An international research team led by Prof. Jakob N. Kather from the Else Kröner Fresenius Centre for Digital Health (TU Dresden) has developed a new AI model that can simultaneously detect multiple genetic changes in colon cancer. This is done directly from standard histological sections. In the future, the technology could contribute to faster and more cost-effective diagnostics.
Previous AI models were usually limited to predicting one genetic abnormality at a time. The multi-target transformation model now developed goes one step further: it can identify multiple biomarkers in a single analysis. This does justice to the complexity of colon cancer, in which multiple mutations and morphological changes often occur simultaneously.
For the study, the team analysed nearly 2,000 digital tissue samples from colon cancer patients from seven independent cohorts in Europe and the United States. The dataset included both complete tissue samples and clinical, demographic and lifestyle information.
Better prediction of clinically relevant biomarkers
‘Our model can detect many different biomarkers simultaneously, including mutations that are not yet directly clinically relevant,’ says first author Marco Gustav (EKFZ, TU Dresden). ‘We also saw that many mutations occur more frequently in microsatellite-unstable tumours (MSI). The model recognises shared visual patterns rather than assessing individual mutations separately.’
Microsatellite instability (MSI) is an important biomarker in colon cancer. Patients with MSI-positive tumours are eligible for immunotherapy. The fact that the new model can accurately predict MSI, in addition to mutations such as BRAF and RNF43, demonstrates the clinical potential of this approach.
Impact on diagnostics and treatment
According to Prof. Kather, this innovation could significantly speed up the diagnostic workflow. ‘AI models such as this make it possible to extract molecular information directly from routinely obtained tissue sections. This allows clinicians to decide more quickly which patients require further molecular testing and which treatment options are appropriate.’
The researchers emphasise that the model is not intended to completely replace existing molecular tests, but rather as a pre-screening tool that makes healthcare more efficient and guides patients to the right therapy more quickly.
International collaboration
The research, published in The Lancet Digital Health, was made possible by intensive collaboration between experts in data science, pathology, oncology and epidemiology. In addition to TU Dresden and Dresden University Hospital, the University of Augsburg, NCT Heidelberg, the Medical University of Vienna, the Fred Hutchinson Cancer Centre (Seattle) and the Mayo Clinic (USA) were also involved.
The researchers now want to expand the approach to other types of cancer. This could make the AI model an important step towards broadly applicable, AI-driven diagnostics in oncology.
AI and colon cancer research
Last year, the COLO-DETECT study was conducted in England to test the GI Genius AI system. This AI tool, integrated into existing colonoscopy equipment, increases the effectiveness of detecting potentially malignant lesions such as adenomas. This also included small or flat polyps that may not be visible to the human eye.
In the Netherlands, similar AI support has been in use at the Catharina Hospital since 2022, where suspicious areas are marked with a “square” during colonoscopies.
