Researchers at the University of Pennsylvania have developed an artificial intelligence system designed to mimic how human pathologists examine tissue samples for cancer. Rather than analyzing pre-selected regions or dividing slides into uniform sections, the new approach teaches AI to scan slides dynamically—zooming in and out while focusing on suspicious areas, much like a pathologist would.
The team created a training method called “Pathology-CoT” by recording how eight pathologists navigated digital slides, capturing their deliberate movements and areas of focus. This behavioral data was then used to train a general-purpose AI system called Pathology-o3, which first scans at low resolution before directing higher-resolution analysis to promising regions. The system also generates explanations for why specific areas warrant examination, allowing pathologists to validate the AI’s reasoning.
Testing on lymph node tissue samples from colorectal cancer cases showed Pathology-o3 identified positive slides with perfect accuracy, though it flagged some negative slides as suspicious—a trade-off the researchers made to avoid missing cancer. Performance varied when tested on unfamiliar data, suggesting the approach shows promise but requires further development. Experts note the system could serve as a prescreening tool rather than an independent diagnostic solution.
The researchers emphasize their next phase will measure whether pathologists actually work faster and catch more cancers when assisted by the technology. The current study did not compare the AI directly against human pathologists in clinical practice.
Originally reported by Livescience →
About the Author
