A research group at University Hospital Cologne has shifted from building AI models for individual cancers to a single pan-tumoral algorithm that segments multiple tumour types with the same accuracy. The approach cuts the time needed to annotate pathology images from more than a year down to days, removing a major bottleneck in developing clinical AI tools.
Accelerating annotation from years to days
Four years ago, the Cologne team published findings on prostate cancer detection with solid accuracy before pivoting to multi-segmentation algorithms for colorectal cancer. Annotations were the main obstacle. "It took us more than one year to prepare all annotations to train it," said Dr Yuri Tolkach, senior attending physician and research group lead at the Institute of Pathology. Today, he said, "there are different ways we can accelerate the annotations," and the group now produces precise algorithms for prostate, lung, and colorectal tumours in a fraction of the time.
Moving toward pan-cancer algorithms
The latest phase asks whether one model can handle many cancers. "The question was, why can't we develop one algorithm for more types of cancers; we have lung, colon and breast AI, but what if we applied these algorithms to another tumour type, why can't we develop one algorithm for all these cancers?" Tolkach said. When the group tested its five existing algorithms on cervical and other tumour types, segmentation performance remained high. This kind of cross-tumour generalisation points to a fast-track annotation principle that the researchers see as a practical breakthrough, though Tolkach stressed that clinical validation remains essential.
Advanced algorithms are already giving pathologists a clearer view of tumours from the outset. Large language models pull structured data from AI outputs to train downstream models and enable large-scale analysis. Such tools are relevant for tasks plagued by interobserver variability, such as grading tumour aggressivity. The group's work highlights a broader trend in AI for Healthcare, where machine learning models are increasingly used to support pathologists rather than replace them.
Modelling tumour evolution with multi-modal AI
Multi-modal models are also helping predict relapse after immune checkpoint therapy stops in malignant melanoma. Tolkach described how extracted cell-level data feeds mathematical simulations of tumour growth. "When we extract all this information from tumour cells, we can create a model on our computers of tumour growth and see how that tumour grows. That intratumoral heterogeneity is so close to the real-world data," he said.
Why this matters for healthcare professionals
For pathologists and oncologists, a single deployable algorithm that works across cancer types reduces the need to validate separate tools for each disease site. Faster annotation pipelines mean hospitals can build or customise models on local data without multi-year timelines. The integration of agentic AI and LLMs into diagnostic workflows may also standardise tasks where human judgement varies, lowering second-opinion delays and helping clinics scale precision oncology.
Your membership also unlocks:
