AI Trained on Pathologist Behavior Shows Promise in Cancer Detection

Artificial intelligence systems designed to detect cancer may achieve greater accuracy when programmed to emulate the dynamic visual search strategies of human pathologists, according to a new study published in July in the journal Nature. While many existing AI tools analyze static, preselected tissue regions or fixed-size image patches, human experts scan slides by panning, zooming, and lingering on areas of concern. The research, led by Zhi Huang, an assistant professor of pathology and laboratory medicine at the University of Pennsylvania, suggests that teaching AI to adopt this flexible approach could improve cancer identification.

“You don’t start by inspecting one square meter of ground,” Huang explained, comparing the process to a search-and-rescue helicopter that surveys a broad landscape before swooping in for a closer look at potential targets. Because a single pathology slide can contain billions of pixels, with cancerous evidence potentially hidden in a minuscule fraction, this hierarchical scanning method is critical.

The researchers developed a training methodology termed “Pathology-CoT,” or chain of thought, which converts observable pathologist behaviors into training data. Using a custom tool, they recorded the movements and magnification changes of eight pathologists as they examined tissue slides. After filtering out incidental movements to isolate deliberate actions, such as sustained pans or pauses, the team validated the data against eye-tracking records to ensure accuracy. Additionally, vision language models (VLMs) generated rationales for why specific regions were worth examining, which human pathologists could then accept, edit, or reject to further refine the AI’s learning.

This approach led to the creation of Pathology-o3, a system that initially scans a slide at low resolution, identifies regions of interest using the behavior-trained model, and then subjects those specific areas to high-resolution VLM analysis. Unlike specialized AI models trained for single diseases, Pathology-o3 was tested as a general-purpose tool capable of navigating complex slides effectively.

In tests using lymph node tissue slides from colorectal cancer cases, Pathology-o3 correctly identified all cancer-positive slides, achieving 100% sensitivity. However, the model produced a false positive rate of 15.5%, meaning some slides flagged as positive were actually negative. In comparison, OpenAI’s o3 model identified 87.5% of positive slides correctly but had a significantly higher false positive rate of 53.3%. Huang noted that the system is designed to err on the side of caution to avoid missing cancers, which accounts for the higher false alarm rate.

When tested on an independent dataset the algorithm had not previously encountered, Pathology-o3 maintained high sensitivity at 97.6%, though the false positive rate rose to 37.1%. Mohammad Asadi, a data scientist at Stanford University not involved in the study, observed that while the system is not precise enough for autonomous diagnosis, it could serve as a valuable prescreening tool. He highlighted the benefit of the AI directing human attention to specific regions rather than labeling entire slides as suspicious.

Asadi emphasized that further research is needed to determine if the tool actually improves pathologist efficiency and accuracy in clinical settings. He called for multi-hospital trials that assess workload, the burden of false alarms, and the ability of doctors to recognize AI errors. Huang agreed, stating that the goal is not to replace pathologists but to provide a system where “a pathologist working with it catches more [cancer cases] and works faster.” The next phase of the research will directly compare pathologist performance with and without the AI assistant. The study also noted that current AI diagnostics often rely on single slides, whereas human diagnoses typically integrate multiple slides, stains, and patient history.