor outcomes. Crucially, STimage does not simply deliver a verdict — it explains which tissue features and cellular characteristics drove each prediction, preserving the pathologist's role as the final decision-maker.
The practical case for the tool is strong. It operates on slides already present in standard workflows, making it fast and low-cost. For patients in regional and remote areas, it could deliver the kind of molecular diagnostic precision that previously required access to major research centers. The team is now expanding the model to cover more cancer types and earlier-stage detection, with clinical trials in pathology laboratories planned as the next phase. If those trials succeed, STimage could enter standard clinical practice within two years — a marker in the broader shift toward digital pathology and treatment decisions shaped by individual molecular profiles.
A team of researchers at QIMR Berghofer in Brisbane has built an artificial intelligence tool that can spot cancers hiding in plain sight within tissue samples. The system, called STimage, works by analyzing the molecular patterns in standard pathology slides—the kind pathologists have been examining under microscopes for over a century. What makes it different is that it can see what human eyes cannot.
Pathologists have long relied on a staining technique called hematoxylin and eosin, or H&E, to examine tissue structure and identify abnormalities. The method works well for revealing physical changes in cells, but it tells only part of the story. The molecular activity beneath the surface—the genetic markers that signal disease—remains invisible. STimage bridges that gap. By applying spatial analysis to these standard slides, the tool can detect the molecular signatures of cancer that exist but cannot be seen with conventional examination.
Associate Professor Quan Nguyen, who led the tool's development with QIMR Berghofer's National Centre for Spatial Tissue and AI Research, describes the capability in stark terms: it gives pathologists the ability to scan millions of invisible biomarkers in a tiny sample and identify the few that matter. The research, published in Nature Communications, demonstrated that STimage could accurately predict breast, skin, and kidney cancers, as well as a liver immune disease. The tool also showed promise in predicting which patients would respond well to existing drugs and which faced higher risk of poor outcomes.
What distinguishes STimage from other AI diagnostic tools is not just its accuracy but its transparency. The system doesn't simply deliver a verdict. It shows its work. When it makes a prediction, it explains which specific tissue features and cellular characteristics led to that conclusion, allowing pathologists to evaluate the reasoning and maintain their role as the final decision-maker. Nguyen is careful to frame the tool not as a replacement for pathologist expertise but as an enhancement—a way to surface information that would otherwise remain hidden, freeing experienced clinicians to focus their judgment where it matters most.
The practical advantages are substantial. The tool works on slides that already exist in standard pathology workflows, making it low-cost and fast. Results are generated quickly and presented in a format pathologists can readily interpret. For patients in regional and remote areas, where access to specialist expertise is limited, the technology could democratize diagnostic precision. A pathologist in a small town could use STimage to access the kind of molecular analysis that previously required sending samples to major research centers.
The research team trained the model using machine learning on de-identified datasets of the cancers it can now detect. They are already working to expand its capabilities—broadening the range of cancer types it can identify, improving its accuracy, and teaching it to spot rarer cancers at earlier stages. The next phase is clinical trials in actual pathology laboratories. If those trials succeed, the researchers believe STimage could become part of standard clinical practice within two years. The tool represents a shift toward what researchers call digital pathology and precision medicine: faster diagnosis, more accurate classification, and treatment decisions tailored to individual molecular profiles rather than broad categories.
Notable Quotes
It's like giving pathologists the super resolution vision of Superman or Superwoman to scan millions of invisible biomarkers in a tiny tissue sample to find the two or three that are showing signs of cancer.— Associate Professor Quan Nguyen, QIMR Berghofer
The STimage tool does not replace the experience and expertise of pathologists. Rather, it assists them in their important and technically challenging work, by providing extra information about cell types and genetic activity that they can't see with their own eyes.— Associate Professor Quan Nguyen
The Hearth Conversation Another angle on the story
So this tool looks at the same slides pathologists have always looked at, but sees something different. What exactly is it detecting that human eyes miss?
The molecular activity inside the tissue. A standard stain shows you the structure—the shape and arrangement of cells. But cancer isn't just about shape. It's about what's happening at the genetic level, which proteins are active, how cells are communicating. That's invisible under a microscope, but it's there in the tissue. STimage can read it.
And it does this by analyzing the same physical slide?
Yes. It applies spatial analysis to the standard slide. It's learning patterns in how molecules are distributed across the tissue, then matching those patterns to what it knows about cancer. It's like reading a map of activity that was always there but couldn't be seen.
The tool explains its reasoning. Why is that important?
Because a pathologist needs to trust it and understand it. If a machine just says "cancer," that's not useful. But if it says "cancer, because these three cell types are clustered this way and these genetic markers are active," then the pathologist can evaluate that reasoning. They stay in control.
What happens next?
Clinical trials in real pathology labs. If those work, it could be in standard use within two years. But they're also expanding it—more cancer types, rarer cancers, immune cells that predict drug response. The foundation is solid. Now it's about scale.