Pan‑cancer AI diagnostics are moving from experimental labs to hospital wards, as researchers in Cologne report faster tissue analysis and broader applicability across tumor types.
From single‑tumor tools to broader platforms
Dr. Yuri Tolkach, head of the digital pathology group at University Hospital Cologne, described a shift that began with prostate cancer detection and now includes lung, colorectal and breast cancers. Early work relied on simple pattern‑recognition, but newer versions incorporate large language models and multi‑modal reasoning to aid clinicians.
Speeding up the annotation bottleneck
Four years ago the team published results on prostate tumors, achieving respectable accuracy. They later introduced a multi‑segmentation approach for colorectal disease, which improved performance but required extensive manual labeling. “Annotations are the major problem for colorectal cancer,” Tolkach said. “It took us more than one year to prepare all annotations to train it.”
Related: Non-surgical solution provides lasting knee pain relief
Recent advances have cut that timeline dramatically. The group now reports that, for certain cancers, the labeling phase can be completed in days rather than months. The acceleration stems from a suite of precise tools tailored to each organ system, allowing faster data collection for training.
One model, many cancers?
In the latest phase the researchers tested five separate models on a variety of tumor types. They found that the same segmentation quality held up when the tools were applied to cervical and other malignancies. “Why not develop one algorithm for more types of cancers,” Tolkach asked, noting that existing solutions for lung, colon and breast could be extended.
The findings suggest a path toward a single, pan‑tumor system, though clinical validation remains essential. Independent experts caution that broader deployment will require rigorous trials across diverse patient populations.
Extracting data and predicting outcomes
Beyond identification, AI is being used to address interobserver variability in grading aggressiveness. Large language models pull parameters from imaging data, enabling large‑scale analyses that were previously impractical. Multi‑modal frameworks also forecast relapse after immune checkpoint inhibition in malignant melanoma, using mathematical simulations of tumor growth.
Related: Should I Heed My GP’s Advice to See a Pain Management Doctor
“When we extract all this information from tumor cells, we can create a model on our computers of tumor growth and see how that tumor grows,” Tolkach explained. “That intratumoral heterogeneity is so close to the real‑world data.”
External perspective
Dr. Anika Schreiber, a pathologist not involved with the project, said the promise is clear but emphasized the need for transparency. “Clinicians need to understand how these tools arrive at a diagnosis,” she noted. “Without that, adoption will be slow.”
Clinical integration and future steps
The Cologne platform currently offers fully automated analysis of lung cancer tissue sections, delivering rapid results that could streamline diagnostic workflows. Researchers plan to expand the system to include molecular characterization, linking histology with genetic data for a fuller view.
Regulatory pathways are being explored, and the team hopes that the fast‑track annotation principle will satisfy quality‑control requirements. If successful, the approach could set a benchmark for other institutions seeking to combine AI with digital pathology.
Related: 3D map reveals new insights into Tetralogy of Fallot
Key numbers
Annotation time reduced from over a year to days; five models tested across multiple cancers; accuracy comparable to dedicated organ‑specific tools.
AI is reshaping pathology.
While the technology is still evolving, the move toward pan‑cancer AI diagnostics marks a notable step in the integration of artificial intelligence with everyday pathology practice.