Researchers have developed a new way to train artificial intelligence (AI) to mimic how human pathologists search for cancer in tissue samples, potentially improving the AI’s ability to detect cancer. Unlike many AI systems that examine fixed areas of a pathology slide, pathologists dynamically scan the tissue, zooming in and out and focusing on areas that seem suspicious. A single pathology slide can contain billions of pixels, but cancer may be present in only a very small portion of it. Zhi Huang, a co-author of the study and assistant professor of pathology at the University of Pennsylvania, compared the search for cancer to a search-and-rescue helicopter. "You don’t start by inspecting one square meter of ground," he said. "You scan the landscape first and then swoop in for a closer look." This approach was used to train a new AI system called Pathology-o3, which learns from how pathologists navigate slides rather than from labeled cancer regions or final diagnoses. The researchers trained their AI on the search behavior of eight pathologists, recording how they moved across slides and adjusted magnification levels. They focused on moments that appeared to reflect deliberate attention, such as lingering over a specific area or making a sustained scan. These observations were matched with eye-tracking data to confirm where the pathologists were actually looking. For each area inspected, the AI also generated a short explanation of why the region was of interest, which human pathologists could review, edit, or reject, creating more training data for the AI. Pathology-o3 scans a slide at low resolution first, then selects areas that appear suspicious for closer examination by a more detailed AI model. In tests, the system correctly identified 100% of slides that contained cancer, but it also flagged 15.5% of healthy slides as potentially cancerous. While this rate of false positives is higher than some other AI models, the researchers designed the system to prioritize flagging areas for further review rather than risking missing cancer. The system is not yet precise enough to make a standalone diagnosis, but it could help pathologists focus on the most relevant parts of a slide. Future studies will test whether using Pathology-o3 improves the speed or accuracy of pathologists in real-world scenarios. Researchers believe the key takeaway is that the missing piece for AI in pathology has been available all along — the way pathologists naturally explore slides.