QIMR Berghofer scientists have developed an AI screening tool that harnesses the power of cutting-edge spatial biology analysis to give pathologists ‘super vision’ to detect hidden genetic markers of cancer in standard patient tissue samples.
New research published in Nature Communications shows how the machine learning tool, STimage, accurately predicted breast, skin and kidney cancers and a liver immune disease. It found the tool was reliable, low cost and rapidly generated results that were easy for pathologists to interpret.
The breakthrough could help deliver a new era of digital pathology and precision medicine, saving lives through faster and more accurate diagnoses, personalised treatments, and improved access to specialist expertise for patients in regional and remote areas.
“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. This capability is critical for earlier detection, more precise diagnosis, and better-informed treatment decisions,” said Associate Professor Quan Nguyen, who led development of the tool with QIMR Berghofer’s National Centre for Spatial Tissue and AI Research (NCSTAR).
The team hopes this tool could help pathologists across rural, regional and metropolitan areas manage high demand and workload, enhance diagnostic precision and reduce the time involved in screening and analysing samples by providing a way for them to access crucial molecular information currently limited to specialist research centres.
“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,” A/Prof Nguyen said.
Spatial biology is a new field that looks at the complex molecular activity within the tissue microenvironment to help uncover the causes of cancer and other diseases. It provides insights that are not accessible using the standard pathology approach of examining hematoxylin and eosin (H&E)-stained slides under the microscope.
H&E staining is the routine, low-cost method used by pathologists worldwide for more than a century. The characteristic blue and pink staining highlights cell features and tissue architecture allowing doctors to identify structural abnormalities, but it does not directly reveal the underlying molecular activity within the tissue.
A/Prof Nguyen said the STimage tool applies spatial analysis over a H&E slide to generate a biologically-grounded prediction of disease based on the molecular patterns detected in the tissue.
“It makes a diagnostic prediction and mathematically computes the level of certainty about the result. In a first, there is transparency about the result with the tool showing the reasons that led to the prediction, like specific tissue or cellular features, to help pathologists evaluate findings,” A/Prof Nguyen said.
The tool also generated accurate predictions about prognosis and treatment response, correctly classifying patients as high or low risk of survival and likely to have a complete or partial response to existing drugs. These features are at an early stage and are being developed further.
The researchers trained the model using machine learning and statistical algorithms to spatially learn from de-identified data sets of breast, skin and kidney cancers, and the liver disease primary sclerosing cholangitis.
Only a few comparable tools exist in this pioneering field, but the team’s research showed STimage outperformed those models, while adding critical features about reliability and interpretability of the model prediction.
The scientists are continuing to improve the tool by broadening the cancer types it can detect, increasing its accuracy, and integrating more data sets to identify rarer cancer cells at early stage and important immune cell types that determine cancer progression and response to drugs.
The next stage is to trial the model in pathology labs, with the research team hoping the STimage tool could be part of clinical practice within two years.
The study is available at this link https://www.nature.com/articles/s41467-026-68487-0 in Nature Communications with DOI 10.1038/s41467-026-68487-0.
A/Prof Quan Nguyen, QIMR Berghofer National Centre for Spatial Tissue & AI Research (NCSTAR)