A new artificial intelligence model developed by researchers at Charité – Universitätsmedizin Berlin could transform how brain tumors are diagnosed, especially in cases where traditional biopsies are risky or impossible. The innovation comes in response to complex cases like that of a patient who first sought medical help for double vision. An MRI revealed a tumor located in a part of the brain that made surgical biopsy highly dangerous.
Confronted with such challenges, the team of researchers turned to an alternative approach. Instead of relying on tissue samples, they developed a method that uses an AI model to analyze the epigenetic fingerprint of tumors—chemical modifications in the genetic material that act like cellular memory and regulate gene activity. These fingerprints can be collected from body fluids such as cerebrospinal fluid, making the process minimally invasive.
Published in the journal Nature Cancer, the research highlights how this AI model can rapidly and accurately classify tumors based on these unique molecular patterns. This represents a shift from traditional diagnostic techniques, which often rely on visual examination of tissue under a microscope. While this standard method provides important information, it may fall short in precision, especially when dealing with rare or complex tumor types.
Modern oncology increasingly depends on accurately identifying tumor types at the molecular level. Tumors today are classified not just by the organ in which they appear but also by their unique molecular, structural, and metabolic characteristics. Effective treatment—including targeted therapies, chemotherapy, or enrollment in clinical trials—depends on a correct and specific diagnosis.
To meet this demand, the research team developed crossNN, a neural network model trained on the epigenetic profiles of over 8,000 tumors. When presented with data from an unknown tumor, crossNN compares it against its vast reference library and identifies the most likely match. In testing, the model correctly diagnosed brain tumors in over 99% of cases. A broader version of the same model, trained to classify over 170 types of tumors from all organs, achieved an accuracy rate of 97.8%.
This level of precision was achieved despite the model being based on a relatively simple neural network architecture. The simplicity makes its results more interpretable, a key requirement for clinical adoption. Understanding how the AI arrives at a diagnosis is essential in gaining regulatory approval and the trust of healthcare providers.
In practice, the model can analyze either tissue samples or fluids such as cerebrospinal fluid. In the case of the patient with double vision, researchers were able to extract cerebrospinal fluid and analyze it using nanopore sequencing—a rapid, high-resolution genetic sequencing method. The AI model identified the tumor as a lymphoma of the central nervous system, enabling doctors to begin targeted chemotherapy without ever needing to perform surgery.
This success is already influencing clinical practice. The Department of Neuropathology at Charité currently offers this non-invasive diagnostic technique for specific brain tumors, marking a significant step forward in patient care. The method not only avoids the risks of invasive procedures but also accelerates diagnosis and treatment, which is critical for aggressive cancers.
The team is now preparing for clinical trials of crossNN across all eight sites of the German Cancer Consortium (DKTK). They also plan to explore the AI model’s potential use during surgery to provide real-time tumor analysis. The long-term goal is to make accurate, cost-effective tumor classification using DNA samples a routine part of clinical care.
By combining molecular biology with artificial intelligence, this approach offers a safer, faster, and more precise pathway to cancer diagnosis—especially for patients for whom traditional biopsies carry unacceptable risks.
By Impact Lab